1 /*
2 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
3 * Copyright (c) 2025, Oracle and/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 #include "gc/shenandoah/shenandoahAgeCensus.hpp"
27 #include "gc/shenandoah/shenandoahClosures.inline.hpp"
28 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
29 #include "gc/shenandoah/shenandoahFreeSet.hpp"
30 #include "gc/shenandoah/shenandoahGeneration.hpp"
31 #include "gc/shenandoah/shenandoahGenerationalControlThread.hpp"
32 #include "gc/shenandoah/shenandoahGenerationalEvacuationTask.hpp"
33 #include "gc/shenandoah/shenandoahGenerationalHeap.hpp"
34 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
35 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
36 #include "gc/shenandoah/shenandoahHeapRegionClosures.hpp"
37 #include "gc/shenandoah/shenandoahInitLogger.hpp"
38 #include "gc/shenandoah/shenandoahMemoryPool.hpp"
39 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
40 #include "gc/shenandoah/shenandoahOldGeneration.hpp"
41 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
42 #include "gc/shenandoah/shenandoahRegulatorThread.hpp"
43 #include "gc/shenandoah/shenandoahScanRemembered.inline.hpp"
44 #include "gc/shenandoah/shenandoahUtils.hpp"
45 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
46 #include "gc/shenandoah/shenandoahYoungGeneration.hpp"
47 #include "logging/log.hpp"
48 #include "utilities/events.hpp"
49
50
51 class ShenandoahGenerationalInitLogger : public ShenandoahInitLogger {
52 public:
53 static void print() {
54 ShenandoahGenerationalInitLogger logger;
55 logger.print_all();
56 }
57 protected:
58 void print_gc_specific() override {
59 ShenandoahInitLogger::print_gc_specific();
60
61 ShenandoahGenerationalHeap* heap = ShenandoahGenerationalHeap::heap();
62 log_info(gc, init)("Young Heuristics: %s", heap->young_generation()->heuristics()->name());
63 log_info(gc, init)("Old Heuristics: %s", heap->old_generation()->heuristics()->name());
64 }
65 };
66
67 size_t ShenandoahGenerationalHeap::calculate_min_plab() {
68 return align_up(PLAB::min_size(), CardTable::card_size_in_words());
69 }
70
71 size_t ShenandoahGenerationalHeap::calculate_max_plab() {
72 size_t MaxTLABSizeWords = ShenandoahHeapRegion::max_tlab_size_words();
73 return align_down(MaxTLABSizeWords, CardTable::card_size_in_words());
74 }
75
76 // Returns size in bytes
77 size_t ShenandoahGenerationalHeap::unsafe_max_tlab_alloc() const {
78 return MIN2(ShenandoahHeapRegion::max_tlab_size_bytes(), young_generation()->available());
79 }
80
81 ShenandoahGenerationalHeap::ShenandoahGenerationalHeap(ShenandoahCollectorPolicy* policy) :
82 ShenandoahHeap(policy),
83 _age_census(nullptr),
84 _min_plab_size(calculate_min_plab()),
85 _max_plab_size(calculate_max_plab()),
86 _regulator_thread(nullptr),
87 _young_gen_memory_pool(nullptr),
88 _old_gen_memory_pool(nullptr) {
89 assert(is_aligned(_min_plab_size, CardTable::card_size_in_words()), "min_plab_size must be aligned");
90 assert(is_aligned(_max_plab_size, CardTable::card_size_in_words()), "max_plab_size must be aligned");
91 }
92
93 void ShenandoahGenerationalHeap::initialize_generations() {
94 ShenandoahHeap::initialize_generations();
95 _young_generation->post_initialize(this);
96 _old_generation->post_initialize(this);
97 }
98
99 void ShenandoahGenerationalHeap::post_initialize() {
100 ShenandoahHeap::post_initialize();
101 _age_census = new ShenandoahAgeCensus();
102 }
103
104 void ShenandoahGenerationalHeap::post_initialize_heuristics() {
105 ShenandoahHeap::post_initialize_heuristics();
106 _young_generation->post_initialize_heuristics();
107 _old_generation->post_initialize_heuristics();
108 }
109
110 void ShenandoahGenerationalHeap::print_init_logger() const {
111 ShenandoahGenerationalInitLogger logger;
112 logger.print_all();
113 }
114
115 void ShenandoahGenerationalHeap::initialize_heuristics() {
116 // Initialize global generation and heuristics even in generational mode.
117 ShenandoahHeap::initialize_heuristics();
118
119 _young_generation = new ShenandoahYoungGeneration(max_workers());
120 _old_generation = new ShenandoahOldGeneration(max_workers());
121 _young_generation->initialize_heuristics(mode());
122 _old_generation->initialize_heuristics(mode());
123 }
124
125 void ShenandoahGenerationalHeap::initialize_serviceability() {
126 assert(mode()->is_generational(), "Only for the generational mode");
127 _young_gen_memory_pool = new ShenandoahYoungGenMemoryPool(this);
128 _old_gen_memory_pool = new ShenandoahOldGenMemoryPool(this);
129 cycle_memory_manager()->add_pool(_young_gen_memory_pool);
130 cycle_memory_manager()->add_pool(_old_gen_memory_pool);
131 stw_memory_manager()->add_pool(_young_gen_memory_pool);
132 stw_memory_manager()->add_pool(_old_gen_memory_pool);
133 }
134
135 GrowableArray<MemoryPool*> ShenandoahGenerationalHeap::memory_pools() {
136 assert(mode()->is_generational(), "Only for the generational mode");
137 GrowableArray<MemoryPool*> memory_pools(2);
138 memory_pools.append(_young_gen_memory_pool);
139 memory_pools.append(_old_gen_memory_pool);
140 return memory_pools;
141 }
142
143 void ShenandoahGenerationalHeap::initialize_controller() {
144 auto control_thread = new ShenandoahGenerationalControlThread();
145 _control_thread = control_thread;
146 _regulator_thread = new ShenandoahRegulatorThread(control_thread);
147 }
148
149 void ShenandoahGenerationalHeap::gc_threads_do(ThreadClosure* tcl) const {
150 if (!shenandoah_policy()->is_at_shutdown()) {
151 ShenandoahHeap::gc_threads_do(tcl);
152 tcl->do_thread(regulator_thread());
153 }
154 }
155
156 void ShenandoahGenerationalHeap::stop() {
157 ShenandoahHeap::stop();
158 regulator_thread()->stop();
159 }
160
161 void ShenandoahGenerationalHeap::start_idle_span() {
162 young_generation()->heuristics()->start_idle_span();
163 }
164
165 bool ShenandoahGenerationalHeap::requires_barriers(stackChunkOop obj) const {
166 if (is_idle()) {
167 return false;
168 }
169
170 if (is_concurrent_young_mark_in_progress() && is_in_young(obj) && !marking_context()->allocated_after_mark_start(obj)) {
171 // We are marking young, this object is in young, and it is below the TAMS
172 return true;
173 }
174
175 if (is_in_old(obj)) {
176 // Card marking barriers are required for objects in the old generation
177 return true;
178 }
179
180 if (has_forwarded_objects()) {
181 // Object may have pointers that need to be updated
182 return true;
183 }
184
185 return false;
186 }
187
188 void ShenandoahGenerationalHeap::evacuate_collection_set(ShenandoahGeneration* generation, bool concurrent) {
189 ShenandoahRegionIterator regions;
190 ShenandoahGenerationalEvacuationTask task(this, generation, ®ions, concurrent, false /* only promote regions */);
191 workers()->run_task(&task);
192 }
193
194 void ShenandoahGenerationalHeap::promote_regions_in_place(ShenandoahGeneration* generation, bool concurrent) {
195 ShenandoahRegionIterator regions;
196 ShenandoahGenerationalEvacuationTask task(this, generation, ®ions, concurrent, true /* only promote regions */);
197 workers()->run_task(&task);
198 }
199
200 oop ShenandoahGenerationalHeap::evacuate_object(oop p, Thread* thread) {
201 assert(thread == Thread::current(), "Expected thread parameter to be current thread.");
202 if (ShenandoahThreadLocalData::is_oom_during_evac(thread)) {
203 // This thread went through the OOM during evac protocol and it is safe to return
204 // the forward pointer. It must not attempt to evacuate anymore.
205 return ShenandoahBarrierSet::resolve_forwarded(p);
206 }
207
208 assert(ShenandoahThreadLocalData::is_evac_allowed(thread), "must be enclosed in oom-evac scope");
209
210 ShenandoahHeapRegion* from_region = heap_region_containing(p);
211 assert(!from_region->is_humongous(), "never evacuate humongous objects");
212
213 // Try to keep the object in the same generation
214 const ShenandoahAffiliation target_gen = from_region->affiliation();
215
216 if (target_gen == YOUNG_GENERATION) {
217 markWord mark = p->mark();
218 if (mark.is_marked()) {
219 // Already forwarded.
220 return ShenandoahBarrierSet::resolve_forwarded(p);
221 }
222
223 if (mark.has_displaced_mark_helper()) {
224 // We don't want to deal with MT here just to ensure we read the right mark word.
225 // Skip the potential promotion attempt for this one.
226 } else if (age_census()->is_tenurable(from_region->age() + mark.age())) {
227 // If the object is tenurable, try to promote it
228 oop result = try_evacuate_object<YOUNG_GENERATION, OLD_GENERATION>(p, thread, from_region->age());
229
230 // If we failed to promote this aged object, we'll fall through to code below and evacuate to young-gen.
231 if (result != nullptr) {
232 return result;
233 }
234 }
235 return try_evacuate_object<YOUNG_GENERATION, YOUNG_GENERATION>(p, thread, from_region->age());
236 }
237
238 assert(target_gen == OLD_GENERATION, "Expected evacuation to old");
239 return try_evacuate_object<OLD_GENERATION, OLD_GENERATION>(p, thread, from_region->age());
240 }
241
242 // try_evacuate_object registers the object and dirties the associated remembered set information when evacuating
243 // to OLD_GENERATION.
244 template<ShenandoahAffiliation FROM_GENERATION, ShenandoahAffiliation TO_GENERATION>
245 oop ShenandoahGenerationalHeap::try_evacuate_object(oop p, Thread* thread, uint from_region_age) {
246 bool alloc_from_lab = true;
247 bool has_plab = false;
248 HeapWord* copy = nullptr;
249 size_t size = ShenandoahForwarding::size(p);
250 constexpr bool is_promotion = (TO_GENERATION == OLD_GENERATION) && (FROM_GENERATION == YOUNG_GENERATION);
251
252 #ifdef ASSERT
253 if (ShenandoahOOMDuringEvacALot &&
254 (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call
255 copy = nullptr;
256 } else {
257 #endif
258 if (UseTLAB) {
259 switch (TO_GENERATION) {
260 case YOUNG_GENERATION: {
261 copy = allocate_from_gclab(thread, size);
262 if ((copy == nullptr) && (size < ShenandoahThreadLocalData::gclab_size(thread))) {
263 // GCLAB allocation failed because we are bumping up against the limit on young evacuation reserve. Try resetting
264 // the desired GCLAB size and retry GCLAB allocation to avoid cascading of shared memory allocations.
265 ShenandoahThreadLocalData::set_gclab_size(thread, PLAB::min_size());
266 copy = allocate_from_gclab(thread, size);
267 // If we still get nullptr, we'll try a shared allocation below.
268 }
269 break;
270 }
271 case OLD_GENERATION: {
272 PLAB* plab = ShenandoahThreadLocalData::plab(thread);
273 if (plab != nullptr) {
274 has_plab = true;
275 copy = allocate_from_plab(thread, size, is_promotion);
276 if ((copy == nullptr) && (size < ShenandoahThreadLocalData::plab_size(thread)) &&
277 ShenandoahThreadLocalData::plab_retries_enabled(thread)) {
278 // PLAB allocation failed because we are bumping up against the limit on old evacuation reserve or because
279 // the requested object does not fit within the current plab but the plab still has an "abundance" of memory,
280 // where abundance is defined as >= ShenGenHeap::plab_min_size(). In the former case, we try shrinking the
281 // desired PLAB size to the minimum and retry PLAB allocation to avoid cascading of shared memory allocations.
282 // Shrinking the desired PLAB size may allow us to eke out a small PLAB while staying beneath evacuation reserve.
283 if (plab->words_remaining() < plab_min_size()) {
284 ShenandoahThreadLocalData::set_plab_size(thread, plab_min_size());
285 copy = allocate_from_plab(thread, size, is_promotion);
286 // If we still get nullptr, we'll try a shared allocation below.
287 if (copy == nullptr) {
288 // If retry fails, don't continue to retry until we have success (probably in next GC pass)
289 ShenandoahThreadLocalData::disable_plab_retries(thread);
290 }
291 }
292 // else, copy still equals nullptr. this causes shared allocation below, preserving this plab for future needs.
293 }
294 }
295 break;
296 }
297 default: {
298 ShouldNotReachHere();
299 break;
300 }
301 }
302 }
303
304 if (copy == nullptr) {
305 // If we failed to allocate in LAB, we'll try a shared allocation.
306 if (!is_promotion || !has_plab || (size > PLAB::min_size())) {
307 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size, TO_GENERATION, is_promotion);
308 copy = allocate_memory(req);
309 alloc_from_lab = false;
310 }
311 // else, we leave copy equal to nullptr, signaling a promotion failure below if appropriate.
312 // We choose not to promote objects smaller than size_threshold by way of shared allocations as this is too
313 // costly. Instead, we'll simply "evacuate" to young-gen memory (using a GCLAB) and will promote in a future
314 // evacuation pass. This condition is denoted by: is_promotion && has_plab && (size <= size_threshhold).
315 }
316 #ifdef ASSERT
317 }
318 #endif
319
320 if (copy == nullptr) {
321 if (TO_GENERATION == OLD_GENERATION) {
322 if (FROM_GENERATION == YOUNG_GENERATION) {
323 // Signal that promotion failed. Will evacuate this old object somewhere in young gen.
324 old_generation()->handle_failed_promotion(thread, size);
325 return nullptr;
326 } else {
327 // Remember that evacuation to old gen failed. We'll want to trigger a full gc to recover from this
328 // after the evacuation threads have finished.
329 old_generation()->handle_failed_evacuation();
330 }
331 }
332
333 control_thread()->handle_alloc_failure_evac(size);
334 oom_evac_handler()->handle_out_of_memory_during_evacuation();
335 return ShenandoahBarrierSet::resolve_forwarded(p);
336 }
337
338 if (ShenandoahEvacTracking) {
339 evac_tracker()->begin_evacuation(thread, size * HeapWordSize, FROM_GENERATION, TO_GENERATION);
340 }
341
342 // Copy the object:
343 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(p), copy, size);
344 oop copy_val = cast_to_oop(copy);
345
346 // Update the age of the evacuated object
347 if (TO_GENERATION == YOUNG_GENERATION && is_aging_cycle()) {
348 increase_object_age(copy_val, from_region_age + 1);
349 }
350
351 // Try to install the new forwarding pointer.
352 oop result = ShenandoahForwarding::try_update_forwardee(p, copy_val);
353 if (result == copy_val) {
354 // Successfully evacuated. Our copy is now the public one!
355
356 // This is necessary for virtual thread support. This uses the mark word without
357 // considering that it may now be a forwarding pointer (and could therefore crash).
358 // Secondarily, we do not want to spend cycles relativizing stack chunks for oops
359 // that lost the evacuation race (and will therefore not become visible). It is
360 // safe to do this on the public copy (this is also done during concurrent mark).
361 ContinuationGCSupport::relativize_stack_chunk(copy_val);
362
363 if (ShenandoahEvacTracking) {
364 // Record that the evacuation succeeded
365 evac_tracker()->end_evacuation(thread, size * HeapWordSize, FROM_GENERATION, TO_GENERATION);
366 }
367
368 if (TO_GENERATION == OLD_GENERATION) {
369 old_generation()->handle_evacuation(copy, size);
370 }
371 } else {
372 // Failed to evacuate. We need to deal with the object that is left behind. Since this
373 // new allocation is certainly after TAMS, it will be considered live in the next cycle.
374 // But if it happens to contain references to evacuated regions, those references would
375 // not get updated for this stale copy during this cycle, and we will crash while scanning
376 // it the next cycle.
377 if (alloc_from_lab) {
378 // For LAB allocations, it is enough to rollback the allocation ptr. Either the next
379 // object will overwrite this stale copy, or the filler object on LAB retirement will
380 // do this.
381 switch (TO_GENERATION) {
382 case YOUNG_GENERATION: {
383 ShenandoahThreadLocalData::gclab(thread)->undo_allocation(copy, size);
384 break;
385 }
386 case OLD_GENERATION: {
387 ShenandoahThreadLocalData::plab(thread)->undo_allocation(copy, size);
388 if (is_promotion) {
389 ShenandoahThreadLocalData::subtract_from_plab_promoted(thread, size * HeapWordSize);
390 }
391 break;
392 }
393 default: {
394 ShouldNotReachHere();
395 break;
396 }
397 }
398 } else {
399 // For non-LAB allocations, we have no way to retract the allocation, and
400 // have to explicitly overwrite the copy with the filler object. With that overwrite,
401 // we have to keep the fwdptr initialized and pointing to our (stale) copy.
402 assert(size >= ShenandoahHeap::min_fill_size(), "previously allocated object known to be larger than min_size");
403 fill_with_object(copy, size);
404 }
405 }
406 shenandoah_assert_correct(nullptr, result);
407 return result;
408 }
409
410 template oop ShenandoahGenerationalHeap::try_evacuate_object<YOUNG_GENERATION, YOUNG_GENERATION>(oop p, Thread* thread, uint from_region_age);
411 template oop ShenandoahGenerationalHeap::try_evacuate_object<YOUNG_GENERATION, OLD_GENERATION>(oop p, Thread* thread, uint from_region_age);
412 template oop ShenandoahGenerationalHeap::try_evacuate_object<OLD_GENERATION, OLD_GENERATION>(oop p, Thread* thread, uint from_region_age);
413
414 inline HeapWord* ShenandoahGenerationalHeap::allocate_from_plab(Thread* thread, size_t size, bool is_promotion) {
415 assert(UseTLAB, "TLABs should be enabled");
416
417 PLAB* plab = ShenandoahThreadLocalData::plab(thread);
418 HeapWord* obj;
419
420 if (plab == nullptr) {
421 assert(!thread->is_Java_thread() && !thread->is_Worker_thread(), "Performance: thread should have PLAB: %s", thread->name());
422 // No PLABs in this thread, fallback to shared allocation
423 return nullptr;
424 } else if (is_promotion && !ShenandoahThreadLocalData::allow_plab_promotions(thread)) {
425 return nullptr;
426 }
427 // if plab->word_size() <= 0, thread's plab not yet initialized for this pass, so allow_plab_promotions() is not trustworthy
428 obj = plab->allocate(size);
429 if ((obj == nullptr) && (plab->words_remaining() < plab_min_size())) {
430 // allocate_from_plab_slow will establish allow_plab_promotions(thread) for future invocations
431 obj = allocate_from_plab_slow(thread, size, is_promotion);
432 }
433 // if plab->words_remaining() >= ShenGenHeap::heap()->plab_min_size(), just return nullptr so we can use a shared allocation
434 if (obj == nullptr) {
435 return nullptr;
436 }
437
438 if (is_promotion) {
439 ShenandoahThreadLocalData::add_to_plab_promoted(thread, size * HeapWordSize);
440 }
441 return obj;
442 }
443
444 // Establish a new PLAB and allocate size HeapWords within it.
445 HeapWord* ShenandoahGenerationalHeap::allocate_from_plab_slow(Thread* thread, size_t size, bool is_promotion) {
446 assert(mode()->is_generational(), "PLABs only relevant to generational GC");
447
448 const size_t plab_min_size = this->plab_min_size();
449 // PLABs are aligned to card boundaries to avoid synchronization with concurrent
450 // allocations in other PLABs.
451 const size_t min_size = (size > plab_min_size)? align_up(size, CardTable::card_size_in_words()): plab_min_size;
452
453 // Figure out size of new PLAB, using value determined at last refill.
454 size_t cur_size = ShenandoahThreadLocalData::plab_size(thread);
455 if (cur_size == 0) {
456 cur_size = plab_min_size;
457 }
458
459 // Expand aggressively, doubling at each refill in this epoch, ceiling at plab_max_size()
460 const size_t future_size = MIN2(cur_size * 2, plab_max_size());
461 // Doubling, starting at a card-multiple, should give us a card-multiple. (Ceiling and floor
462 // are card multiples.)
463 assert(is_aligned(future_size, CardTable::card_size_in_words()), "Card multiple by construction, future_size: %zu"
464 ", card_size: %u, cur_size: %zu, max: %zu",
465 future_size, CardTable::card_size_in_words(), cur_size, plab_max_size());
466
467 // Record new heuristic value even if we take any shortcut. This captures
468 // the case when moderately-sized objects always take a shortcut. At some point,
469 // heuristics should catch up with them. Note that the requested cur_size may
470 // not be honored, but we remember that this is the preferred size.
471 log_debug(gc, plab)("Set next PLAB refill size: %zu bytes", future_size * HeapWordSize);
472 ShenandoahThreadLocalData::set_plab_size(thread, future_size);
473
474 if (cur_size < size) {
475 // The PLAB to be allocated is still not large enough to hold the object. Fall back to shared allocation.
476 // This avoids retiring perfectly good PLABs in order to represent a single large object allocation.
477 log_debug(gc, plab)("Current PLAB size (%zu) is too small for %zu", cur_size * HeapWordSize, size * HeapWordSize);
478 return nullptr;
479 }
480
481 // Retire current PLAB, and allocate a new one.
482 PLAB* plab = ShenandoahThreadLocalData::plab(thread);
483 if (plab->words_remaining() < plab_min_size) {
484 // Retire current PLAB. This takes care of any PLAB book-keeping.
485 // retire_plab() registers the remnant filler object with the remembered set scanner without a lock.
486 // Since PLABs are card-aligned, concurrent registrations in other PLABs don't interfere.
487 retire_plab(plab, thread);
488
489 size_t actual_size = 0;
490 HeapWord* plab_buf = allocate_new_plab(min_size, cur_size, &actual_size);
491 if (plab_buf == nullptr) {
492 if (min_size == plab_min_size) {
493 // Disable PLAB promotions for this thread because we cannot even allocate a minimal PLAB. This allows us
494 // to fail faster on subsequent promotion attempts.
495 ShenandoahThreadLocalData::disable_plab_promotions(thread);
496 }
497 return nullptr;
498 } else {
499 ShenandoahThreadLocalData::enable_plab_retries(thread);
500 }
501 // Since the allocated PLAB may have been down-sized for alignment, plab->allocate(size) below may still fail.
502 if (ZeroTLAB) {
503 // ... and clear it.
504 Copy::zero_to_words(plab_buf, actual_size);
505 } else {
506 // ...and zap just allocated object.
507 #ifdef ASSERT
508 // Skip mangling the space corresponding to the object header to
509 // ensure that the returned space is not considered parsable by
510 // any concurrent GC thread.
511 size_t hdr_size = oopDesc::header_size();
512 Copy::fill_to_words(plab_buf + hdr_size, actual_size - hdr_size, badHeapWordVal);
513 #endif // ASSERT
514 }
515 assert(is_aligned(actual_size, CardTable::card_size_in_words()), "Align by design");
516 plab->set_buf(plab_buf, actual_size);
517 if (is_promotion && !ShenandoahThreadLocalData::allow_plab_promotions(thread)) {
518 return nullptr;
519 }
520 return plab->allocate(size);
521 } else {
522 // If there's still at least min_size() words available within the current plab, don't retire it. Let's nibble
523 // away on this plab as long as we can. Meanwhile, return nullptr to force this particular allocation request
524 // to be satisfied with a shared allocation. By packing more promotions into the previously allocated PLAB, we
525 // reduce the likelihood of evacuation failures, and we reduce the need for downsizing our PLABs.
526 return nullptr;
527 }
528 }
529
530 HeapWord* ShenandoahGenerationalHeap::allocate_new_plab(size_t min_size, size_t word_size, size_t* actual_size) {
531 // Align requested sizes to card-sized multiples. Align down so that we don't violate max size of TLAB.
532 assert(is_aligned(min_size, CardTable::card_size_in_words()), "Align by design");
533 assert(word_size >= min_size, "Requested PLAB is too small");
534
535 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_plab(min_size, word_size);
536 // Note that allocate_memory() sets a thread-local flag to prohibit further promotions by this thread
537 // if we are at risk of infringing on the old-gen evacuation budget.
538 HeapWord* res = allocate_memory(req);
539 if (res != nullptr) {
540 *actual_size = req.actual_size();
541 } else {
542 *actual_size = 0;
543 }
544 assert(is_aligned(res, CardTable::card_size_in_words()), "Align by design");
545 return res;
546 }
547
548 void ShenandoahGenerationalHeap::retire_plab(PLAB* plab, Thread* thread) {
549 // We don't enforce limits on plab evacuations. We let it consume all available old-gen memory in order to reduce
550 // probability of an evacuation failure. We do enforce limits on promotion, to make sure that excessive promotion
551 // does not result in an old-gen evacuation failure. Note that a failed promotion is relatively harmless. Any
552 // object that fails to promote in the current cycle will be eligible for promotion in a subsequent cycle.
553
554 // When the plab was instantiated, its entirety was treated as if the entire buffer was going to be dedicated to
555 // promotions. Now that we are retiring the buffer, we adjust for the reality that the plab is not entirely promotions.
556 // 1. Some of the plab may have been dedicated to evacuations.
557 // 2. Some of the plab may have been abandoned due to waste (at the end of the plab).
558 size_t not_promoted =
559 ShenandoahThreadLocalData::get_plab_actual_size(thread) - ShenandoahThreadLocalData::get_plab_promoted(thread);
560 ShenandoahThreadLocalData::reset_plab_promoted(thread);
561 ShenandoahThreadLocalData::set_plab_actual_size(thread, 0);
562 if (not_promoted > 0) {
563 log_debug(gc, plab)("Retire PLAB, unexpend unpromoted: %zu", not_promoted * HeapWordSize);
564 old_generation()->unexpend_promoted(not_promoted);
565 }
566 const size_t original_waste = plab->waste();
567 HeapWord* const top = plab->top();
568
569 // plab->retire() overwrites unused memory between plab->top() and plab->hard_end() with a dummy object to make memory parsable.
570 // It adds the size of this unused memory, in words, to plab->waste().
571 plab->retire();
572 if (top != nullptr && plab->waste() > original_waste && is_in_old(top)) {
573 // If retiring the plab created a filler object, then we need to register it with our card scanner so it can
574 // safely walk the region backing the plab.
575 log_debug(gc, plab)("retire_plab() is registering remnant of size %zu at " PTR_FORMAT,
576 (plab->waste() - original_waste) * HeapWordSize, p2i(top));
577 // No lock is necessary because the PLAB memory is aligned on card boundaries.
578 old_generation()->card_scan()->register_object_without_lock(top);
579 }
580 }
581
582 void ShenandoahGenerationalHeap::retire_plab(PLAB* plab) {
583 Thread* thread = Thread::current();
584 retire_plab(plab, thread);
585 }
586
587 // Make sure old-generation is large enough, but no larger than is necessary, to hold mixed evacuations
588 // and promotions, if we anticipate either. Any deficit is provided by the young generation, subject to
589 // mutator_xfer_limit, and any surplus is transferred to the young generation. mutator_xfer_limit is
590 // the maximum we're able to transfer from young to old. This is called at the end of GC, as we prepare
591 // for the idle span that precedes the next GC.
592 void ShenandoahGenerationalHeap::compute_old_generation_balance(size_t mutator_xfer_limit,
593 size_t old_trashed_regions, size_t young_trashed_regions) {
594 shenandoah_assert_heaplocked();
595 // We can limit the old reserve to the size of anticipated promotions:
596 // max_old_reserve is an upper bound on memory evacuated from old and promoted to old,
597 // clamped by the old generation space available.
598 //
599 // Here's the algebra.
600 // Let SOEP = ShenandoahOldEvacPercent,
601 // OE = old evac,
602 // YE = young evac, and
603 // TE = total evac = OE + YE
604 // By definition:
605 // SOEP/100 = OE/TE
606 // = OE/(OE+YE)
607 // => SOEP/(100-SOEP) = OE/((OE+YE)-OE) // componendo-dividendo: If a/b = c/d, then a/(b-a) = c/(d-c)
608 // = OE/YE
609 // => OE = YE*SOEP/(100-SOEP)
610
611 // We have to be careful in the event that SOEP is set to 100 by the user.
612 assert(ShenandoahOldEvacPercent <= 100, "Error");
613 const size_t region_size_bytes = ShenandoahHeapRegion::region_size_bytes();
614
615 ShenandoahOldGeneration* old_gen = old_generation();
616 size_t old_capacity = old_gen->max_capacity();
617 size_t old_usage = old_gen->used(); // includes humongous waste
618 size_t old_currently_available =
619 ((old_capacity >= old_usage)? old_capacity - old_usage: 0) + old_trashed_regions * region_size_bytes;
620
621 ShenandoahYoungGeneration* young_gen = young_generation();
622 size_t young_capacity = young_gen->max_capacity();
623 size_t young_usage = young_gen->used(); // includes humongous waste
624 size_t young_available = ((young_capacity >= young_usage)? young_capacity - young_usage: 0);
625 size_t freeset_available = free_set()->available_locked();
626 if (young_available > freeset_available) {
627 young_available = freeset_available;
628 }
629 young_available += young_trashed_regions * region_size_bytes;
630
631 // The free set will reserve this amount of memory to hold young evacuations (initialized to the ideal reserve)
632 size_t young_reserve = (young_generation()->max_capacity() * ShenandoahEvacReserve) / 100;
633
634 // If ShenandoahOldEvacPercent equals 100, max_old_reserve is limited only by mutator_xfer_limit and young_reserve
635 const size_t bound_on_old_reserve =
636 ((old_currently_available + mutator_xfer_limit + young_reserve) * ShenandoahOldEvacPercent) / 100;
637 size_t proposed_max_old = ((ShenandoahOldEvacPercent == 100)?
638 bound_on_old_reserve:
639 MIN2((young_reserve * ShenandoahOldEvacPercent) / (100 - ShenandoahOldEvacPercent),
640 bound_on_old_reserve));
641 if (young_reserve > young_available) {
642 young_reserve = young_available;
643 }
644
645 // Decide how much old space we should reserve for a mixed collection
646 size_t proposed_reserve_for_mixed = 0;
647 const size_t old_fragmented_available =
648 old_currently_available - (old_generation()->free_unaffiliated_regions() + old_trashed_regions) * region_size_bytes;
649
650 if (old_fragmented_available > proposed_max_old) {
651 // In this case, the old_fragmented_available is greater than the desired amount of evacuation to old.
652 // We'll use all of this memory to hold results of old evacuation, and we'll give back to the young generation
653 // any old regions that are not fragmented.
654 //
655 // This scenario may happen after we have promoted many regions in place, and each of these regions had non-zero
656 // unused memory, so there is now an abundance of old-fragmented available memory, even more than the desired
657 // percentage for old reserve. We cannot transfer these fragmented regions back to young. Instead we make the
658 // best of the situation by using this fragmented memory for both promotions and evacuations.
659
660 proposed_max_old = old_fragmented_available;
661 }
662 // Otherwise: old_fragmented_available <= proposed_max_old. Do not shrink proposed_max_old from the original computation.
663
664 // Though we initially set proposed_reserve_for_promo to equal the entirety of old fragmented available, we have the
665 // opportunity below to shift some of this memory into the proposed_reserve_for_mixed.
666 size_t proposed_reserve_for_promo = old_fragmented_available;
667 const size_t max_old_reserve = proposed_max_old;
668
669 const size_t mixed_candidate_live_memory = old_generation()->unprocessed_collection_candidates_live_memory();
670 const bool doing_mixed = (mixed_candidate_live_memory > 0);
671 if (doing_mixed) {
672 // In the ideal, all of the memory reserved for mixed evacuation would be unfragmented, but we don't enforce
673 // this. Note that the initial value of max_evac_need is conservative because we may not evacuate all of the
674 // remaining mixed evacuation candidates in a single cycle.
675 const size_t max_evac_need = (size_t) (mixed_candidate_live_memory * ShenandoahOldEvacWaste);
676 assert(old_currently_available >= old_generation()->free_unaffiliated_regions() * region_size_bytes,
677 "Unaffiliated available must be less than total available");
678
679 // We prefer to evacuate all of mixed into unfragmented memory, and will expand old in order to do so, unless
680 // we already have too much fragmented available memory in old.
681 proposed_reserve_for_mixed = max_evac_need;
682 if (proposed_reserve_for_mixed + proposed_reserve_for_promo > max_old_reserve) {
683 // We're trying to reserve more memory than is available. So we need to shrink our reserves.
684 size_t excess_reserves = (proposed_reserve_for_mixed + proposed_reserve_for_promo) - max_old_reserve;
685 // We need to shrink reserves by excess_reserves. We prefer to shrink by reducing promotion, giving priority to mixed
686 // evacuation. If the promotion reserve is larger than the amount we need to shrink by, do all the shrinkage there.
687 if (proposed_reserve_for_promo > excess_reserves) {
688 proposed_reserve_for_promo -= excess_reserves;
689 } else {
690 // Otherwise, we'll shrink promotion reserve to zero and we'll shrink the mixed-evac reserve by the remaining excess.
691 excess_reserves -= proposed_reserve_for_promo;
692 proposed_reserve_for_promo = 0;
693 proposed_reserve_for_mixed -= excess_reserves;
694 }
695 }
696 }
697 assert(proposed_reserve_for_mixed + proposed_reserve_for_promo <= max_old_reserve,
698 "Reserve for mixed (%zu) plus reserve for promotions (%zu) must be less than maximum old reserve (%zu)",
699 proposed_reserve_for_mixed, proposed_reserve_for_promo, max_old_reserve);
700
701 // Decide how much additional space we should reserve for promotions from young. We give priority to mixed evacations
702 // over promotions.
703 const size_t promo_load = old_generation()->get_promotion_potential();
704 const bool doing_promotions = promo_load > 0;
705
706 // promo_load represents the combined total of live memory within regions that have reached tenure age. The true
707 // promotion potential is larger than this, because individual objects within regions that have not yet reached tenure
708 // age may be promotable. On the other hand, some of the objects that we intend to promote in the next GC cycle may
709 // die before they are next marked. In the future, the promo_load will include the total size of tenurable objects
710 // residing in regions that have not yet reached tenure age.
711
712 if (doing_promotions) {
713 // We are always doing promotions, even when old_generation->get_promotion_potential() returns 0. As currently implemented,
714 // get_promotion_potential() only knows the total live memory contained within young-generation regions whose age is
715 // tenurable. It does not know whether that memory will still be live at the end of the next mark cycle, and it doesn't
716 // know how much memory is contained within objects whose individual ages are tenurable, which reside in regions with
717 // non-tenurable age. We use this, as adjusted by ShenandoahPromoEvacWaste, as an approximation of the total amount of
718 // memory to be promoted. In the near future, we expect to implement a change that will allow get_promotion_potential()
719 // to account also for the total memory contained within individual objects that are tenure-ready even when they do
720 // not reside in aged regions. This will represent a conservative over approximation of promotable memory because
721 // some of these objects may die before the next GC cycle executes.
722
723 // Be careful not to ask for too much promotion reserves. We have observed jtreg test failures under which a greedy
724 // promotion reserve causes a humongous allocation which is awaiting a full GC to fail (specifically
725 // gc/TestAllocHumongousFragment.java). This happens if too much of the memory reclaimed by the full GC
726 // is immediately reserved so that it cannot be allocated by the waiting mutator. It's not clear that this
727 // particular test is representative of the needs of typical GenShen users. It is really a test of high frequency
728 // Full GCs under heap fragmentation stress.
729
730 size_t promo_need = (size_t) (promo_load * ShenandoahPromoEvacWaste);
731 if (promo_need > proposed_reserve_for_promo) {
732 const size_t available_for_additional_promotions =
733 max_old_reserve - (proposed_reserve_for_mixed + proposed_reserve_for_promo);
734 if (proposed_reserve_for_promo + available_for_additional_promotions >= promo_need) {
735 proposed_reserve_for_promo = promo_need;
736 } else {
737 proposed_reserve_for_promo += available_for_additional_promotions;
738 }
739 }
740 }
741 // else, leave proposed_reserve_for_promo as is. By default, it is initialized to represent old_fragmented_available.
742
743 // This is the total old we want to reserve (initialized to the ideal reserve)
744 size_t proposed_old_reserve = proposed_reserve_for_mixed + proposed_reserve_for_promo;
745
746 // We now check if the old generation is running a surplus or a deficit.
747 size_t old_region_deficit = 0;
748 size_t old_region_surplus = 0;
749
750 size_t mutator_region_xfer_limit = mutator_xfer_limit / region_size_bytes;
751 // align the mutator_xfer_limit on region size
752 mutator_xfer_limit = mutator_region_xfer_limit * region_size_bytes;
753
754 if (old_currently_available >= proposed_old_reserve) {
755 // We are running a surplus, so the old region surplus can go to young
756 const size_t old_surplus = old_currently_available - proposed_old_reserve;
757 old_region_surplus = old_surplus / region_size_bytes;
758 const size_t unaffiliated_old_regions = old_generation()->free_unaffiliated_regions() + old_trashed_regions;
759 old_region_surplus = MIN2(old_region_surplus, unaffiliated_old_regions);
760 old_generation()->set_region_balance(checked_cast<ssize_t>(old_region_surplus));
761 old_currently_available -= old_region_surplus * region_size_bytes;
762 young_available += old_region_surplus * region_size_bytes;
763 } else if (old_currently_available + mutator_xfer_limit >= proposed_old_reserve) {
764 // We know that old_currently_available < proposed_old_reserve because above test failed. Expand old_currently_available.
765 // Mutator's xfer limit is sufficient to satisfy our need: transfer all memory from there.
766 size_t old_deficit = proposed_old_reserve - old_currently_available;
767 old_region_deficit = (old_deficit + region_size_bytes - 1) / region_size_bytes;
768 old_generation()->set_region_balance(0 - checked_cast<ssize_t>(old_region_deficit));
769 old_currently_available += old_region_deficit * region_size_bytes;
770 young_available -= old_region_deficit * region_size_bytes;
771 } else {
772 // We know that (old_currently_available < proposed_old_reserve) and
773 // (old_currently_available + mutator_xfer_limit < proposed_old_reserve) because above tests failed.
774 // We need to shrink proposed_old_reserves.
775
776 // We could potentially shrink young_reserves in order to further expand proposed_old_reserves. Let's not bother. The
777 // important thing is that we keep a total amount of memory in reserve in preparation for the next GC cycle. At
778 // the time we choose the next collection set, we'll have an opportunity to shift some of these young reserves
779 // into old reserves if that makes sense.
780
781 // Start by taking all of mutator_xfer_limit into old_currently_available.
782 size_t old_region_deficit = mutator_region_xfer_limit;
783 old_generation()->set_region_balance(0 - checked_cast<ssize_t>(old_region_deficit));
784 old_currently_available += old_region_deficit * region_size_bytes;
785 young_available -= old_region_deficit * region_size_bytes;
786
787 assert(old_currently_available < proposed_old_reserve,
788 "Old currently available (%zu) must be less than old reserve (%zu)", old_currently_available, proposed_old_reserve);
789
790 // There's not enough memory to satisfy our desire. Scale back our old-gen intentions. We prefer to satisfy
791 // the budget_overrun entirely from the promotion reserve, if that is large enough. Otherwise, we'll satisfy
792 // the overrun from a combination of promotion and mixed-evacuation reserves.
793 size_t budget_overrun = proposed_old_reserve - old_currently_available;
794 if (proposed_reserve_for_promo > budget_overrun) {
795 proposed_reserve_for_promo -= budget_overrun;
796 // Dead code:
797 // proposed_old_reserve -= budget_overrun;
798 } else {
799 budget_overrun -= proposed_reserve_for_promo;
800 proposed_reserve_for_promo = 0;
801 proposed_reserve_for_mixed = (proposed_reserve_for_mixed > budget_overrun)? proposed_reserve_for_mixed - budget_overrun: 0;
802 // Dead code:
803 // Note: proposed_reserve_for_promo is 0 and proposed_reserve_for_mixed may equal 0.
804 // proposed_old_reserve = proposed_reserve_for_mixed;
805 }
806 }
807
808 assert(old_region_deficit == 0 || old_region_surplus == 0,
809 "Only surplus (%zu) or deficit (%zu), never both", old_region_surplus, old_region_deficit);
810 assert(young_reserve + proposed_reserve_for_mixed + proposed_reserve_for_promo <= old_currently_available + young_available,
811 "Cannot reserve more memory than is available: %zu + %zu + %zu <= %zu + %zu",
812 young_reserve, proposed_reserve_for_mixed, proposed_reserve_for_promo, old_currently_available, young_available);
813
814 // deficit/surplus adjustments to generation sizes will precede rebuild
815 young_generation()->set_evacuation_reserve(young_reserve);
816 old_generation()->set_evacuation_reserve(proposed_reserve_for_mixed);
817 old_generation()->set_promoted_reserve(proposed_reserve_for_promo);
818 }
819
820 void ShenandoahGenerationalHeap::coalesce_and_fill_old_regions(bool concurrent) {
821 class ShenandoahGlobalCoalesceAndFill : public WorkerTask {
822 private:
823 ShenandoahPhaseTimings::Phase _phase;
824 ShenandoahRegionIterator _regions;
825 public:
826 explicit ShenandoahGlobalCoalesceAndFill(ShenandoahPhaseTimings::Phase phase) :
827 WorkerTask("Shenandoah Global Coalesce"),
828 _phase(phase) {}
829
830 void work(uint worker_id) override {
831 ShenandoahWorkerTimingsTracker timer(_phase,
832 ShenandoahPhaseTimings::ScanClusters,
833 worker_id, true);
834 ShenandoahHeapRegion* region;
835 while ((region = _regions.next()) != nullptr) {
836 // old region is not in the collection set and was not immediately trashed
837 if (region->is_old() && region->is_active() && !region->is_humongous()) {
838 // Reset the coalesce and fill boundary because this is a global collect
839 // and cannot be preempted by young collects. We want to be sure the entire
840 // region is coalesced here and does not resume from a previously interrupted
841 // or completed coalescing.
842 region->begin_preemptible_coalesce_and_fill();
843 region->oop_coalesce_and_fill(false);
844 }
845 }
846 }
847 };
848
849 ShenandoahPhaseTimings::Phase phase = concurrent ?
850 ShenandoahPhaseTimings::conc_coalesce_and_fill :
851 ShenandoahPhaseTimings::degen_gc_coalesce_and_fill;
852
853 // This is not cancellable
854 ShenandoahGlobalCoalesceAndFill coalesce(phase);
855 workers()->run_task(&coalesce);
856 old_generation()->set_parsable(true);
857 }
858
859 template<bool CONCURRENT>
860 class ShenandoahGenerationalUpdateHeapRefsTask : public WorkerTask {
861 private:
862 // For update refs, _generation will be young or global. Mixed collections use the young generation.
863 ShenandoahGeneration* _generation;
864 ShenandoahGenerationalHeap* _heap;
865 ShenandoahRegionIterator* _regions;
866 ShenandoahRegionChunkIterator* _work_chunks;
867
868 public:
869 ShenandoahGenerationalUpdateHeapRefsTask(ShenandoahGeneration* generation,
870 ShenandoahRegionIterator* regions,
871 ShenandoahRegionChunkIterator* work_chunks) :
872 WorkerTask("Shenandoah Update References"),
873 _generation(generation),
874 _heap(ShenandoahGenerationalHeap::heap()),
875 _regions(regions),
876 _work_chunks(work_chunks)
877 {
878 const bool old_bitmap_stable = _heap->old_generation()->is_mark_complete();
879 log_debug(gc, remset)("Update refs, scan remembered set using bitmap: %s", BOOL_TO_STR(old_bitmap_stable));
880 }
881
882 void work(uint worker_id) override {
883 if (CONCURRENT) {
884 ShenandoahConcurrentWorkerSession worker_session(worker_id);
885 ShenandoahSuspendibleThreadSetJoiner stsj;
886 do_work<ShenandoahConcUpdateRefsClosure>(worker_id);
887 } else {
888 ShenandoahParallelWorkerSession worker_session(worker_id);
889 do_work<ShenandoahNonConcUpdateRefsClosure>(worker_id);
890 }
891 }
892
893 private:
894 template<class T>
895 void do_work(uint worker_id) {
896 T cl;
897
898 if (CONCURRENT && (worker_id == 0)) {
899 // We ask the first worker to replenish the Mutator free set by moving regions previously reserved to hold the
900 // results of evacuation. These reserves are no longer necessary because evacuation has completed.
901 size_t cset_regions = _heap->collection_set()->count();
902
903 // Now that evacuation is done, we can reassign any regions that had been reserved to hold the results of evacuation
904 // to the mutator free set. At the end of GC, we will have cset_regions newly evacuated fully empty regions from
905 // which we will be able to replenish the Collector free set and the OldCollector free set in preparation for the
906 // next GC cycle.
907 _heap->free_set()->move_regions_from_collector_to_mutator(cset_regions);
908 }
909 // If !CONCURRENT, there's no value in expanding Mutator free set
910
911 ShenandoahHeapRegion* r = _regions->next();
912 // We update references for global, mixed, and young collections.
913 assert(_generation->is_mark_complete(), "Expected complete marking");
914 ShenandoahMarkingContext* const ctx = _heap->marking_context();
915 bool is_mixed = _heap->collection_set()->has_old_regions();
916 while (r != nullptr) {
917 HeapWord* update_watermark = r->get_update_watermark();
918 assert(update_watermark >= r->bottom(), "sanity");
919
920 log_debug(gc)("Update refs worker " UINT32_FORMAT ", looking at region %zu", worker_id, r->index());
921 if (r->is_active() && !r->is_cset()) {
922 if (r->is_young()) {
923 _heap->marked_object_oop_iterate(r, &cl, update_watermark);
924 } else if (r->is_old()) {
925 if (_generation->is_global()) {
926
927 _heap->marked_object_oop_iterate(r, &cl, update_watermark);
928 }
929 // Otherwise, this is an old region in a young or mixed cycle. Process it during a second phase, below.
930 } else {
931 // Because updating of references runs concurrently, it is possible that a FREE inactive region transitions
932 // to a non-free active region while this loop is executing. Whenever this happens, the changing of a region's
933 // active status may propagate at a different speed than the changing of the region's affiliation.
934
935 // When we reach this control point, it is because a race has allowed a region's is_active() status to be seen
936 // by this thread before the region's affiliation() is seen by this thread.
937
938 // It's ok for this race to occur because the newly transformed region does not have any references to be
939 // updated.
940
941 assert(r->get_update_watermark() == r->bottom(),
942 "%s Region %zu is_active but not recognized as YOUNG or OLD so must be newly transitioned from FREE",
943 r->affiliation_name(), r->index());
944 }
945 }
946
947 if (_heap->check_cancelled_gc_and_yield(CONCURRENT)) {
948 return;
949 }
950
951 r = _regions->next();
952 }
953
954 if (_generation->is_young()) {
955 // Since this is generational and not GLOBAL, we have to process the remembered set. There's no remembered
956 // set processing if not in generational mode or if GLOBAL mode.
957
958 // After this thread has exhausted its traditional update-refs work, it continues with updating refs within
959 // remembered set. The remembered set workload is better balanced between threads, so threads that are "behind"
960 // can catch up with other threads during this phase, allowing all threads to work more effectively in parallel.
961 update_references_in_remembered_set(worker_id, cl, ctx, is_mixed);
962 }
963 }
964
965 template<class T>
966 void update_references_in_remembered_set(uint worker_id, T &cl, const ShenandoahMarkingContext* ctx, bool is_mixed) {
967
968 struct ShenandoahRegionChunk assignment;
969 ShenandoahScanRemembered* scanner = _heap->old_generation()->card_scan();
970
971 while (!_heap->check_cancelled_gc_and_yield(CONCURRENT) && _work_chunks->next(&assignment)) {
972 // Keep grabbing next work chunk to process until finished, or asked to yield
973 ShenandoahHeapRegion* r = assignment._r;
974 if (r->is_active() && !r->is_cset() && r->is_old()) {
975 HeapWord* start_of_range = r->bottom() + assignment._chunk_offset;
976 HeapWord* end_of_range = r->get_update_watermark();
977 if (end_of_range > start_of_range + assignment._chunk_size) {
978 end_of_range = start_of_range + assignment._chunk_size;
979 }
980
981 if (start_of_range >= end_of_range) {
982 continue;
983 }
984
985 // Old region in a young cycle or mixed cycle.
986 if (is_mixed) {
987 if (r->is_humongous()) {
988 // Need to examine both dirty and clean cards during mixed evac.
989 r->oop_iterate_humongous_slice_all(&cl,start_of_range, assignment._chunk_size);
990 } else {
991 // Since this is mixed evacuation, old regions that are candidates for collection have not been coalesced
992 // and filled. This will use mark bits to find objects that need to be updated.
993 update_references_in_old_region(cl, ctx, scanner, r, start_of_range, end_of_range);
994 }
995 } else {
996 // This is a young evacuation
997 size_t cluster_size = CardTable::card_size_in_words() * ShenandoahCardCluster::CardsPerCluster;
998 size_t clusters = assignment._chunk_size / cluster_size;
999 assert(clusters * cluster_size == assignment._chunk_size, "Chunk assignment must align on cluster boundaries");
1000 scanner->process_region_slice(r, assignment._chunk_offset, clusters, end_of_range, &cl, true, worker_id);
1001 }
1002 }
1003 }
1004 }
1005
1006 template<class T>
1007 void update_references_in_old_region(T &cl, const ShenandoahMarkingContext* ctx, ShenandoahScanRemembered* scanner,
1008 const ShenandoahHeapRegion* r, HeapWord* start_of_range,
1009 HeapWord* end_of_range) const {
1010 // In case last object in my range spans boundary of my chunk, I may need to scan all the way to top()
1011 ShenandoahObjectToOopBoundedClosure<T> objs(&cl, start_of_range, r->top());
1012
1013 // Any object that begins in a previous range is part of a different scanning assignment. Any object that
1014 // starts after end_of_range is also not my responsibility. (Either allocated during evacuation, so does
1015 // not hold pointers to from-space, or is beyond the range of my assigned work chunk.)
1016
1017 // Find the first object that begins in my range, if there is one. Note that `p` will be set to `end_of_range`
1018 // when no live object is found in the range.
1019 HeapWord* tams = ctx->top_at_mark_start(r);
1020 HeapWord* p = get_first_object_start_word(ctx, scanner, tams, start_of_range, end_of_range);
1021
1022 while (p < end_of_range) {
1023 // p is known to point to the beginning of marked object obj
1024 oop obj = cast_to_oop(p);
1025 objs.do_object(obj);
1026 HeapWord* prev_p = p;
1027 p += obj->size();
1028 if (p < tams) {
1029 p = ctx->get_next_marked_addr(p, tams);
1030 // If there are no more marked objects before tams, this returns tams. Note that tams is
1031 // either >= end_of_range, or tams is the start of an object that is marked.
1032 }
1033 assert(p != prev_p, "Lack of forward progress");
1034 }
1035 }
1036
1037 HeapWord* get_first_object_start_word(const ShenandoahMarkingContext* ctx, ShenandoahScanRemembered* scanner, HeapWord* tams,
1038 HeapWord* start_of_range, HeapWord* end_of_range) const {
1039 HeapWord* p = start_of_range;
1040
1041 if (p >= tams) {
1042 // We cannot use ctx->is_marked(obj) to test whether an object begins at this address. Instead,
1043 // we need to use the remembered set crossing map to advance p to the first object that starts
1044 // within the enclosing card.
1045 size_t card_index = scanner->card_index_for_addr(start_of_range);
1046 while (true) {
1047 HeapWord* first_object = scanner->first_object_in_card(card_index);
1048 if (first_object != nullptr) {
1049 p = first_object;
1050 break;
1051 } else if (scanner->addr_for_card_index(card_index + 1) < end_of_range) {
1052 card_index++;
1053 } else {
1054 // Signal that no object was found in range
1055 p = end_of_range;
1056 break;
1057 }
1058 }
1059 } else if (!ctx->is_marked(cast_to_oop(p))) {
1060 p = ctx->get_next_marked_addr(p, tams);
1061 // If there are no more marked objects before tams, this returns tams.
1062 // Note that tams is either >= end_of_range, or tams is the start of an object that is marked.
1063 }
1064 return p;
1065 }
1066 };
1067
1068 void ShenandoahGenerationalHeap::update_heap_references(ShenandoahGeneration* generation, bool concurrent) {
1069 assert(!is_full_gc_in_progress(), "Only for concurrent and degenerated GC");
1070 const uint nworkers = workers()->active_workers();
1071 ShenandoahRegionChunkIterator work_list(nworkers);
1072 if (concurrent) {
1073 ShenandoahGenerationalUpdateHeapRefsTask<true> task(generation, &_update_refs_iterator, &work_list);
1074 workers()->run_task(&task);
1075 } else {
1076 ShenandoahGenerationalUpdateHeapRefsTask<false> task(generation, &_update_refs_iterator, &work_list);
1077 workers()->run_task(&task);
1078 }
1079
1080 if (ShenandoahEnableCardStats) {
1081 // Only do this if we are collecting card stats
1082 ShenandoahScanRemembered* card_scan = old_generation()->card_scan();
1083 assert(card_scan != nullptr, "Card table must exist when card stats are enabled");
1084 card_scan->log_card_stats(nworkers, CARD_STAT_UPDATE_REFS);
1085 }
1086 }
1087
1088 struct ShenandoahCompositeRegionClosure {
1089 template<typename C1, typename C2>
1090 class Closure : public ShenandoahHeapRegionClosure {
1091 private:
1092 C1 &_c1;
1093 C2 &_c2;
1094
1095 public:
1096 Closure(C1 &c1, C2 &c2) : ShenandoahHeapRegionClosure(), _c1(c1), _c2(c2) {}
1097
1098 void heap_region_do(ShenandoahHeapRegion* r) override {
1099 _c1.heap_region_do(r);
1100 _c2.heap_region_do(r);
1101 }
1102
1103 bool is_thread_safe() override {
1104 return _c1.is_thread_safe() && _c2.is_thread_safe();
1105 }
1106 };
1107
1108 template<typename C1, typename C2>
1109 static Closure<C1, C2> of(C1 &c1, C2 &c2) {
1110 return Closure<C1, C2>(c1, c2);
1111 }
1112 };
1113
1114 class ShenandoahUpdateRegionAges : public ShenandoahHeapRegionClosure {
1115 private:
1116 ShenandoahMarkingContext* _ctx;
1117
1118 public:
1119 explicit ShenandoahUpdateRegionAges(ShenandoahMarkingContext* ctx) : _ctx(ctx) { }
1120
1121 void heap_region_do(ShenandoahHeapRegion* r) override {
1122 // Maintenance of region age must follow evacuation in order to account for
1123 // evacuation allocations within survivor regions. We consult region age during
1124 // the subsequent evacuation to determine whether certain objects need to
1125 // be promoted.
1126 if (r->is_young() && r->is_active()) {
1127 HeapWord *tams = _ctx->top_at_mark_start(r);
1128 HeapWord *top = r->top();
1129
1130 // Allocations move the watermark when top moves. However, compacting
1131 // objects will sometimes lower top beneath the watermark, after which,
1132 // attempts to read the watermark will assert out (watermark should not be
1133 // higher than top).
1134 if (top > tams) {
1135 // There have been allocations in this region since the start of the cycle.
1136 // Any objects new to this region must not assimilate elevated age.
1137 r->reset_age();
1138 } else if (ShenandoahGenerationalHeap::heap()->is_aging_cycle()) {
1139 r->increment_age();
1140 }
1141 }
1142 }
1143
1144 bool is_thread_safe() override {
1145 return true;
1146 }
1147 };
1148
1149 void ShenandoahGenerationalHeap::final_update_refs_update_region_states() {
1150 ShenandoahSynchronizePinnedRegionStates pins;
1151 ShenandoahUpdateRegionAges ages(marking_context());
1152 auto cl = ShenandoahCompositeRegionClosure::of(pins, ages);
1153 parallel_heap_region_iterate(&cl);
1154 }
1155
1156 void ShenandoahGenerationalHeap::complete_degenerated_cycle() {
1157 shenandoah_assert_heaplocked_or_safepoint();
1158 if (!old_generation()->is_parsable()) {
1159 ShenandoahGCPhase phase(ShenandoahPhaseTimings::degen_gc_coalesce_and_fill);
1160 coalesce_and_fill_old_regions(false);
1161 }
1162
1163 log_info(gc, cset)("Degenerated cycle complete, promotions reserved: %zu, promotions expended: %zu, failed count: %zu, failed bytes: %zu",
1164 old_generation()->get_promoted_reserve(), old_generation()->get_promoted_expended(),
1165 old_generation()->get_promotion_failed_count(), old_generation()->get_promotion_failed_words() * HeapWordSize);
1166 }
1167
1168 void ShenandoahGenerationalHeap::complete_concurrent_cycle() {
1169 if (!old_generation()->is_parsable()) {
1170 // Class unloading may render the card offsets unusable, so we must rebuild them before
1171 // the next remembered set scan. We _could_ let the control thread do this sometime after
1172 // the global cycle has completed and before the next young collection, but under memory
1173 // pressure the control thread may not have the time (that is, because it's running back
1174 // to back GCs). In that scenario, we would have to make the old regions parsable before
1175 // we could start a young collection. This could delay the start of the young cycle and
1176 // throw off the heuristics.
1177 entry_global_coalesce_and_fill();
1178 }
1179
1180 log_info(gc, cset)("Concurrent cycle complete, promotions reserved: %zu, promotions expended: %zu, failed count: %zu, failed bytes: %zu",
1181 old_generation()->get_promoted_reserve(), old_generation()->get_promoted_expended(),
1182 old_generation()->get_promotion_failed_count(), old_generation()->get_promotion_failed_words() * HeapWordSize);
1183 }
1184
1185 void ShenandoahGenerationalHeap::entry_global_coalesce_and_fill() {
1186 const char* msg = "Coalescing and filling old regions";
1187 ShenandoahConcurrentPhase gc_phase(msg, ShenandoahPhaseTimings::conc_coalesce_and_fill);
1188
1189 TraceCollectorStats tcs(monitoring_support()->concurrent_collection_counters());
1190 EventMark em("%s", msg);
1191 ShenandoahWorkerScope scope(workers(),
1192 ShenandoahWorkerPolicy::calc_workers_for_conc_marking(),
1193 "concurrent coalesce and fill");
1194
1195 coalesce_and_fill_old_regions(true);
1196 }
1197
1198 void ShenandoahGenerationalHeap::update_region_ages(ShenandoahMarkingContext* ctx) {
1199 ShenandoahUpdateRegionAges cl(ctx);
1200 parallel_heap_region_iterate(&cl);
1201 }