1 /*
2 * Copyright (c) 2015, Red Hat, Inc. and/or its affiliates.
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_VM_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
25 #define SHARE_VM_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP
26
27 #include "gc_implementation/shared/markBitMap.inline.hpp"
28 #include "memory/threadLocalAllocBuffer.inline.hpp"
29 #include "gc_implementation/shenandoah/brooksPointer.inline.hpp"
30 #include "gc_implementation/shenandoah/shenandoahAsserts.hpp"
31 #include "gc_implementation/shenandoah/shenandoahBarrierSet.inline.hpp"
32 #include "gc_implementation/shenandoah/shenandoahCollectionSet.hpp"
33 #include "gc_implementation/shenandoah/shenandoahCollectionSet.inline.hpp"
34 #include "gc_implementation/shenandoah/shenandoahControlThread.hpp"
35 #include "gc_implementation/shenandoah/shenandoahMarkingContext.inline.hpp"
36 #include "gc_implementation/shenandoah/shenandoahHeap.hpp"
37 #include "gc_implementation/shenandoah/shenandoahHeapRegionSet.hpp"
38 #include "gc_implementation/shenandoah/shenandoahHeapRegion.inline.hpp"
39 #include "gc_implementation/shenandoah/shenandoahUtils.hpp"
40 #include "oops/oop.inline.hpp"
41 #include "runtime/atomic.hpp"
42 #include "runtime/prefetch.hpp"
43 #include "runtime/prefetch.inline.hpp"
44 #include "utilities/copy.hpp"
45 #include "utilities/globalDefinitions.hpp"
46
47 template <class T>
48 void ShenandoahUpdateRefsClosure::do_oop_work(T* p) {
49 T o = oopDesc::load_heap_oop(p);
50 if (! oopDesc::is_null(o)) {
51 oop obj = oopDesc::decode_heap_oop_not_null(o);
52 _heap->update_with_forwarded_not_null(p, obj);
53 }
54 }
55
56 void ShenandoahUpdateRefsClosure::do_oop(oop* p) { do_oop_work(p); }
57 void ShenandoahUpdateRefsClosure::do_oop(narrowOop* p) { do_oop_work(p); }
58
59 inline ShenandoahHeapRegion* ShenandoahRegionIterator::next() {
60 size_t new_index = Atomic::add((size_t) 1, &_index);
61 // get_region() provides the bounds-check and returns NULL on OOB.
62 return _heap->get_region(new_index - 1);
63 }
64
65 inline bool ShenandoahHeap::has_forwarded_objects() const {
66 return _gc_state.is_set(HAS_FORWARDED);
67 }
68
69 inline ShenandoahWorkGang* ShenandoahHeap::workers() const {
70 return _workers;
71 }
72
73 inline size_t ShenandoahHeap::heap_region_index_containing(const void* addr) const {
74 uintptr_t region_start = ((uintptr_t) addr);
75 uintptr_t index = (region_start - (uintptr_t) base()) >> ShenandoahHeapRegion::region_size_bytes_shift();
76 assert(index < num_regions(), err_msg("Region index is in bounds: " PTR_FORMAT, p2i(addr)));
77 return index;
78 }
79
80 inline ShenandoahHeapRegion* const ShenandoahHeap::heap_region_containing(const void* addr) const {
81 size_t index = heap_region_index_containing(addr);
82 ShenandoahHeapRegion* const result = get_region(index);
83 assert(addr >= result->bottom() && addr < result->end(), err_msg("Heap region contains the address: " PTR_FORMAT, p2i(addr)));
84 return result;
85 }
86
87 template <class T>
88 inline oop ShenandoahHeap::update_with_forwarded_not_null(T* p, oop obj) {
89 if (in_collection_set(obj)) {
90 shenandoah_assert_forwarded_except(p, obj, is_full_gc_in_progress() || cancelled_gc() || is_degenerated_gc_in_progress());
91 obj = ShenandoahBarrierSet::resolve_forwarded_not_null(obj);
92 oopDesc::encode_store_heap_oop(p, obj);
93 }
94 #ifdef ASSERT
95 else {
96 shenandoah_assert_not_forwarded(p, obj);
97 }
98 #endif
99 return obj;
100 }
101
102 template <class T>
103 inline oop ShenandoahHeap::maybe_update_with_forwarded(T* p) {
104 T o = oopDesc::load_heap_oop(p);
105 if (! oopDesc::is_null(o)) {
106 oop obj = oopDesc::decode_heap_oop_not_null(o);
107 return maybe_update_with_forwarded_not_null(p, obj);
108 } else {
109 return NULL;
110 }
111 }
112
113 inline oop ShenandoahHeap::atomic_compare_exchange_oop(oop n, oop* addr, oop c) {
114 return (oop) Atomic::cmpxchg_ptr(n, addr, c);
115 }
116
117 inline oop ShenandoahHeap::atomic_compare_exchange_oop(oop n, narrowOop* addr, oop c) {
118 narrowOop cmp = oopDesc::encode_heap_oop(c);
119 narrowOop val = oopDesc::encode_heap_oop(n);
120 return oopDesc::decode_heap_oop((narrowOop) Atomic::cmpxchg(val, addr, cmp));
121 }
122
123 template <class T>
124 inline oop ShenandoahHeap::maybe_update_with_forwarded_not_null(T* p, oop heap_oop) {
125 shenandoah_assert_not_in_cset_loc_except(p, !is_in(p) || is_full_gc_in_progress() || is_degenerated_gc_in_progress());
126 shenandoah_assert_correct(p, heap_oop);
127
128 if (in_collection_set(heap_oop)) {
129 oop forwarded_oop = ShenandoahBarrierSet::resolve_forwarded_not_null(heap_oop);
130
131 shenandoah_assert_forwarded_except(p, heap_oop, is_full_gc_in_progress() || is_degenerated_gc_in_progress());
132 shenandoah_assert_not_in_cset_except(p, forwarded_oop, cancelled_gc());
133
134 // If this fails, another thread wrote to p before us, it will be logged in SATB and the
135 // reference be updated later.
136 oop result = atomic_compare_exchange_oop(forwarded_oop, p, heap_oop);
137
138 if (oopDesc::unsafe_equals(result, heap_oop)) { // CAS successful.
139 return forwarded_oop;
140 } else {
141 // Note: we used to assert the following here. This doesn't work because sometimes, during
142 // marking/updating-refs, it can happen that a Java thread beats us with an arraycopy,
143 // which first copies the array, which potentially contains from-space refs, and only afterwards
144 // updates all from-space refs to to-space refs, which leaves a short window where the new array
145 // elements can be from-space.
146 // assert(oopDesc::is_null(result) ||
147 // oopDesc::unsafe_equals(result, ShenandoahBarrierSet::resolve_oop_static_not_null(result)),
148 // "expect not forwarded");
149 return NULL;
150 }
151 } else {
152 shenandoah_assert_not_forwarded(p, heap_oop);
153 return heap_oop;
154 }
155 }
156
157 inline bool ShenandoahHeap::cancelled_gc() const {
158 return _cancelled_gc.is_set();
159 }
160
161 inline bool ShenandoahHeap::try_cancel_gc() {
162 return _cancelled_gc.try_set();
163 }
164
165 inline void ShenandoahHeap::clear_cancelled_gc() {
166 _cancelled_gc.unset();
167 _oom_evac_handler.clear();
168 }
169
170 inline HeapWord* ShenandoahHeap::allocate_from_gclab(Thread* thread, size_t size) {
171 assert(UseTLAB, "TLABs should be enabled");
172
173 if (!thread->gclab().is_initialized()) {
174 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(),
175 err_msg("Performance: thread should have GCLAB: %s", thread->name()));
176 // No GCLABs in this thread, fallback to shared allocation
177 return NULL;
178 }
179 HeapWord *obj = thread->gclab().allocate(size);
180 if (obj != NULL) {
181 return obj;
182 }
183 // Otherwise...
184 return allocate_from_gclab_slow(thread, size);
185 }
186
187 inline oop ShenandoahHeap::evacuate_object(oop p, Thread* thread, bool& evacuated) {
188 evacuated = false;
189
190 if (Thread::current()->is_oom_during_evac()) {
191 // This thread went through the OOM during evac protocol and it is safe to return
192 // the forward pointer. It must not attempt to evacuate any more.
193 return ShenandoahBarrierSet::resolve_forwarded(p);
194 }
195
196 assert(thread->is_evac_allowed(), "must be enclosed in in oom-evac scope");
197
198 size_t size_no_fwdptr = (size_t) p->size();
199 size_t size_with_fwdptr = size_no_fwdptr + BrooksPointer::word_size();
200
201 assert(!heap_region_containing(p)->is_humongous(), "never evacuate humongous objects");
202
203 bool alloc_from_gclab = true;
204 HeapWord* filler = NULL;
205
206 #ifdef ASSERT
207 if (ShenandoahOOMDuringEvacALot &&
208 (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call
209 filler = NULL;
210 } else {
211 #endif
212 if (UseTLAB) {
213 filler = allocate_from_gclab(thread, size_with_fwdptr);
214 }
215 if (filler == NULL) {
216 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size_with_fwdptr);
217 filler = allocate_memory(req);
218 alloc_from_gclab = false;
219 }
220 #ifdef ASSERT
221 }
222 #endif
223
224 if (filler == NULL) {
225 control_thread()->handle_alloc_failure_evac(size_with_fwdptr);
226
227 _oom_evac_handler.handle_out_of_memory_during_evacuation();
228
229 return ShenandoahBarrierSet::resolve_forwarded(p);
230 }
231
232 // Copy the object and initialize its forwarding ptr:
233 HeapWord* copy = filler + BrooksPointer::word_size();
234 oop copy_val = oop(copy);
235
236 Copy::aligned_disjoint_words((HeapWord*) p, copy, size_no_fwdptr);
237 BrooksPointer::initialize(oop(copy));
238
239 // Try to install the new forwarding pointer.
240 oop result = BrooksPointer::try_update_forwardee(p, copy_val);
241
242 if (oopDesc::unsafe_equals(result, p)) {
243 // Successfully evacuated. Our copy is now the public one!
244 evacuated = true;
245 shenandoah_assert_correct(NULL, copy_val);
246 return copy_val;
247 } else {
248 // Failed to evacuate. We need to deal with the object that is left behind. Since this
249 // new allocation is certainly after TAMS, it will be considered live in the next cycle.
250 // But if it happens to contain references to evacuated regions, those references would
251 // not get updated for this stale copy during this cycle, and we will crash while scanning
252 // it the next cycle.
253 //
254 // For GCLAB allocations, it is enough to rollback the allocation ptr. Either the next
255 // object will overwrite this stale copy, or the filler object on LAB retirement will
256 // do this. For non-GCLAB allocations, we have no way to retract the allocation, and
257 // have to explicitly overwrite the copy with the filler object. With that overwrite,
258 // we have to keep the fwdptr initialized and pointing to our (stale) copy.
259 if (alloc_from_gclab) {
260 thread->gclab().rollback(size_with_fwdptr);
261 } else {
262 fill_with_object(copy, size_no_fwdptr);
263 }
264 shenandoah_assert_correct(NULL, copy_val);
265 shenandoah_assert_correct(NULL, result);
266 return result;
267 }
268 }
269
270 inline bool ShenandoahHeap::requires_marking(const void* entry) const {
271 return !_marking_context->is_marked(oop(entry));
272 }
273
274 template <class T>
275 inline bool ShenandoahHeap::in_collection_set(T p) const {
276 HeapWord* obj = (HeapWord*) p;
277 assert(collection_set() != NULL, "Sanity");
278 assert(is_in(obj), "should be in heap");
279
280 return collection_set()->is_in(obj);
281 }
282
283 inline bool ShenandoahHeap::is_stable() const {
284 return _gc_state.is_clear();
285 }
286
287 inline bool ShenandoahHeap::is_idle() const {
288 return _gc_state.is_unset(MARKING | EVACUATION | UPDATEREFS);
289 }
290
291 inline bool ShenandoahHeap::is_concurrent_mark_in_progress() const {
292 return _gc_state.is_set(MARKING);
293 }
294
295 inline bool ShenandoahHeap::is_evacuation_in_progress() const {
296 return _gc_state.is_set(EVACUATION);
297 }
298
299 inline bool ShenandoahHeap::is_gc_in_progress_mask(uint mask) const {
300 return _gc_state.is_set(mask);
301 }
302
303 inline bool ShenandoahHeap::is_degenerated_gc_in_progress() const {
304 return _degenerated_gc_in_progress.is_set();
305 }
306
307 inline bool ShenandoahHeap::is_full_gc_in_progress() const {
308 return _full_gc_in_progress.is_set();
309 }
310
311 inline bool ShenandoahHeap::is_full_gc_move_in_progress() const {
312 return _full_gc_move_in_progress.is_set();
313 }
314
315 inline bool ShenandoahHeap::is_update_refs_in_progress() const {
316 return _gc_state.is_set(UPDATEREFS);
317 }
318
319 template<class T>
320 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl) {
321 marked_object_iterate(region, cl, region->top());
322 }
323
324 template<class T>
325 inline void ShenandoahHeap::marked_object_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* limit) {
326 assert(BrooksPointer::word_offset() < 0, "skip_delta calculation below assumes the forwarding ptr is before obj");
327
328 ShenandoahMarkingContext* const ctx = complete_marking_context();
329 assert(ctx->is_complete(), "sanity");
330
331 MarkBitMap* mark_bit_map = ctx->mark_bit_map();
332 HeapWord* tams = ctx->top_at_mark_start(region);
333
334 size_t skip_bitmap_delta = BrooksPointer::word_size() + 1;
335 size_t skip_objsize_delta = BrooksPointer::word_size() /* + actual obj.size() below */;
336 HeapWord* start = region->bottom() + BrooksPointer::word_size();
337 HeapWord* end = MIN2(tams + BrooksPointer::word_size(), region->end());
338
339 // Step 1. Scan below the TAMS based on bitmap data.
340 HeapWord* limit_bitmap = MIN2(limit, tams);
341
342 // Try to scan the initial candidate. If the candidate is above the TAMS, it would
343 // fail the subsequent "< limit_bitmap" checks, and fall through to Step 2.
344 HeapWord* cb = mark_bit_map->getNextMarkedWordAddress(start, end);
345
346 intx dist = ShenandoahMarkScanPrefetch;
347 if (dist > 0) {
348 // Batched scan that prefetches the oop data, anticipating the access to
349 // either header, oop field, or forwarding pointer. Not that we cannot
350 // touch anything in oop, while it still being prefetched to get enough
351 // time for prefetch to work. This is why we try to scan the bitmap linearly,
352 // disregarding the object size. However, since we know forwarding pointer
353 // preceeds the object, we can skip over it. Once we cannot trust the bitmap,
354 // there is no point for prefetching the oop contents, as oop->size() will
355 // touch it prematurely.
356
357 // No variable-length arrays in standard C++, have enough slots to fit
358 // the prefetch distance.
359 static const int SLOT_COUNT = 256;
360 guarantee(dist <= SLOT_COUNT, "adjust slot count");
361 HeapWord* slots[SLOT_COUNT];
362
363 int avail;
364 do {
365 avail = 0;
366 for (int c = 0; (c < dist) && (cb < limit_bitmap); c++) {
367 Prefetch::read(cb, BrooksPointer::byte_offset());
368 slots[avail++] = cb;
369 cb += skip_bitmap_delta;
370 if (cb < limit_bitmap) {
371 cb = mark_bit_map->getNextMarkedWordAddress(cb, limit_bitmap);
372 }
373 }
374
375 for (int c = 0; c < avail; c++) {
376 assert (slots[c] < tams, err_msg("only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(tams)));
377 assert (slots[c] < limit, err_msg("only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(slots[c]), p2i(limit)));
378 oop obj = oop(slots[c]);
379 assert(!oopDesc::is_null(obj), "sanity");
380 assert(obj->is_oop(), "sanity");
381 assert(_marking_context->is_marked(obj), "object expected to be marked");
382 cl->do_object(obj);
383 }
384 } while (avail > 0);
385 } else {
386 while (cb < limit_bitmap) {
387 assert (cb < tams, err_msg("only objects below TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(tams)));
388 assert (cb < limit, err_msg("only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cb), p2i(limit)));
389 oop obj = oop(cb);
390 assert(!oopDesc::is_null(obj), "sanity");
391 assert(obj->is_oop(), "sanity");
392 assert(_marking_context->is_marked(obj), "object expected to be marked");
393 cl->do_object(obj);
394 cb += skip_bitmap_delta;
395 if (cb < limit_bitmap) {
396 cb = mark_bit_map->getNextMarkedWordAddress(cb, limit_bitmap);
397 }
398 }
399 }
400
401 // Step 2. Accurate size-based traversal, happens past the TAMS.
402 // This restarts the scan at TAMS, which makes sure we traverse all objects,
403 // regardless of what happened at Step 1.
404 HeapWord* cs = tams + BrooksPointer::word_size();
405 while (cs < limit) {
406 assert (cs > tams, err_msg("only objects past TAMS here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(tams)));
407 assert (cs < limit, err_msg("only objects below limit here: " PTR_FORMAT " (" PTR_FORMAT ")", p2i(cs), p2i(limit)));
408 oop obj = oop(cs);
409 int size = obj->size();
410 assert(!oopDesc::is_null(obj), "sanity");
411 assert(obj->is_oop(), "sanity");
412 assert(_marking_context->is_marked(obj), "object expected to be marked");
413 cl->do_object(obj);
414 cs += size + skip_objsize_delta;
415 }
416 }
417
418 template <class T>
419 class ShenandoahObjectToOopClosure : public ObjectClosure {
420 T* _cl;
421 public:
422 ShenandoahObjectToOopClosure(T* cl) : _cl(cl) {}
423
424 void do_object(oop obj) {
425 obj->oop_iterate(_cl);
426 }
427 };
428
429 template <class T>
430 class ShenandoahObjectToOopBoundedClosure : public ObjectClosure {
431 T* _cl;
432 MemRegion _bounds;
433 public:
434 ShenandoahObjectToOopBoundedClosure(T* cl, HeapWord* bottom, HeapWord* top) :
435 _cl(cl), _bounds(bottom, top) {}
436
437 void do_object(oop obj) {
438 obj->oop_iterate(_cl, _bounds);
439 }
440 };
441
442 template<class T>
443 inline void ShenandoahHeap::marked_object_oop_iterate(ShenandoahHeapRegion* region, T* cl, HeapWord* top) {
444 if (region->is_humongous()) {
445 HeapWord* bottom = region->bottom();
446 if (top > bottom) {
447 region = region->humongous_start_region();
448 ShenandoahObjectToOopBoundedClosure<T> objs(cl, bottom, top);
449 marked_object_iterate(region, &objs);
450 }
451 } else {
452 ShenandoahObjectToOopClosure<T> objs(cl);
453 marked_object_iterate(region, &objs, top);
454 }
455 }
456
457 inline ShenandoahHeapRegion* const ShenandoahHeap::get_region(size_t region_idx) const {
458 if (region_idx < _num_regions) {
459 return _regions[region_idx];
460 } else {
461 return NULL;
462 }
463 }
464
465 inline void ShenandoahHeap::mark_complete_marking_context() {
466 _marking_context->mark_complete();
467 }
468
469 inline void ShenandoahHeap::mark_incomplete_marking_context() {
470 _marking_context->mark_incomplete();
471 }
472
473 inline ShenandoahMarkingContext* ShenandoahHeap::complete_marking_context() const {
474 assert (_marking_context->is_complete()," sanity");
475 return _marking_context;
476 }
477
478 inline ShenandoahMarkingContext* ShenandoahHeap::marking_context() const {
479 return _marking_context;
480 }
481
482 #endif // SHARE_VM_GC_SHENANDOAH_SHENANDOAHHEAP_INLINE_HPP