1 /*
2 * Copyright (c) 2015, 2020, Red Hat, Inc. All rights reserved.
3 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #ifndef SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
27 #define SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
28
29 #include "gc/shenandoah/shenandoahHeap.hpp"
30
31 #include "classfile/javaClasses.inline.hpp"
32 #include "gc/shared/markBitMap.inline.hpp"
33 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
34 #include "gc/shared/continuationGCSupport.inline.hpp"
35 #include "gc/shared/suspendibleThreadSet.hpp"
36 #include "gc/shared/tlab_globals.hpp"
37 #include "gc/shenandoah/shenandoahAsserts.hpp"
38 #include "gc/shenandoah/shenandoahBarrierSet.inline.hpp"
39 #include "gc/shenandoah/shenandoahCollectionSet.inline.hpp"
40 #include "gc/shenandoah/shenandoahForwarding.inline.hpp"
41 #include "gc/shenandoah/shenandoahWorkGroup.hpp"
42 #include "gc/shenandoah/shenandoahHeapRegionSet.inline.hpp"
43 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
44 #include "gc/shenandoah/shenandoahGeneration.hpp"
45 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
46 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
47 #include "gc/shenandoah/mode/shenandoahMode.hpp"
48 #include "oops/compressedOops.inline.hpp"
49 #include "oops/oop.inline.hpp"
50 #include "runtime/atomic.hpp"
51 #include "runtime/javaThread.hpp"
52 #include "runtime/prefetch.inline.hpp"
53 #include "runtime/objectMonitor.inline.hpp"
54 #include "utilities/copy.hpp"
55 #include "utilities/globalDefinitions.hpp"
56
57 inline ShenandoahHeap* ShenandoahHeap::heap() {
58 return named_heap<ShenandoahHeap>(CollectedHeap::Shenandoah);
59 }
60
61 inline ShenandoahHeapRegion* ShenandoahRegionIterator::next() {
62 size_t new_index = Atomic::add(&_index, (size_t) 1, memory_order_relaxed);
63 // get_region() provides the bounds-check and returns null on OOB.
64 return _heap->get_region(new_index - 1);
65 }
66
67 inline WorkerThreads* ShenandoahHeap::workers() const {
68 return _workers;
69 }
70
71 inline WorkerThreads* ShenandoahHeap::safepoint_workers() {
72 return _safepoint_workers;
73 }
74
75 inline void ShenandoahHeap::notify_gc_progress() {
76 Atomic::store(&_gc_no_progress_count, (size_t) 0);
77
78 }
79 inline void ShenandoahHeap::notify_gc_no_progress() {
80 Atomic::inc(&_gc_no_progress_count);
81 }
82
83 inline size_t ShenandoahHeap::get_gc_no_progress_count() const {
84 return Atomic::load(&_gc_no_progress_count);
85 }
86
87 inline size_t ShenandoahHeap::heap_region_index_containing(const void* addr) const {
88 uintptr_t region_start = ((uintptr_t) addr);
89 uintptr_t index = (region_start - (uintptr_t) base()) >> ShenandoahHeapRegion::region_size_bytes_shift();
90 assert(index < num_regions(), "Region index is in bounds: " PTR_FORMAT, p2i(addr));
91 return index;
92 }
93
94 inline ShenandoahHeapRegion* ShenandoahHeap::heap_region_containing(const void* addr) const {
95 size_t index = heap_region_index_containing(addr);
96 ShenandoahHeapRegion* const result = get_region(index);
97 assert(addr >= result->bottom() && addr < result->end(), "Heap region contains the address: " PTR_FORMAT, p2i(addr));
98 return result;
99 }
100
101 inline void ShenandoahHeap::enter_evacuation(Thread* t) {
102 _oom_evac_handler.enter_evacuation(t);
103 }
104
105 inline void ShenandoahHeap::leave_evacuation(Thread* t) {
106 _oom_evac_handler.leave_evacuation(t);
107 }
108
109 template <class T>
110 inline void ShenandoahHeap::update_with_forwarded(T* p) {
111 T o = RawAccess<>::oop_load(p);
112 if (!CompressedOops::is_null(o)) {
113 oop obj = CompressedOops::decode_not_null(o);
114 if (in_collection_set(obj)) {
115 // Corner case: when evacuation fails, there are objects in collection
116 // set that are not really forwarded. We can still go and try and update them
117 // (uselessly) to simplify the common path.
118 shenandoah_assert_forwarded_except(p, obj, cancelled_gc());
119 oop fwd = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
120 shenandoah_assert_not_in_cset_except(p, fwd, cancelled_gc());
121
122 // Unconditionally store the update: no concurrent updates expected.
123 RawAccess<IS_NOT_NULL>::oop_store(p, fwd);
124 }
125 }
126 }
127
128 template <class T>
129 inline void ShenandoahHeap::conc_update_with_forwarded(T* p) {
130 T o = RawAccess<>::oop_load(p);
131 if (!CompressedOops::is_null(o)) {
132 oop obj = CompressedOops::decode_not_null(o);
133 if (in_collection_set(obj)) {
134 // Corner case: when evacuation fails, there are objects in collection
135 // set that are not really forwarded. We can still go and try CAS-update them
136 // (uselessly) to simplify the common path.
137 shenandoah_assert_forwarded_except(p, obj, cancelled_gc());
138 oop fwd = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
139 shenandoah_assert_not_in_cset_except(p, fwd, cancelled_gc());
140
141 // Sanity check: we should not be updating the cset regions themselves,
142 // unless we are recovering from the evacuation failure.
143 shenandoah_assert_not_in_cset_loc_except(p, !is_in(p) || cancelled_gc());
144
145 // Either we succeed in updating the reference, or something else gets in our way.
146 // We don't care if that is another concurrent GC update, or another mutator update.
147 atomic_update_oop(fwd, p, obj);
148 }
149 }
150 }
151
152 // Atomic updates of heap location. This is only expected to work with updating the same
153 // logical object with its forwardee. The reason why we need stronger-than-relaxed memory
154 // ordering has to do with coordination with GC barriers and mutator accesses.
155 //
156 // In essence, stronger CAS access is required to maintain the transitive chains that mutator
157 // accesses build by themselves. To illustrate this point, consider the following example.
158 //
159 // Suppose "o" is the object that has a field "x" and the reference to "o" is stored
160 // to field at "addr", which happens to be Java volatile field. Normally, the accesses to volatile
161 // field at "addr" would be matched with release/acquire barriers. This changes when GC moves
162 // the object under mutator feet.
163 //
164 // Thread 1 (Java)
165 // // --- previous access starts here
166 // ...
167 // T1.1: store(&o.x, 1, mo_relaxed)
168 // T1.2: store(&addr, o, mo_release) // volatile store
169 //
170 // // --- new access starts here
171 // // LRB: copy and install the new copy to fwdptr
172 // T1.3: var copy = copy(o)
173 // T1.4: cas(&fwd, t, copy, mo_release) // pointer-mediated publication
174 // <access continues>
175 //
176 // Thread 2 (GC updater)
177 // T2.1: var f = load(&fwd, mo_{consume|acquire}) // pointer-mediated acquisition
178 // T2.2: cas(&addr, o, f, mo_release) // this method
179 //
180 // Thread 3 (Java)
181 // T3.1: var o = load(&addr, mo_acquire) // volatile read
182 // T3.2: if (o != null)
183 // T3.3: var r = load(&o.x, mo_relaxed)
184 //
185 // r is guaranteed to contain "1".
186 //
187 // Without GC involvement, there is synchronizes-with edge from T1.2 to T3.1,
188 // which guarantees this. With GC involvement, when LRB copies the object and
189 // another thread updates the reference to it, we need to have the transitive edge
190 // from T1.4 to T2.1 (that one is guaranteed by forwarding accesses), plus the edge
191 // from T2.2 to T3.1 (which is brought by this CAS).
192 //
193 // Note that we do not need to "acquire" in these methods, because we do not read the
194 // failure witnesses contents on any path, and "release" is enough.
195 //
196
197 inline void ShenandoahHeap::atomic_update_oop(oop update, oop* addr, oop compare) {
198 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr));
199 Atomic::cmpxchg(addr, compare, update, memory_order_release);
200 }
201
202 inline void ShenandoahHeap::atomic_update_oop(oop update, narrowOop* addr, narrowOop compare) {
203 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
204 narrowOop u = CompressedOops::encode(update);
205 Atomic::cmpxchg(addr, compare, u, memory_order_release);
206 }
207
208 inline void ShenandoahHeap::atomic_update_oop(oop update, narrowOop* addr, oop compare) {
209 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
210 narrowOop c = CompressedOops::encode(compare);
211 narrowOop u = CompressedOops::encode(update);
212 Atomic::cmpxchg(addr, c, u, memory_order_release);
213 }
214
215 inline bool ShenandoahHeap::atomic_update_oop_check(oop update, oop* addr, oop compare) {
216 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr));
217 return (oop) Atomic::cmpxchg(addr, compare, update, memory_order_release) == compare;
218 }
219
220 inline bool ShenandoahHeap::atomic_update_oop_check(oop update, narrowOop* addr, narrowOop compare) {
221 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
222 narrowOop u = CompressedOops::encode(update);
223 return (narrowOop) Atomic::cmpxchg(addr, compare, u, memory_order_release) == compare;
224 }
225
226 inline bool ShenandoahHeap::atomic_update_oop_check(oop update, narrowOop* addr, oop compare) {
227 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
228 narrowOop c = CompressedOops::encode(compare);
229 narrowOop u = CompressedOops::encode(update);
230 return CompressedOops::decode(Atomic::cmpxchg(addr, c, u, memory_order_release)) == compare;
231 }
232
233 // The memory ordering discussion above does not apply for methods that store nulls:
234 // then, there is no transitive reads in mutator (as we see nulls), and we can do
235 // relaxed memory ordering there.
236
237 inline void ShenandoahHeap::atomic_clear_oop(oop* addr, oop compare) {
238 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr));
239 Atomic::cmpxchg(addr, compare, oop(), memory_order_relaxed);
240 }
241
242 inline void ShenandoahHeap::atomic_clear_oop(narrowOop* addr, oop compare) {
243 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
244 narrowOop cmp = CompressedOops::encode(compare);
245 Atomic::cmpxchg(addr, cmp, narrowOop(), memory_order_relaxed);
246 }
247
248 inline void ShenandoahHeap::atomic_clear_oop(narrowOop* addr, narrowOop compare) {
249 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
250 Atomic::cmpxchg(addr, compare, narrowOop(), memory_order_relaxed);
251 }
252
253 inline bool ShenandoahHeap::cancelled_gc() const {
254 return _cancelled_gc.get() != GCCause::_no_gc;
255 }
256
257 inline bool ShenandoahHeap::check_cancelled_gc_and_yield(bool sts_active) {
258 if (sts_active && !cancelled_gc()) {
259 if (SuspendibleThreadSet::should_yield()) {
260 SuspendibleThreadSet::yield();
261 }
262 }
263 return cancelled_gc();
264 }
265
266 inline GCCause::Cause ShenandoahHeap::cancelled_cause() const {
267 return _cancelled_gc.get();
268 }
269
270 inline void ShenandoahHeap::clear_cancelled_gc(bool clear_oom_handler) {
271 _cancelled_gc.set(GCCause::_no_gc);
272 if (_cancel_requested_time > 0) {
273 log_debug(gc)("GC cancellation took %.3fs", (os::elapsedTime() - _cancel_requested_time));
274 _cancel_requested_time = 0;
275 }
276
277 if (clear_oom_handler) {
278 _oom_evac_handler.clear();
279 }
280 }
281
282 inline HeapWord* ShenandoahHeap::allocate_from_gclab(Thread* thread, size_t size) {
283 assert(UseTLAB, "TLABs should be enabled");
284
285 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
286 if (gclab == nullptr) {
287 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(),
288 "Performance: thread should have GCLAB: %s", thread->name());
289 // No GCLABs in this thread, fallback to shared allocation
290 return nullptr;
291 }
292 HeapWord* obj = gclab->allocate(size);
293 if (obj != nullptr) {
294 return obj;
295 }
296 return allocate_from_gclab_slow(thread, size);
297 }
298
299 void ShenandoahHeap::increase_object_age(oop obj, uint additional_age) {
300 // This operates on new copy of an object. This means that the object's mark-word
301 // is thread-local and therefore safe to access. However, when the mark is
302 // displaced (i.e. stack-locked or monitor-locked), then it must be considered
303 // a shared memory location. It can be accessed by other threads.
304 // In particular, a competing evacuating thread can succeed to install its copy
305 // as the forwardee and continue to unlock the object, at which point 'our'
306 // write to the foreign stack-location would potentially over-write random
307 // information on that stack. Writing to a monitor is less problematic,
308 // but still not safe: while the ObjectMonitor would not randomly disappear,
309 // the other thread would also write to the same displaced header location,
310 // possibly leading to increase the age twice.
311 // For all these reasons, we take the conservative approach and not attempt
312 // to increase the age when the header is displaced.
313 markWord w = obj->mark();
314 // The mark-word has been copied from the original object. It can not be
315 // inflating, because inflation can not be interrupted by a safepoint,
316 // and after a safepoint, a Java thread would first have to successfully
317 // evacuate the object before it could inflate the monitor.
318 assert(!w.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT, "must not inflate monitor before evacuation of object succeeds");
319 // It is possible that we have copied the object after another thread has
320 // already successfully completed evacuation. While harmless (we would never
321 // publish our copy), don't even attempt to modify the age when that
322 // happens.
323 if (!w.has_displaced_mark_helper() && !w.is_marked()) {
324 w = w.set_age(MIN2(markWord::max_age, w.age() + additional_age));
325 obj->set_mark(w);
326 }
327 }
328
329 // Return the object's age, or a sentinel value when the age can't
330 // necessarily be determined because of concurrent locking by the
331 // mutator
332 uint ShenandoahHeap::get_object_age(oop obj) {
333 markWord w = obj->mark();
334 assert(!w.is_marked(), "must not be forwarded");
335 if (w.has_monitor()) {
336 w = w.monitor()->header();
337 } else if (w.is_being_inflated() || w.has_displaced_mark_helper()) {
338 // Informs caller that we aren't able to determine the age
339 return markWord::max_age + 1; // sentinel
340 }
341 assert(w.age() <= markWord::max_age, "Impossible!");
342 return w.age();
343 }
344
345 inline bool ShenandoahHeap::is_in_active_generation(oop obj) const {
346 if (!mode()->is_generational()) {
347 // everything is the same single generation
348 assert(is_in(obj), "Otherwise shouldn't return true below");
349 return true;
350 }
351
352 ShenandoahGeneration* const gen = active_generation();
353
354 if (gen == nullptr) {
355 // no collection is happening: only expect this to be called
356 // when concurrent processing is active, but that could change
357 return false;
358 }
359
360 assert(is_in(obj), "only check if is in active generation for objects (" PTR_FORMAT ") in heap", p2i(obj));
361 assert(gen->is_old() || gen->is_young() || gen->is_global(),
362 "Active generation must be old, young, or global");
363
364 size_t index = heap_region_containing(obj)->index();
365
366 // No flickering!
367 assert(gen == active_generation(), "Race?");
368
369 switch (region_affiliation(index)) {
370 case ShenandoahAffiliation::FREE:
371 // Free regions are in old, young, and global collections
372 return true;
373 case ShenandoahAffiliation::YOUNG_GENERATION:
374 // Young regions are in young and global collections, not in old collections
375 return !gen->is_old();
376 case ShenandoahAffiliation::OLD_GENERATION:
377 // Old regions are in old and global collections, not in young collections
378 return !gen->is_young();
379 default:
380 assert(false, "Bad affiliation (%d) for region " SIZE_FORMAT, region_affiliation(index), index);
381 return false;
382 }
383 }
384
385 inline bool ShenandoahHeap::is_in_young(const void* p) const {
386 return is_in(p) && (_affiliations[heap_region_index_containing(p)] == ShenandoahAffiliation::YOUNG_GENERATION);
387 }
388
389 inline bool ShenandoahHeap::is_in_old(const void* p) const {
390 return is_in(p) && (_affiliations[heap_region_index_containing(p)] == ShenandoahAffiliation::OLD_GENERATION);
391 }
392
393 inline bool ShenandoahHeap::is_in_old_during_young_collection(oop obj) const {
394 return active_generation()->is_young() && is_in_old(obj);
395 }
396
397 inline ShenandoahAffiliation ShenandoahHeap::region_affiliation(const ShenandoahHeapRegion *r) const {
398 return region_affiliation(r->index());
399 }
400
401 inline void ShenandoahHeap::assert_lock_for_affiliation(ShenandoahAffiliation orig_affiliation,
402 ShenandoahAffiliation new_affiliation) {
403 // A lock is required when changing from FREE to NON-FREE. Though it may be possible to elide the lock when
404 // transitioning from in-use to FREE, the current implementation uses a lock for this transition. A lock is
405 // not required to change from YOUNG to OLD (i.e. when promoting humongous region).
406 //
407 // new_affiliation is: FREE YOUNG OLD
408 // orig_affiliation is: FREE X L L
409 // YOUNG L X
410 // OLD L X X
411 // X means state transition won't happen (so don't care)
412 // L means lock should be held
413 // Blank means no lock required because affiliation visibility will not be required until subsequent safepoint
414 //
415 // Note: during full GC, all transitions between states are possible. During Full GC, we should be in a safepoint.
416
417 if (orig_affiliation == ShenandoahAffiliation::FREE) {
418 shenandoah_assert_heaplocked_or_safepoint();
419 }
420 }
421
422 inline void ShenandoahHeap::set_affiliation(ShenandoahHeapRegion* r, ShenandoahAffiliation new_affiliation) {
423 #ifdef ASSERT
424 assert_lock_for_affiliation(region_affiliation(r), new_affiliation);
425 #endif
426 Atomic::store(_affiliations + r->index(), (uint8_t) new_affiliation);
427 }
428
429 inline ShenandoahAffiliation ShenandoahHeap::region_affiliation(size_t index) const {
430 return (ShenandoahAffiliation) Atomic::load(_affiliations + index);
431 }
432
433 inline bool ShenandoahHeap::requires_marking(const void* entry) const {
434 oop obj = cast_to_oop(entry);
435 return !_marking_context->is_marked_strong(obj);
436 }
437
438 inline bool ShenandoahHeap::in_collection_set(oop p) const {
439 assert(collection_set() != nullptr, "Sanity");
440 return collection_set()->is_in(p);
441 }
442
443 inline bool ShenandoahHeap::in_collection_set_loc(void* p) const {
444 assert(collection_set() != nullptr, "Sanity");
445 return collection_set()->is_in_loc(p);
446 }
447
448 inline bool ShenandoahHeap::is_idle() const {
449 return _gc_state_changed ? _gc_state.is_clear() : ShenandoahThreadLocalData::gc_state(Thread::current()) == 0;
450 }
451
452 inline bool ShenandoahHeap::has_forwarded_objects() const {
453 return is_gc_state(HAS_FORWARDED);
454 }
455
456 inline bool ShenandoahHeap::is_concurrent_mark_in_progress() const {
457 return is_gc_state(MARKING);
458 }
459
460 inline bool ShenandoahHeap::is_concurrent_young_mark_in_progress() const {
461 return is_gc_state(YOUNG_MARKING);
462 }
463
464 inline bool ShenandoahHeap::is_concurrent_old_mark_in_progress() const {
465 return is_gc_state(OLD_MARKING);
466 }
467
468 inline bool ShenandoahHeap::is_evacuation_in_progress() const {
469 return is_gc_state(EVACUATION);
470 }
471
472 inline bool ShenandoahHeap::is_update_refs_in_progress() const {
473 return is_gc_state(UPDATE_REFS);
474 }
475
476 inline bool ShenandoahHeap::is_concurrent_weak_root_in_progress() const {
477 return is_gc_state(WEAK_ROOTS);
478 }
479
480 inline bool ShenandoahHeap::is_degenerated_gc_in_progress() const {
481 return _degenerated_gc_in_progress.is_set();
482 }
483
484 inline bool ShenandoahHeap::is_full_gc_in_progress() const {
485 return _full_gc_in_progress.is_set();
486 }
487
488 inline bool ShenandoahHeap::is_full_gc_move_in_progress() const {
489 return _full_gc_move_in_progress.is_set();
490 }
491
492 inline bool ShenandoahHeap::is_stw_gc_in_progress() const {
493 return is_full_gc_in_progress() || is_degenerated_gc_in_progress();
494 }
495
496 inline bool ShenandoahHeap::is_concurrent_strong_root_in_progress() const {
497 return _concurrent_strong_root_in_progress.is_set();
498 }
499
500 template<class T>
501 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl) {
502 marked_object_iterate(region, cl, region->top());
503 }
504
505 template<class T>
506 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* limit) {
507 assert(! region->is_humongous_continuation(), "no humongous continuation regions here");
508
509 ShenandoahMarkingContext* const ctx = marking_context();
510
511 HeapWord* tams = ctx->top_at_mark_start(region);
512
513 size_t skip_bitmap_delta = 1;
514 HeapWord* start = region->bottom();
515 HeapWord* end = MIN2(tams, region->end());
516
517 // Step 1. Scan below the TAMS based on bitmap data.
518 HeapWord* limit_bitmap = MIN2(limit, tams);
519
520 // Try to scan the initial candidate. If the candidate is above the TAMS, it would
521 // fail the subsequent "< limit_bitmap" checks, and fall through to Step 2.
522 HeapWord* cb = ctx->get_next_marked_addr(start, end);
523
524 intx dist = ShenandoahMarkScanPrefetch;
525 if (dist > 0) {
526 // Batched scan that prefetches the oop data, anticipating the access to
527 // either header, oop field, or forwarding pointer. Not that we cannot
528 // touch anything in oop, while it still being prefetched to get enough
529 // time for prefetch to work. This is why we try to scan the bitmap linearly,
530 // disregarding the object size. However, since we know forwarding pointer
531 // precedes the object, we can skip over it. Once we cannot trust the bitmap,
532 // there is no point for prefetching the oop contents, as oop->size() will
533 // touch it prematurely.
534
535 // No variable-length arrays in standard C++, have enough slots to fit
536 // the prefetch distance.
537 static const int SLOT_COUNT = 256;
538 guarantee(dist <= SLOT_COUNT, "adjust slot count");
539 HeapWord* slots[SLOT_COUNT];
540
541 int avail;
542 do {
543 avail = 0;
544 for (int c = 0; (c < dist) && (cb < limit_bitmap); c++) {
545 Prefetch::read(cb, oopDesc::mark_offset_in_bytes());
546 slots[avail++] = cb;
547 cb += skip_bitmap_delta;
548 if (cb < limit_bitmap) {
549 cb = ctx->get_next_marked_addr(cb, limit_bitmap);
550 }
551 }
552
553 for (int c = 0; c < avail; c++) {
554 assert (slots[c] < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(tams));
555 assert (slots[c] < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(limit));
556 oop obj = cast_to_oop(slots[c]);
557 assert(oopDesc::is_oop(obj), "sanity");
558 assert(ctx->is_marked(obj), "object expected to be marked");
559 cl->do_object(obj);
560 }
561 } while (avail > 0);
562 } else {
563 while (cb < limit_bitmap) {
564 assert (cb < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(tams));
565 assert (cb < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(limit));
566 oop obj = cast_to_oop(cb);
567 assert(oopDesc::is_oop(obj), "sanity");
568 assert(ctx->is_marked(obj), "object expected to be marked");
569 cl->do_object(obj);
570 cb += skip_bitmap_delta;
571 if (cb < limit_bitmap) {
572 cb = ctx->get_next_marked_addr(cb, limit_bitmap);
573 }
574 }
575 }
576
577 // Step 2. Accurate size-based traversal, happens past the TAMS.
578 // This restarts the scan at TAMS, which makes sure we traverse all objects,
579 // regardless of what happened at Step 1.
580 HeapWord* cs = tams;
581 while (cs < limit) {
582 assert (cs >= tams, "only objects past TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(tams));
583 assert (cs < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(limit));
584 oop obj = cast_to_oop(cs);
585 assert(oopDesc::is_oop(obj), "sanity");
586 assert(ctx->is_marked(obj), "object expected to be marked");
587 size_t size = obj->size();
588 cl->do_object(obj);
589 cs += size;
590 }
591 }
592
593 template <class T>
594 class ShenandoahObjectToOopClosure : public ObjectClosure {
595 T* _cl;
596 public:
597 ShenandoahObjectToOopClosure(T* cl) : _cl(cl) {}
598
599 void do_object(oop obj) {
600 obj->oop_iterate(_cl);
601 }
602 };
603
604 template <class T>
605 class ShenandoahObjectToOopBoundedClosure : public ObjectClosure {
606 T* _cl;
607 MemRegion _bounds;
608 public:
609 ShenandoahObjectToOopBoundedClosure(T* cl, HeapWord* bottom, HeapWord* top) :
610 _cl(cl), _bounds(bottom, top) {}
611
612 void do_object(oop obj) {
613 obj->oop_iterate(_cl, _bounds);
614 }
615 };
616
617 template<class T>
618 inline void ShenandoahHeap::marked_object_oop_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* top) {
619 if (region->is_humongous()) {
620 HeapWord* bottom = region->bottom();
621 if (top > bottom) {
622 region = region->humongous_start_region();
623 ShenandoahObjectToOopBoundedClosure<T> objs(cl, bottom, top);
624 marked_object_iterate(region, &objs);
625 }
626 } else {
627 ShenandoahObjectToOopClosure<T> objs(cl);
628 marked_object_iterate(region, &objs, top);
629 }
630 }
631
632 inline ShenandoahHeapRegion* ShenandoahHeap::get_region(size_t region_idx) const {
633 if (region_idx < _num_regions) {
634 return _regions[region_idx];
635 } else {
636 return nullptr;
637 }
638 }
639
640 inline ShenandoahMarkingContext* ShenandoahHeap::marking_context() const {
641 return _marking_context;
642 }
643
644 #endif // SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP