1 /*
2 * Copyright (c) 2015, 2020, Red Hat, Inc. 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 #ifndef SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
26 #define SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
27
28 #include "gc/shenandoah/shenandoahHeap.hpp"
29
30 #include "classfile/javaClasses.inline.hpp"
31 #include "gc/shared/markBitMap.inline.hpp"
32 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
33 #include "gc/shared/continuationGCSupport.inline.hpp"
34 #include "gc/shared/suspendibleThreadSet.hpp"
35 #include "gc/shared/tlab_globals.hpp"
36 #include "gc/shenandoah/shenandoahAsserts.hpp"
37 #include "gc/shenandoah/shenandoahBarrierSet.inline.hpp"
38 #include "gc/shenandoah/shenandoahCollectionSet.inline.hpp"
39 #include "gc/shenandoah/shenandoahForwarding.inline.hpp"
40 #include "gc/shenandoah/shenandoahWorkGroup.hpp"
41 #include "gc/shenandoah/shenandoahHeapRegionSet.inline.hpp"
42 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
43 #include "gc/shenandoah/shenandoahControlThread.hpp"
44 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
45 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
46 #include "oops/compressedOops.inline.hpp"
47 #include "oops/oop.inline.hpp"
48 #include "runtime/atomic.hpp"
49 #include "runtime/javaThread.hpp"
50 #include "runtime/prefetch.inline.hpp"
51 #include "utilities/copy.hpp"
52 #include "utilities/globalDefinitions.hpp"
53
54 inline ShenandoahHeap* ShenandoahHeap::heap() {
55 return named_heap<ShenandoahHeap>(CollectedHeap::Shenandoah);
56 }
57
58 inline ShenandoahHeapRegion* ShenandoahRegionIterator::next() {
59 size_t new_index = Atomic::add(&_index, (size_t) 1, memory_order_relaxed);
60 // get_region() provides the bounds-check and returns null on OOB.
61 return _heap->get_region(new_index - 1);
62 }
63
64 inline bool ShenandoahHeap::has_forwarded_objects() const {
65 return _gc_state.is_set(HAS_FORWARDED);
66 }
67
68 inline WorkerThreads* ShenandoahHeap::workers() const {
69 return _workers;
70 }
71
72 inline WorkerThreads* ShenandoahHeap::safepoint_workers() {
73 return _safepoint_workers;
74 }
75
76 inline void ShenandoahHeap::notify_gc_progress() {
77 Atomic::store(&_gc_no_progress_count, (size_t) 0);
78
79 }
80 inline void ShenandoahHeap::notify_gc_no_progress() {
81 Atomic::inc(&_gc_no_progress_count);
82 }
83
84 inline size_t ShenandoahHeap::get_gc_no_progress_count() const {
85 return Atomic::load(&_gc_no_progress_count);
86 }
87
88 inline size_t ShenandoahHeap::heap_region_index_containing(const void* addr) const {
89 uintptr_t region_start = ((uintptr_t) addr);
90 uintptr_t index = (region_start - (uintptr_t) base()) >> ShenandoahHeapRegion::region_size_bytes_shift();
91 assert(index < num_regions(), "Region index is in bounds: " PTR_FORMAT, p2i(addr));
92 return index;
93 }
94
95 inline ShenandoahHeapRegion* ShenandoahHeap::heap_region_containing(const void* addr) const {
96 size_t index = heap_region_index_containing(addr);
97 ShenandoahHeapRegion* const result = get_region(index);
98 assert(addr >= result->bottom() && addr < result->end(), "Heap region contains the address: " PTR_FORMAT, p2i(addr));
99 return result;
100 }
101
102 inline void ShenandoahHeap::enter_evacuation(Thread* t) {
103 _oom_evac_handler.enter_evacuation(t);
104 }
105
106 inline void ShenandoahHeap::leave_evacuation(Thread* t) {
107 _oom_evac_handler.leave_evacuation(t);
108 }
109
110 template <class T>
111 inline void ShenandoahHeap::non_conc_update_with_forwarded(T* p) {
112 T o = RawAccess<>::oop_load(p);
113 if (!CompressedOops::is_null(o)) {
114 oop obj = CompressedOops::decode_not_null(o);
115 if (in_collection_set(obj)) {
116 // Corner case: when evacuation fails, there are objects in collection
117 // set that are not really forwarded. We can still go and try and update them
118 // (uselessly) to simplify the common path.
119 shenandoah_assert_forwarded_except(p, obj, cancelled_gc());
120 oop fwd = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
121 shenandoah_assert_not_in_cset_except(p, fwd, cancelled_gc());
122
123 // Unconditionally store the update: no concurrent updates expected.
124 RawAccess<IS_NOT_NULL>::oop_store(p, fwd);
125 }
126 }
127 }
128
129 template <class T>
130 inline void ShenandoahHeap::conc_update_with_forwarded(T* p) {
131 T o = RawAccess<>::oop_load(p);
132 if (!CompressedOops::is_null(o)) {
133 oop obj = CompressedOops::decode_not_null(o);
134 if (in_collection_set(obj)) {
135 // Corner case: when evacuation fails, there are objects in collection
136 // set that are not really forwarded. We can still go and try CAS-update them
137 // (uselessly) to simplify the common path.
138 shenandoah_assert_forwarded_except(p, obj, cancelled_gc());
139 oop fwd = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
140 shenandoah_assert_not_in_cset_except(p, fwd, cancelled_gc());
141
142 // Sanity check: we should not be updating the cset regions themselves,
143 // unless we are recovering from the evacuation failure.
144 shenandoah_assert_not_in_cset_loc_except(p, !is_in(p) || cancelled_gc());
145
146 // Either we succeed in updating the reference, or something else gets in our way.
147 // We don't care if that is another concurrent GC update, or another mutator update.
148 atomic_update_oop(fwd, p, obj);
149 }
150 }
151 }
152
153 // Atomic updates of heap location. This is only expected to work with updating the same
154 // logical object with its forwardee. The reason why we need stronger-than-relaxed memory
155 // ordering has to do with coordination with GC barriers and mutator accesses.
156 //
157 // In essence, stronger CAS access is required to maintain the transitive chains that mutator
158 // accesses build by themselves. To illustrate this point, consider the following example.
159 //
160 // Suppose "o" is the object that has a field "x" and the reference to "o" is stored
161 // to field at "addr", which happens to be Java volatile field. Normally, the accesses to volatile
162 // field at "addr" would be matched with release/acquire barriers. This changes when GC moves
163 // the object under mutator feet.
164 //
165 // Thread 1 (Java)
166 // // --- previous access starts here
167 // ...
168 // T1.1: store(&o.x, 1, mo_relaxed)
169 // T1.2: store(&addr, o, mo_release) // volatile store
170 //
171 // // --- new access starts here
172 // // LRB: copy and install the new copy to fwdptr
173 // T1.3: var copy = copy(o)
174 // T1.4: cas(&fwd, t, copy, mo_release) // pointer-mediated publication
175 // <access continues>
176 //
177 // Thread 2 (GC updater)
178 // T2.1: var f = load(&fwd, mo_{consume|acquire}) // pointer-mediated acquisition
179 // T2.2: cas(&addr, o, f, mo_release) // this method
180 //
181 // Thread 3 (Java)
182 // T3.1: var o = load(&addr, mo_acquire) // volatile read
183 // T3.2: if (o != null)
184 // T3.3: var r = load(&o.x, mo_relaxed)
185 //
186 // r is guaranteed to contain "1".
187 //
188 // Without GC involvement, there is synchronizes-with edge from T1.2 to T3.1,
189 // which guarantees this. With GC involvement, when LRB copies the object and
190 // another thread updates the reference to it, we need to have the transitive edge
191 // from T1.4 to T2.1 (that one is guaranteed by forwarding accesses), plus the edge
192 // from T2.2 to T3.1 (which is brought by this CAS).
193 //
194 // Note that we do not need to "acquire" in these methods, because we do not read the
195 // failure witnesses contents on any path, and "release" is enough.
196 //
197
198 inline void ShenandoahHeap::atomic_update_oop(oop update, oop* addr, oop compare) {
199 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr));
200 Atomic::cmpxchg(addr, compare, update, memory_order_release);
201 }
202
203 inline void ShenandoahHeap::atomic_update_oop(oop update, narrowOop* addr, narrowOop compare) {
204 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
205 narrowOop u = CompressedOops::encode(update);
206 Atomic::cmpxchg(addr, compare, u, memory_order_release);
207 }
208
209 inline void ShenandoahHeap::atomic_update_oop(oop update, narrowOop* addr, oop compare) {
210 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
211 narrowOop c = CompressedOops::encode(compare);
212 narrowOop u = CompressedOops::encode(update);
213 Atomic::cmpxchg(addr, c, u, memory_order_release);
214 }
215
216 inline bool ShenandoahHeap::atomic_update_oop_check(oop update, oop* addr, oop compare) {
217 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr));
218 return (oop) Atomic::cmpxchg(addr, compare, update, memory_order_release) == compare;
219 }
220
221 inline bool ShenandoahHeap::atomic_update_oop_check(oop update, narrowOop* addr, narrowOop compare) {
222 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
223 narrowOop u = CompressedOops::encode(update);
224 return (narrowOop) Atomic::cmpxchg(addr, compare, u, memory_order_release) == compare;
225 }
226
227 inline bool ShenandoahHeap::atomic_update_oop_check(oop update, narrowOop* addr, oop compare) {
228 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
229 narrowOop c = CompressedOops::encode(compare);
230 narrowOop u = CompressedOops::encode(update);
231 return CompressedOops::decode(Atomic::cmpxchg(addr, c, u, memory_order_release)) == compare;
232 }
233
234 // The memory ordering discussion above does not apply for methods that store nulls:
235 // then, there is no transitive reads in mutator (as we see nulls), and we can do
236 // relaxed memory ordering there.
237
238 inline void ShenandoahHeap::atomic_clear_oop(oop* addr, oop compare) {
239 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr));
240 Atomic::cmpxchg(addr, compare, oop(), memory_order_relaxed);
241 }
242
243 inline void ShenandoahHeap::atomic_clear_oop(narrowOop* addr, oop compare) {
244 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
245 narrowOop cmp = CompressedOops::encode(compare);
246 Atomic::cmpxchg(addr, cmp, narrowOop(), memory_order_relaxed);
247 }
248
249 inline void ShenandoahHeap::atomic_clear_oop(narrowOop* addr, narrowOop compare) {
250 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
251 Atomic::cmpxchg(addr, compare, narrowOop(), memory_order_relaxed);
252 }
253
254 inline bool ShenandoahHeap::cancelled_gc() const {
255 return _cancelled_gc.get() == CANCELLED;
256 }
257
258 inline bool ShenandoahHeap::check_cancelled_gc_and_yield(bool sts_active) {
259 if (sts_active && !cancelled_gc()) {
260 if (SuspendibleThreadSet::should_yield()) {
261 SuspendibleThreadSet::yield();
262 }
263 }
264 return cancelled_gc();
265 }
266
267 inline void ShenandoahHeap::clear_cancelled_gc() {
268 _cancelled_gc.set(CANCELLABLE);
269 _oom_evac_handler.clear();
270 }
271
272 inline HeapWord* ShenandoahHeap::allocate_from_gclab(Thread* thread, size_t size) {
273 assert(UseTLAB, "TLABs should be enabled");
274
275 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
276 if (gclab == nullptr) {
277 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(),
278 "Performance: thread should have GCLAB: %s", thread->name());
279 // No GCLABs in this thread, fallback to shared allocation
280 return nullptr;
281 }
282 HeapWord* obj = gclab->allocate(size);
283 if (obj != nullptr) {
284 return obj;
285 }
286 // Otherwise...
287 return allocate_from_gclab_slow(thread, size);
288 }
289
290 inline bool ShenandoahHeap::requires_marking(const void* entry) const {
291 oop obj = cast_to_oop(entry);
292 return !_marking_context->is_marked_strong(obj);
293 }
294
295 inline bool ShenandoahHeap::in_collection_set(oop p) const {
296 assert(collection_set() != nullptr, "Sanity");
297 return collection_set()->is_in(p);
298 }
299
300 inline bool ShenandoahHeap::in_collection_set_loc(void* p) const {
301 assert(collection_set() != nullptr, "Sanity");
302 return collection_set()->is_in_loc(p);
303 }
304
305 inline bool ShenandoahHeap::is_stable() const {
306 return _gc_state.is_clear();
307 }
308
309 inline bool ShenandoahHeap::is_idle() const {
310 return _gc_state.is_unset(MARKING | EVACUATION | UPDATEREFS);
311 }
312
313 inline bool ShenandoahHeap::is_concurrent_mark_in_progress() const {
314 return _gc_state.is_set(MARKING);
315 }
316
317 inline bool ShenandoahHeap::is_evacuation_in_progress() const {
318 return _gc_state.is_set(EVACUATION);
319 }
320
321 inline bool ShenandoahHeap::is_degenerated_gc_in_progress() const {
322 return _degenerated_gc_in_progress.is_set();
323 }
324
325 inline bool ShenandoahHeap::is_full_gc_in_progress() const {
326 return _full_gc_in_progress.is_set();
327 }
328
329 inline bool ShenandoahHeap::is_full_gc_move_in_progress() const {
330 return _full_gc_move_in_progress.is_set();
331 }
332
333 inline bool ShenandoahHeap::is_update_refs_in_progress() const {
334 return _gc_state.is_set(UPDATEREFS);
335 }
336
337 inline bool ShenandoahHeap::is_stw_gc_in_progress() const {
338 return is_full_gc_in_progress() || is_degenerated_gc_in_progress();
339 }
340
341 inline bool ShenandoahHeap::is_concurrent_strong_root_in_progress() const {
342 return _concurrent_strong_root_in_progress.is_set();
343 }
344
345 inline bool ShenandoahHeap::is_concurrent_weak_root_in_progress() const {
346 return _gc_state.is_set(WEAK_ROOTS);
347 }
348
349 template<class T>
350 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl) {
351 marked_object_iterate(region, cl, region->top());
352 }
353
354 template<class T>
355 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* limit) {
356 assert(! region->is_humongous_continuation(), "no humongous continuation regions here");
357
358 ShenandoahMarkingContext* const ctx = complete_marking_context();
359 assert(ctx->is_complete(), "sanity");
360
361 HeapWord* tams = ctx->top_at_mark_start(region);
362
363 size_t skip_bitmap_delta = 1;
364 HeapWord* start = region->bottom();
365 HeapWord* end = MIN2(tams, region->end());
366
367 // Step 1. Scan below the TAMS based on bitmap data.
368 HeapWord* limit_bitmap = MIN2(limit, tams);
369
370 // Try to scan the initial candidate. If the candidate is above the TAMS, it would
371 // fail the subsequent "< limit_bitmap" checks, and fall through to Step 2.
372 HeapWord* cb = ctx->get_next_marked_addr(start, end);
373
374 intx dist = ShenandoahMarkScanPrefetch;
375 if (dist > 0) {
376 // Batched scan that prefetches the oop data, anticipating the access to
377 // either header, oop field, or forwarding pointer. Not that we cannot
378 // touch anything in oop, while it still being prefetched to get enough
379 // time for prefetch to work. This is why we try to scan the bitmap linearly,
380 // disregarding the object size. However, since we know forwarding pointer
381 // precedes the object, we can skip over it. Once we cannot trust the bitmap,
382 // there is no point for prefetching the oop contents, as oop->size() will
383 // touch it prematurely.
384
385 // No variable-length arrays in standard C++, have enough slots to fit
386 // the prefetch distance.
387 static const int SLOT_COUNT = 256;
388 guarantee(dist <= SLOT_COUNT, "adjust slot count");
389 HeapWord* slots[SLOT_COUNT];
390
391 int avail;
392 do {
393 avail = 0;
394 for (int c = 0; (c < dist) && (cb < limit_bitmap); c++) {
395 Prefetch::read(cb, oopDesc::mark_offset_in_bytes());
396 slots[avail++] = cb;
397 cb += skip_bitmap_delta;
398 if (cb < limit_bitmap) {
399 cb = ctx->get_next_marked_addr(cb, limit_bitmap);
400 }
401 }
402
403 for (int c = 0; c < avail; c++) {
404 assert (slots[c] < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(tams));
405 assert (slots[c] < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(limit));
406 oop obj = cast_to_oop(slots[c]);
407 assert(oopDesc::is_oop(obj), "sanity");
408 assert(ctx->is_marked(obj), "object expected to be marked");
409 cl->do_object(obj);
410 }
411 } while (avail > 0);
412 } else {
413 while (cb < limit_bitmap) {
414 assert (cb < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(tams));
415 assert (cb < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(limit));
416 oop obj = cast_to_oop(cb);
417 assert(oopDesc::is_oop(obj), "sanity");
418 assert(ctx->is_marked(obj), "object expected to be marked");
419 cl->do_object(obj);
420 cb += skip_bitmap_delta;
421 if (cb < limit_bitmap) {
422 cb = ctx->get_next_marked_addr(cb, limit_bitmap);
423 }
424 }
425 }
426
427 // Step 2. Accurate size-based traversal, happens past the TAMS.
428 // This restarts the scan at TAMS, which makes sure we traverse all objects,
429 // regardless of what happened at Step 1.
430 HeapWord* cs = tams;
431 while (cs < limit) {
432 assert (cs >= tams, "only objects past TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(tams));
433 assert (cs < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(limit));
434 oop obj = cast_to_oop(cs);
435 assert(oopDesc::is_oop(obj), "sanity");
436 assert(ctx->is_marked(obj), "object expected to be marked");
437 size_t size = ShenandoahForwarding::size(obj);
438 cl->do_object(obj);
439 cs += size;
440 }
441 }
442
443 template <class T>
444 class ShenandoahObjectToOopClosure : public ObjectClosure {
445 T* _cl;
446 public:
447 ShenandoahObjectToOopClosure(T* cl) : _cl(cl) {}
448
449 void do_object(oop obj) {
450 obj->oop_iterate(_cl);
451 }
452 };
453
454 template <class T>
455 class ShenandoahObjectToOopBoundedClosure : public ObjectClosure {
456 T* _cl;
457 MemRegion _bounds;
458 public:
459 ShenandoahObjectToOopBoundedClosure(T* cl, HeapWord* bottom, HeapWord* top) :
460 _cl(cl), _bounds(bottom, top) {}
461
462 void do_object(oop obj) {
463 obj->oop_iterate(_cl, _bounds);
464 }
465 };
466
467 template<class T>
468 inline void ShenandoahHeap::marked_object_oop_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* top) {
469 if (region->is_humongous()) {
470 HeapWord* bottom = region->bottom();
471 if (top > bottom) {
472 region = region->humongous_start_region();
473 ShenandoahObjectToOopBoundedClosure<T> objs(cl, bottom, top);
474 marked_object_iterate(region, &objs);
475 }
476 } else {
477 ShenandoahObjectToOopClosure<T> objs(cl);
478 marked_object_iterate(region, &objs, top);
479 }
480 }
481
482 inline ShenandoahHeapRegion* ShenandoahHeap::get_region(size_t region_idx) const {
483 if (region_idx < _num_regions) {
484 return _regions[region_idx];
485 } else {
486 return nullptr;
487 }
488 }
489
490 inline void ShenandoahHeap::mark_complete_marking_context() {
491 _marking_context->mark_complete();
492 }
493
494 inline void ShenandoahHeap::mark_incomplete_marking_context() {
495 _marking_context->mark_incomplete();
496 }
497
498 inline ShenandoahMarkingContext* ShenandoahHeap::complete_marking_context() const {
499 assert (_marking_context->is_complete()," sanity");
500 return _marking_context;
501 }
502
503 inline ShenandoahMarkingContext* ShenandoahHeap::marking_context() const {
504 return _marking_context;
505 }
506
507 #endif // SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP