1 /*
2 * Copyright (c) 2015, 2019, Red Hat, Inc. All rights reserved.
3 *
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation.
7 *
8 * This code is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11 * version 2 for more details (a copy is included in the LICENSE file that
12 * accompanied this code).
13 *
14 * You should have received a copy of the GNU General Public License version
15 * 2 along with this work; if not, write to the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19 * or visit www.oracle.com if you need additional information or have any
20 * questions.
21 *
22 */
23
24 #ifndef SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
25 #define SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
26
27 #include "classfile/javaClasses.inline.hpp"
28 #include "gc/shared/markBitMap.inline.hpp"
29 #include "gc/shared/threadLocalAllocBuffer.inline.hpp"
30 #include "gc/shared/suspendibleThreadSet.hpp"
31 #include "gc/shenandoah/shenandoahAsserts.hpp"
32 #include "gc/shenandoah/shenandoahBarrierSet.inline.hpp"
33 #include "gc/shenandoah/shenandoahCollectionSet.inline.hpp"
34 #include "gc/shenandoah/shenandoahForwarding.inline.hpp"
35 #include "gc/shenandoah/shenandoahWorkGroup.hpp"
36 #include "gc/shenandoah/shenandoahHeap.hpp"
37 #include "gc/shenandoah/shenandoahHeapRegionSet.inline.hpp"
38 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
39 #include "gc/shenandoah/shenandoahControlThread.hpp"
40 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
41 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
42 #include "oops/compressedOops.inline.hpp"
43 #include "oops/oop.inline.hpp"
44 #include "runtime/atomic.hpp"
45 #include "runtime/prefetch.inline.hpp"
46 #include "runtime/thread.hpp"
47 #include "utilities/copy.hpp"
48 #include "utilities/globalDefinitions.hpp"
49
50
51 inline ShenandoahHeapRegion* ShenandoahRegionIterator::next() {
52 size_t new_index = Atomic::add(&_index, (size_t) 1);
53 // get_region() provides the bounds-check and returns NULL on OOB.
54 return _heap->get_region(new_index - 1);
55 }
56
57 inline bool ShenandoahHeap::has_forwarded_objects() const {
58 return _gc_state.is_set(HAS_FORWARDED);
59 }
60
61 inline WorkGang* ShenandoahHeap::workers() const {
62 return _workers;
63 }
64
65 inline WorkGang* ShenandoahHeap::get_safepoint_workers() {
66 return _safepoint_workers;
67 }
68
69 inline size_t ShenandoahHeap::heap_region_index_containing(const void* addr) const {
70 uintptr_t region_start = ((uintptr_t) addr);
71 uintptr_t index = (region_start - (uintptr_t) base()) >> ShenandoahHeapRegion::region_size_bytes_shift();
72 assert(index < num_regions(), "Region index is in bounds: " PTR_FORMAT, p2i(addr));
73 return index;
74 }
75
76 inline ShenandoahHeapRegion* const ShenandoahHeap::heap_region_containing(const void* addr) const {
77 size_t index = heap_region_index_containing(addr);
78 ShenandoahHeapRegion* const result = get_region(index);
79 assert(addr >= result->bottom() && addr < result->end(), "Heap region contains the address: " PTR_FORMAT, p2i(addr));
80 return result;
81 }
82
83 template <class T>
84 inline oop ShenandoahHeap::update_with_forwarded_not_null(T* p, oop obj) {
85 if (in_collection_set(obj)) {
86 shenandoah_assert_forwarded_except(p, obj, is_full_gc_in_progress() || cancelled_gc() || is_degenerated_gc_in_progress());
87 obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
88 RawAccess<IS_NOT_NULL>::oop_store(p, obj);
89 }
90 #ifdef ASSERT
91 else {
92 shenandoah_assert_not_forwarded(p, obj);
93 }
94 #endif
95 return obj;
96 }
97
98 template <class T>
99 inline oop ShenandoahHeap::maybe_update_with_forwarded(T* p) {
100 T o = RawAccess<>::oop_load(p);
101 if (!CompressedOops::is_null(o)) {
102 oop obj = CompressedOops::decode_not_null(o);
103 return maybe_update_with_forwarded_not_null(p, obj);
104 } else {
105 return NULL;
106 }
107 }
108
109 template <class T>
110 inline oop ShenandoahHeap::evac_update_with_forwarded(T* p) {
111 T o = RawAccess<>::oop_load(p);
112 if (!CompressedOops::is_null(o)) {
113 oop heap_oop = CompressedOops::decode_not_null(o);
114 if (in_collection_set(heap_oop)) {
115 oop forwarded_oop = ShenandoahBarrierSet::resolve_forwarded_not_null(heap_oop);
116 if (forwarded_oop == heap_oop) {
117 forwarded_oop = evacuate_object(heap_oop, Thread::current());
118 }
119 oop prev = cas_oop(forwarded_oop, p, heap_oop);
120 if (prev == heap_oop) {
121 return forwarded_oop;
122 } else {
123 return NULL;
124 }
125 }
126 return heap_oop;
127 } else {
128 return NULL;
129 }
130 }
131
132 inline oop ShenandoahHeap::cas_oop(oop n, oop* addr, oop c) {
133 assert(is_aligned(addr, HeapWordSize), "Address should be aligned: " PTR_FORMAT, p2i(addr));
134 return (oop) Atomic::cmpxchg(addr, c, n);
135 }
136
137 inline oop ShenandoahHeap::cas_oop(oop n, narrowOop* addr, narrowOop c) {
138 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
139 narrowOop val = CompressedOops::encode(n);
140 return CompressedOops::decode((narrowOop) Atomic::cmpxchg(addr, c, val));
141 }
142
143 inline oop ShenandoahHeap::cas_oop(oop n, narrowOop* addr, oop c) {
144 assert(is_aligned(addr, sizeof(narrowOop)), "Address should be aligned: " PTR_FORMAT, p2i(addr));
145 narrowOop cmp = CompressedOops::encode(c);
146 narrowOop val = CompressedOops::encode(n);
147 return CompressedOops::decode((narrowOop) Atomic::cmpxchg(addr, cmp, val));
148 }
149
150 template <class T>
151 inline oop ShenandoahHeap::maybe_update_with_forwarded_not_null(T* p, oop heap_oop) {
152 shenandoah_assert_not_in_cset_loc_except(p, !is_in(p) || is_full_gc_in_progress() || is_degenerated_gc_in_progress());
153 shenandoah_assert_correct(p, heap_oop);
154
155 if (in_collection_set(heap_oop)) {
156 oop forwarded_oop = ShenandoahBarrierSet::resolve_forwarded_not_null(heap_oop);
157 if (forwarded_oop == heap_oop) {
158 // E.g. during evacuation.
159 return forwarded_oop;
160 }
161
162 shenandoah_assert_forwarded_except(p, heap_oop, is_full_gc_in_progress() || is_degenerated_gc_in_progress());
163 shenandoah_assert_not_forwarded(p, forwarded_oop);
164 shenandoah_assert_not_in_cset_except(p, forwarded_oop, cancelled_gc());
165
166 // If this fails, another thread wrote to p before us, it will be logged in SATB and the
167 // reference be updated later.
168 oop witness = cas_oop(forwarded_oop, p, heap_oop);
169
170 if (witness != heap_oop) {
171 // CAS failed, someone had beat us to it. Normally, we would return the failure witness,
172 // because that would be the proper write of to-space object, enforced by strong barriers.
173 // However, there is a corner case with arraycopy. It can happen that a Java thread
174 // beats us with an arraycopy, which first copies the array, which potentially contains
175 // from-space refs, and only afterwards updates all from-space refs to to-space refs,
176 // which leaves a short window where the new array elements can be from-space.
177 // In this case, we can just resolve the result again. As we resolve, we need to consider
178 // the contended write might have been NULL.
179 oop result = witness;
180 if (!CompressedOops::is_null(witness) && in_collection_set(witness)) {
181 result = ShenandoahBarrierSet::resolve_forwarded(witness);
182 }
183 shenandoah_assert_not_forwarded_except(p, result, (result == NULL));
184 shenandoah_assert_not_in_cset_except(p, result, (result == NULL) || cancelled_gc());
185 return result;
186 } else {
187 // Success! We have updated with known to-space copy. We have already asserted it is sane.
188 return forwarded_oop;
189 }
190 } else {
191 shenandoah_assert_not_forwarded(p, heap_oop);
192 return heap_oop;
193 }
194 }
195
196 inline bool ShenandoahHeap::cancelled_gc() const {
197 return _cancelled_gc.get() == CANCELLED;
198 }
199
200 inline bool ShenandoahHeap::check_cancelled_gc_and_yield(bool sts_active) {
201 if (! (sts_active && ShenandoahSuspendibleWorkers)) {
202 return cancelled_gc();
203 }
204
205 jbyte prev = _cancelled_gc.cmpxchg(NOT_CANCELLED, CANCELLABLE);
206 if (prev == CANCELLABLE || prev == NOT_CANCELLED) {
207 if (SuspendibleThreadSet::should_yield()) {
208 SuspendibleThreadSet::yield();
209 }
210
211 // Back to CANCELLABLE. The thread that poked NOT_CANCELLED first gets
212 // to restore to CANCELLABLE.
213 if (prev == CANCELLABLE) {
214 _cancelled_gc.set(CANCELLABLE);
215 }
216 return false;
217 } else {
218 return true;
219 }
220 }
221
222 inline void ShenandoahHeap::clear_cancelled_gc() {
223 _cancelled_gc.set(CANCELLABLE);
224 _oom_evac_handler.clear();
225 }
226
227 inline HeapWord* ShenandoahHeap::allocate_from_gclab(Thread* thread, size_t size) {
228 assert(UseTLAB, "TLABs should be enabled");
229
230 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
231 if (gclab == NULL) {
232 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(),
233 "Performance: thread should have GCLAB: %s", thread->name());
234 // No GCLABs in this thread, fallback to shared allocation
235 return NULL;
236 }
237 HeapWord* obj = gclab->allocate(size);
238 if (obj != NULL) {
239 return obj;
240 }
241 // Otherwise...
242 return allocate_from_gclab_slow(thread, size);
243 }
244
245 inline oop ShenandoahHeap::evacuate_object(oop p, Thread* thread) {
246 if (ShenandoahThreadLocalData::is_oom_during_evac(Thread::current())) {
247 // This thread went through the OOM during evac protocol and it is safe to return
248 // the forward pointer. It must not attempt to evacuate any more.
249 return ShenandoahBarrierSet::resolve_forwarded(p);
250 }
251
252 assert(ShenandoahThreadLocalData::is_evac_allowed(thread), "must be enclosed in oom-evac scope");
253
254 size_t size = p->size();
255
256 assert(!heap_region_containing(p)->is_humongous(), "never evacuate humongous objects");
257
258 bool alloc_from_gclab = true;
259 HeapWord* copy = NULL;
260
261 #ifdef ASSERT
262 if (ShenandoahOOMDuringEvacALot &&
263 (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call
264 copy = NULL;
265 } else {
266 #endif
267 if (UseTLAB) {
268 copy = allocate_from_gclab(thread, size);
269 }
270 if (copy == NULL) {
271 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size);
272 copy = allocate_memory(req);
273 alloc_from_gclab = false;
274 }
275 #ifdef ASSERT
276 }
277 #endif
278
279 if (copy == NULL) {
280 control_thread()->handle_alloc_failure_evac(size);
281
282 _oom_evac_handler.handle_out_of_memory_during_evacuation();
283
284 return ShenandoahBarrierSet::resolve_forwarded(p);
285 }
286
287 // Copy the object:
288 Copy::aligned_disjoint_words((HeapWord*) p, copy, size);
289
290 // Try to install the new forwarding pointer.
291 oop copy_val = oop(copy);
292 oop result = ShenandoahForwarding::try_update_forwardee(p, copy_val);
293 if (result == copy_val) {
294 // Successfully evacuated. Our copy is now the public one!
295 shenandoah_assert_correct(NULL, copy_val);
296 return copy_val;
297 } else {
298 // Failed to evacuate. We need to deal with the object that is left behind. Since this
299 // new allocation is certainly after TAMS, it will be considered live in the next cycle.
300 // But if it happens to contain references to evacuated regions, those references would
301 // not get updated for this stale copy during this cycle, and we will crash while scanning
302 // it the next cycle.
303 //
304 // For GCLAB allocations, it is enough to rollback the allocation ptr. Either the next
305 // object will overwrite this stale copy, or the filler object on LAB retirement will
306 // do this. For non-GCLAB allocations, we have no way to retract the allocation, and
307 // have to explicitly overwrite the copy with the filler object. With that overwrite,
308 // we have to keep the fwdptr initialized and pointing to our (stale) copy.
309 if (alloc_from_gclab) {
310 ShenandoahThreadLocalData::gclab(thread)->undo_allocation(copy, size);
311 } else {
312 fill_with_object(copy, size);
313 shenandoah_assert_correct(NULL, copy_val);
314 }
315 shenandoah_assert_correct(NULL, result);
316 return result;
317 }
318 }
319
320 template<bool RESOLVE>
321 inline bool ShenandoahHeap::requires_marking(const void* entry) const {
322 oop obj = oop(entry);
323 if (RESOLVE && in_collection_set(obj)) {
324 obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
325 }
326 return !_marking_context->is_marked(obj);
327 }
328
329 template <class T>
330 inline bool ShenandoahHeap::in_collection_set(T p) const {
331 HeapWord* obj = (HeapWord*) p;
332 assert(collection_set() != NULL, "Sanity");
333 assert(is_in(obj), "should be in heap");
334
335 return collection_set()->is_in(obj);
336 }
337
338 inline bool ShenandoahHeap::is_stable() const {
339 return _gc_state.is_clear();
340 }
341
342 inline bool ShenandoahHeap::is_idle() const {
343 return _gc_state.is_unset(MARKING | EVACUATION | UPDATEREFS | TRAVERSAL);
344 }
345
346 inline bool ShenandoahHeap::is_concurrent_mark_in_progress() const {
347 return _gc_state.is_set(MARKING);
348 }
349
350 inline bool ShenandoahHeap::is_concurrent_traversal_in_progress() const {
351 return _gc_state.is_set(TRAVERSAL);
352 }
353
354 inline bool ShenandoahHeap::is_evacuation_in_progress() const {
355 return _gc_state.is_set(EVACUATION);
356 }
357
358 inline bool ShenandoahHeap::is_gc_in_progress_mask(uint mask) const {
359 return _gc_state.is_set(mask);
360 }
361
362 inline bool ShenandoahHeap::is_degenerated_gc_in_progress() const {
363 return _degenerated_gc_in_progress.is_set();
364 }
365
366 inline bool ShenandoahHeap::is_full_gc_in_progress() const {
367 return _full_gc_in_progress.is_set();
368 }
369
370 inline bool ShenandoahHeap::is_full_gc_move_in_progress() const {
371 return _full_gc_move_in_progress.is_set();
372 }
373
374 inline bool ShenandoahHeap::is_update_refs_in_progress() const {
375 return _gc_state.is_set(UPDATEREFS);
376 }
377
378 inline bool ShenandoahHeap::is_stw_gc_in_progress() const {
379 return is_full_gc_in_progress() || is_degenerated_gc_in_progress();
380 }
381
382 inline bool ShenandoahHeap::is_concurrent_root_in_progress() const {
383 return _concurrent_root_in_progress.is_set();
384 }
385
386 template<class T>
387 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl) {
388 marked_object_iterate(region, cl, region->top());
389 }
390
391 template<class T>
392 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* limit) {
393 assert(! region->is_humongous_continuation(), "no humongous continuation regions here");
394
395 ShenandoahMarkingContext* const ctx = complete_marking_context();
396 assert(ctx->is_complete(), "sanity");
397
398 MarkBitMap* mark_bit_map = ctx->mark_bit_map();
399 HeapWord* tams = ctx->top_at_mark_start(region);
400
401 size_t skip_bitmap_delta = 1;
402 HeapWord* start = region->bottom();
403 HeapWord* end = MIN2(tams, region->end());
404
405 // Step 1. Scan below the TAMS based on bitmap data.
406 HeapWord* limit_bitmap = MIN2(limit, tams);
407
408 // Try to scan the initial candidate. If the candidate is above the TAMS, it would
409 // fail the subsequent "< limit_bitmap" checks, and fall through to Step 2.
410 HeapWord* cb = mark_bit_map->get_next_marked_addr(start, end);
411
412 intx dist = ShenandoahMarkScanPrefetch;
413 if (dist > 0) {
414 // Batched scan that prefetches the oop data, anticipating the access to
415 // either header, oop field, or forwarding pointer. Not that we cannot
416 // touch anything in oop, while it still being prefetched to get enough
417 // time for prefetch to work. This is why we try to scan the bitmap linearly,
418 // disregarding the object size. However, since we know forwarding pointer
419 // preceeds the object, we can skip over it. Once we cannot trust the bitmap,
420 // there is no point for prefetching the oop contents, as oop->size() will
421 // touch it prematurely.
422
423 // No variable-length arrays in standard C++, have enough slots to fit
424 // the prefetch distance.
425 static const int SLOT_COUNT = 256;
426 guarantee(dist <= SLOT_COUNT, "adjust slot count");
427 HeapWord* slots[SLOT_COUNT];
428
429 int avail;
430 do {
431 avail = 0;
432 for (int c = 0; (c < dist) && (cb < limit_bitmap); c++) {
433 Prefetch::read(cb, oopDesc::mark_offset_in_bytes());
434 slots[avail++] = cb;
435 cb += skip_bitmap_delta;
436 if (cb < limit_bitmap) {
437 cb = mark_bit_map->get_next_marked_addr(cb, limit_bitmap);
438 }
439 }
440
441 for (int c = 0; c < avail; c++) {
442 assert (slots[c] < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(tams));
443 assert (slots[c] < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(limit));
444 oop obj = oop(slots[c]);
445 assert(oopDesc::is_oop(obj), "sanity");
446 assert(ctx->is_marked(obj), "object expected to be marked");
447 cl->do_object(obj);
448 }
449 } while (avail > 0);
450 } else {
451 while (cb < limit_bitmap) {
452 assert (cb < tams, "only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(tams));
453 assert (cb < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(limit));
454 oop obj = oop(cb);
455 assert(oopDesc::is_oop(obj), "sanity");
456 assert(ctx->is_marked(obj), "object expected to be marked");
457 cl->do_object(obj);
458 cb += skip_bitmap_delta;
459 if (cb < limit_bitmap) {
460 cb = mark_bit_map->get_next_marked_addr(cb, limit_bitmap);
461 }
462 }
463 }
464
465 // Step 2. Accurate size-based traversal, happens past the TAMS.
466 // This restarts the scan at TAMS, which makes sure we traverse all objects,
467 // regardless of what happened at Step 1.
468 HeapWord* cs = tams;
469 while (cs < limit) {
470 assert (cs >= tams, "only objects past TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(tams));
471 assert (cs < limit, "only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(limit));
472 oop obj = oop(cs);
473 assert(oopDesc::is_oop(obj), "sanity");
474 assert(ctx->is_marked(obj), "object expected to be marked");
475 int size = obj->size();
476 cl->do_object(obj);
477 cs += size;
478 }
479 }
480
481 template <class T>
482 class ShenandoahObjectToOopClosure : public ObjectClosure {
483 T* _cl;
484 public:
485 ShenandoahObjectToOopClosure(T* cl) : _cl(cl) {}
486
487 void do_object(oop obj) {
488 obj->oop_iterate(_cl);
489 }
490 };
491
492 template <class T>
493 class ShenandoahObjectToOopBoundedClosure : public ObjectClosure {
494 T* _cl;
495 MemRegion _bounds;
496 public:
497 ShenandoahObjectToOopBoundedClosure(T* cl, HeapWord* bottom, HeapWord* top) :
498 _cl(cl), _bounds(bottom, top) {}
499
500 void do_object(oop obj) {
501 obj->oop_iterate(_cl, _bounds);
502 }
503 };
504
505 template<class T>
506 inline void ShenandoahHeap::marked_object_oop_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* top) {
507 if (region->is_humongous()) {
508 HeapWord* bottom = region->bottom();
509 if (top > bottom) {
510 region = region->humongous_start_region();
511 ShenandoahObjectToOopBoundedClosure<T> objs(cl, bottom, top);
512 marked_object_iterate(region, &objs);
513 }
514 } else {
515 ShenandoahObjectToOopClosure<T> objs(cl);
516 marked_object_iterate(region, &objs, top);
517 }
518 }
519
520 inline ShenandoahHeapRegion* const ShenandoahHeap::get_region(size_t region_idx) const {
521 if (region_idx < _num_regions) {
522 return _regions[region_idx];
523 } else {
524 return NULL;
525 }
526 }
527
528 inline void ShenandoahHeap::mark_complete_marking_context() {
529 _marking_context->mark_complete();
530 }
531
532 inline void ShenandoahHeap::mark_incomplete_marking_context() {
533 _marking_context->mark_incomplete();
534 }
535
536 inline ShenandoahMarkingContext* ShenandoahHeap::complete_marking_context() const {
537 assert (_marking_context->is_complete()," sanity");
538 return _marking_context;
539 }
540
541 inline ShenandoahMarkingContext* ShenandoahHeap::marking_context() const {
542 return _marking_context;
543 }
544
545 #endif // SHARE_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP