1 /*
2 * Copyright (c) 2023, 2024, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2013, 2022, Red Hat, Inc. All rights reserved.
4 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 *
7 * This code is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 only, as
9 * published by the Free Software Foundation.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 *
25 */
26
27 #include "precompiled.hpp"
28 #include "memory/allocation.hpp"
29 #include "memory/universe.hpp"
30
31 #include "gc/shared/classUnloadingContext.hpp"
32 #include "gc/shared/gcArguments.hpp"
33 #include "gc/shared/gcTimer.hpp"
34 #include "gc/shared/gcTraceTime.inline.hpp"
35 #include "gc/shared/locationPrinter.inline.hpp"
36 #include "gc/shared/memAllocator.hpp"
37 #include "gc/shared/plab.hpp"
38 #include "gc/shared/tlab_globals.hpp"
39
40 #include "gc/shenandoah/heuristics/shenandoahOldHeuristics.hpp"
41 #include "gc/shenandoah/heuristics/shenandoahYoungHeuristics.hpp"
42 #include "gc/shenandoah/shenandoahAllocRequest.hpp"
43 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
44 #include "gc/shenandoah/shenandoahClosures.inline.hpp"
45 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
46 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
47 #include "gc/shenandoah/shenandoahConcurrentMark.hpp"
48 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
49 #include "gc/shenandoah/shenandoahControlThread.hpp"
50 #include "gc/shenandoah/shenandoahFreeSet.hpp"
51 #include "gc/shenandoah/shenandoahGenerationalEvacuationTask.hpp"
52 #include "gc/shenandoah/shenandoahGenerationalHeap.hpp"
53 #include "gc/shenandoah/shenandoahGlobalGeneration.hpp"
54 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
55 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
56 #include "gc/shenandoah/shenandoahHeapRegionClosures.hpp"
57 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
58 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
59 #include "gc/shenandoah/shenandoahInitLogger.hpp"
60 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
61 #include "gc/shenandoah/shenandoahMemoryPool.hpp"
62 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
63 #include "gc/shenandoah/shenandoahOldGeneration.hpp"
64 #include "gc/shenandoah/shenandoahOopClosures.inline.hpp"
65 #include "gc/shenandoah/shenandoahPacer.inline.hpp"
66 #include "gc/shenandoah/shenandoahPadding.hpp"
67 #include "gc/shenandoah/shenandoahParallelCleaning.inline.hpp"
68 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
69 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
70 #include "gc/shenandoah/shenandoahScanRemembered.inline.hpp"
71 #include "gc/shenandoah/shenandoahSTWMark.hpp"
72 #include "gc/shenandoah/shenandoahUtils.hpp"
73 #include "gc/shenandoah/shenandoahVerifier.hpp"
74 #include "gc/shenandoah/shenandoahCodeRoots.hpp"
75 #include "gc/shenandoah/shenandoahVMOperations.hpp"
76 #include "gc/shenandoah/shenandoahWorkGroup.hpp"
77 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
78 #include "gc/shenandoah/shenandoahYoungGeneration.hpp"
79 #include "gc/shenandoah/mode/shenandoahGenerationalMode.hpp"
80 #include "gc/shenandoah/mode/shenandoahIUMode.hpp"
81 #include "gc/shenandoah/mode/shenandoahPassiveMode.hpp"
82 #include "gc/shenandoah/mode/shenandoahSATBMode.hpp"
83 #include "utilities/globalDefinitions.hpp"
84
85 #if INCLUDE_JFR
86 #include "gc/shenandoah/shenandoahJfrSupport.hpp"
87 #endif
88
89 #include "classfile/systemDictionary.hpp"
90 #include "code/codeCache.hpp"
91 #include "memory/classLoaderMetaspace.hpp"
92 #include "memory/metaspaceUtils.hpp"
93 #include "oops/compressedOops.inline.hpp"
94 #include "prims/jvmtiTagMap.hpp"
95 #include "runtime/atomic.hpp"
96 #include "runtime/globals.hpp"
97 #include "runtime/interfaceSupport.inline.hpp"
98 #include "runtime/java.hpp"
99 #include "runtime/orderAccess.hpp"
100 #include "runtime/safepointMechanism.hpp"
101 #include "runtime/vmThread.hpp"
102 #include "services/mallocTracker.hpp"
103 #include "services/memTracker.hpp"
104 #include "utilities/events.hpp"
105 #include "utilities/powerOfTwo.hpp"
106
107 class ShenandoahPretouchHeapTask : public WorkerTask {
108 private:
109 ShenandoahRegionIterator _regions;
110 const size_t _page_size;
111 public:
112 ShenandoahPretouchHeapTask(size_t page_size) :
113 WorkerTask("Shenandoah Pretouch Heap"),
114 _page_size(page_size) {}
115
116 virtual void work(uint worker_id) {
117 ShenandoahHeapRegion* r = _regions.next();
118 while (r != nullptr) {
119 if (r->is_committed()) {
120 os::pretouch_memory(r->bottom(), r->end(), _page_size);
121 }
122 r = _regions.next();
123 }
124 }
125 };
126
127 class ShenandoahPretouchBitmapTask : public WorkerTask {
128 private:
129 ShenandoahRegionIterator _regions;
130 char* _bitmap_base;
131 const size_t _bitmap_size;
132 const size_t _page_size;
133 public:
134 ShenandoahPretouchBitmapTask(char* bitmap_base, size_t bitmap_size, size_t page_size) :
135 WorkerTask("Shenandoah Pretouch Bitmap"),
136 _bitmap_base(bitmap_base),
137 _bitmap_size(bitmap_size),
138 _page_size(page_size) {}
139
140 virtual void work(uint worker_id) {
141 ShenandoahHeapRegion* r = _regions.next();
142 while (r != nullptr) {
143 size_t start = r->index() * ShenandoahHeapRegion::region_size_bytes() / MarkBitMap::heap_map_factor();
144 size_t end = (r->index() + 1) * ShenandoahHeapRegion::region_size_bytes() / MarkBitMap::heap_map_factor();
145 assert (end <= _bitmap_size, "end is sane: " SIZE_FORMAT " < " SIZE_FORMAT, end, _bitmap_size);
146
147 if (r->is_committed()) {
148 os::pretouch_memory(_bitmap_base + start, _bitmap_base + end, _page_size);
149 }
150
151 r = _regions.next();
152 }
153 }
154 };
155
156 jint ShenandoahHeap::initialize() {
157 //
158 // Figure out heap sizing
159 //
160
161 size_t init_byte_size = InitialHeapSize;
162 size_t min_byte_size = MinHeapSize;
163 size_t max_byte_size = MaxHeapSize;
164 size_t heap_alignment = HeapAlignment;
165
166 size_t reg_size_bytes = ShenandoahHeapRegion::region_size_bytes();
167
168 Universe::check_alignment(max_byte_size, reg_size_bytes, "Shenandoah heap");
169 Universe::check_alignment(init_byte_size, reg_size_bytes, "Shenandoah heap");
170
171 _num_regions = ShenandoahHeapRegion::region_count();
172 assert(_num_regions == (max_byte_size / reg_size_bytes),
173 "Regions should cover entire heap exactly: " SIZE_FORMAT " != " SIZE_FORMAT "/" SIZE_FORMAT,
174 _num_regions, max_byte_size, reg_size_bytes);
175
176 size_t num_committed_regions = init_byte_size / reg_size_bytes;
177 num_committed_regions = MIN2(num_committed_regions, _num_regions);
178 assert(num_committed_regions <= _num_regions, "sanity");
179 _initial_size = num_committed_regions * reg_size_bytes;
180
181 size_t num_min_regions = min_byte_size / reg_size_bytes;
182 num_min_regions = MIN2(num_min_regions, _num_regions);
183 assert(num_min_regions <= _num_regions, "sanity");
184 _minimum_size = num_min_regions * reg_size_bytes;
185
186 // Default to max heap size.
187 _soft_max_size = _num_regions * reg_size_bytes;
188
189 _committed = _initial_size;
190
191 size_t heap_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
192 size_t bitmap_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
193 size_t region_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
194
195 //
196 // Reserve and commit memory for heap
197 //
198
199 ReservedHeapSpace heap_rs = Universe::reserve_heap(max_byte_size, heap_alignment);
200 initialize_reserved_region(heap_rs);
201 _heap_region = MemRegion((HeapWord*)heap_rs.base(), heap_rs.size() / HeapWordSize);
202 _heap_region_special = heap_rs.special();
203
204 assert((((size_t) base()) & ShenandoahHeapRegion::region_size_bytes_mask()) == 0,
205 "Misaligned heap: " PTR_FORMAT, p2i(base()));
206 os::trace_page_sizes_for_requested_size("Heap",
207 max_byte_size, heap_rs.page_size(), heap_alignment,
208 heap_rs.base(), heap_rs.size());
209
210 #if SHENANDOAH_OPTIMIZED_MARKTASK
211 // The optimized ShenandoahMarkTask takes some bits away from the full object bits.
212 // Fail if we ever attempt to address more than we can.
213 if ((uintptr_t)heap_rs.end() >= ShenandoahMarkTask::max_addressable()) {
214 FormatBuffer<512> buf("Shenandoah reserved [" PTR_FORMAT ", " PTR_FORMAT") for the heap, \n"
215 "but max object address is " PTR_FORMAT ". Try to reduce heap size, or try other \n"
216 "VM options that allocate heap at lower addresses (HeapBaseMinAddress, AllocateHeapAt, etc).",
217 p2i(heap_rs.base()), p2i(heap_rs.end()), ShenandoahMarkTask::max_addressable());
218 vm_exit_during_initialization("Fatal Error", buf);
219 }
220 #endif
221
222 ReservedSpace sh_rs = heap_rs.first_part(max_byte_size);
223 if (!_heap_region_special) {
224 os::commit_memory_or_exit(sh_rs.base(), _initial_size, heap_alignment, false,
225 "Cannot commit heap memory");
226 }
227
228 BarrierSet::set_barrier_set(new ShenandoahBarrierSet(this, _heap_region));
229
230 // Now we know the number of regions and heap sizes, initialize the heuristics.
231 initialize_heuristics();
232
233 assert(_heap_region.byte_size() == heap_rs.size(), "Need to know reserved size for card table");
234
235 //
236 // Worker threads must be initialized after the barrier is configured
237 //
238 _workers = new ShenandoahWorkerThreads("Shenandoah GC Threads", _max_workers);
239 if (_workers == nullptr) {
240 vm_exit_during_initialization("Failed necessary allocation.");
241 } else {
242 _workers->initialize_workers();
243 }
244
245 if (ParallelGCThreads > 1) {
246 _safepoint_workers = new ShenandoahWorkerThreads("Safepoint Cleanup Thread", ParallelGCThreads);
247 _safepoint_workers->initialize_workers();
248 }
249
250 //
251 // Reserve and commit memory for bitmap(s)
252 //
253
254 size_t bitmap_size_orig = ShenandoahMarkBitMap::compute_size(heap_rs.size());
255 _bitmap_size = align_up(bitmap_size_orig, bitmap_page_size);
256
257 size_t bitmap_bytes_per_region = reg_size_bytes / ShenandoahMarkBitMap::heap_map_factor();
258
259 guarantee(bitmap_bytes_per_region != 0,
260 "Bitmap bytes per region should not be zero");
261 guarantee(is_power_of_2(bitmap_bytes_per_region),
262 "Bitmap bytes per region should be power of two: " SIZE_FORMAT, bitmap_bytes_per_region);
263
264 if (bitmap_page_size > bitmap_bytes_per_region) {
265 _bitmap_regions_per_slice = bitmap_page_size / bitmap_bytes_per_region;
266 _bitmap_bytes_per_slice = bitmap_page_size;
267 } else {
268 _bitmap_regions_per_slice = 1;
269 _bitmap_bytes_per_slice = bitmap_bytes_per_region;
270 }
271
272 guarantee(_bitmap_regions_per_slice >= 1,
273 "Should have at least one region per slice: " SIZE_FORMAT,
274 _bitmap_regions_per_slice);
275
276 guarantee(((_bitmap_bytes_per_slice) % bitmap_page_size) == 0,
277 "Bitmap slices should be page-granular: bps = " SIZE_FORMAT ", page size = " SIZE_FORMAT,
278 _bitmap_bytes_per_slice, bitmap_page_size);
279
280 ReservedSpace bitmap(_bitmap_size, bitmap_page_size);
281 os::trace_page_sizes_for_requested_size("Mark Bitmap",
282 bitmap_size_orig, bitmap.page_size(), bitmap_page_size,
283 bitmap.base(),
284 bitmap.size());
285 MemTracker::record_virtual_memory_type(bitmap.base(), mtGC);
286 _bitmap_region = MemRegion((HeapWord*) bitmap.base(), bitmap.size() / HeapWordSize);
287 _bitmap_region_special = bitmap.special();
288
289 size_t bitmap_init_commit = _bitmap_bytes_per_slice *
290 align_up(num_committed_regions, _bitmap_regions_per_slice) / _bitmap_regions_per_slice;
291 bitmap_init_commit = MIN2(_bitmap_size, bitmap_init_commit);
292 if (!_bitmap_region_special) {
293 os::commit_memory_or_exit((char *) _bitmap_region.start(), bitmap_init_commit, bitmap_page_size, false,
294 "Cannot commit bitmap memory");
295 }
296
297 _marking_context = new ShenandoahMarkingContext(_heap_region, _bitmap_region, _num_regions);
298
299 if (ShenandoahVerify) {
300 ReservedSpace verify_bitmap(_bitmap_size, bitmap_page_size);
301 os::trace_page_sizes_for_requested_size("Verify Bitmap",
302 bitmap_size_orig, verify_bitmap.page_size(), bitmap_page_size,
303 verify_bitmap.base(),
304 verify_bitmap.size());
305 if (!verify_bitmap.special()) {
306 os::commit_memory_or_exit(verify_bitmap.base(), verify_bitmap.size(), bitmap_page_size, false,
307 "Cannot commit verification bitmap memory");
308 }
309 MemTracker::record_virtual_memory_type(verify_bitmap.base(), mtGC);
310 MemRegion verify_bitmap_region = MemRegion((HeapWord *) verify_bitmap.base(), verify_bitmap.size() / HeapWordSize);
311 _verification_bit_map.initialize(_heap_region, verify_bitmap_region);
312 _verifier = new ShenandoahVerifier(this, &_verification_bit_map);
313 }
314
315 // Reserve aux bitmap for use in object_iterate(). We don't commit it here.
316 size_t aux_bitmap_page_size = bitmap_page_size;
317
318 ReservedSpace aux_bitmap(_bitmap_size, aux_bitmap_page_size);
319 os::trace_page_sizes_for_requested_size("Aux Bitmap",
320 bitmap_size_orig, aux_bitmap.page_size(), aux_bitmap_page_size,
321 aux_bitmap.base(), aux_bitmap.size());
322 MemTracker::record_virtual_memory_type(aux_bitmap.base(), mtGC);
323 _aux_bitmap_region = MemRegion((HeapWord*) aux_bitmap.base(), aux_bitmap.size() / HeapWordSize);
324 _aux_bitmap_region_special = aux_bitmap.special();
325 _aux_bit_map.initialize(_heap_region, _aux_bitmap_region);
326
327 //
328 // Create regions and region sets
329 //
330 size_t region_align = align_up(sizeof(ShenandoahHeapRegion), SHENANDOAH_CACHE_LINE_SIZE);
331 size_t region_storage_size_orig = region_align * _num_regions;
332 size_t region_storage_size = align_up(region_storage_size_orig,
333 MAX2(region_page_size, os::vm_allocation_granularity()));
334
335 ReservedSpace region_storage(region_storage_size, region_page_size);
336 os::trace_page_sizes_for_requested_size("Region Storage",
337 region_storage_size_orig, region_storage.page_size(), region_page_size,
338 region_storage.base(), region_storage.size());
339 MemTracker::record_virtual_memory_type(region_storage.base(), mtGC);
340 if (!region_storage.special()) {
341 os::commit_memory_or_exit(region_storage.base(), region_storage_size, region_page_size, false,
342 "Cannot commit region memory");
343 }
344
345 // Try to fit the collection set bitmap at lower addresses. This optimizes code generation for cset checks.
346 // Go up until a sensible limit (subject to encoding constraints) and try to reserve the space there.
347 // If not successful, bite a bullet and allocate at whatever address.
348 {
349 const size_t cset_align = MAX2<size_t>(os::vm_page_size(), os::vm_allocation_granularity());
350 const size_t cset_size = align_up(((size_t) sh_rs.base() + sh_rs.size()) >> ShenandoahHeapRegion::region_size_bytes_shift(), cset_align);
351 const size_t cset_page_size = os::vm_page_size();
352
353 uintptr_t min = round_up_power_of_2(cset_align);
354 uintptr_t max = (1u << 30u);
355 ReservedSpace cset_rs;
356
357 for (uintptr_t addr = min; addr <= max; addr <<= 1u) {
358 char* req_addr = (char*)addr;
359 assert(is_aligned(req_addr, cset_align), "Should be aligned");
360 cset_rs = ReservedSpace(cset_size, cset_align, cset_page_size, req_addr);
361 if (cset_rs.is_reserved()) {
362 assert(cset_rs.base() == req_addr, "Allocated where requested: " PTR_FORMAT ", " PTR_FORMAT, p2i(cset_rs.base()), addr);
363 _collection_set = new ShenandoahCollectionSet(this, cset_rs, sh_rs.base());
364 break;
365 }
366 }
367
368 if (_collection_set == nullptr) {
369 cset_rs = ReservedSpace(cset_size, cset_align, os::vm_page_size());
370 _collection_set = new ShenandoahCollectionSet(this, cset_rs, sh_rs.base());
371 }
372 os::trace_page_sizes_for_requested_size("Collection Set",
373 cset_size, cset_rs.page_size(), cset_page_size,
374 cset_rs.base(),
375 cset_rs.size());
376 }
377
378 _regions = NEW_C_HEAP_ARRAY(ShenandoahHeapRegion*, _num_regions, mtGC);
379 _affiliations = NEW_C_HEAP_ARRAY(uint8_t, _num_regions, mtGC);
380 _free_set = new ShenandoahFreeSet(this, _num_regions);
381
382 {
383 ShenandoahHeapLocker locker(lock());
384
385 for (size_t i = 0; i < _num_regions; i++) {
386 HeapWord* start = (HeapWord*)sh_rs.base() + ShenandoahHeapRegion::region_size_words() * i;
387 bool is_committed = i < num_committed_regions;
388 void* loc = region_storage.base() + i * region_align;
389
390 ShenandoahHeapRegion* r = new (loc) ShenandoahHeapRegion(start, i, is_committed);
391 assert(is_aligned(r, SHENANDOAH_CACHE_LINE_SIZE), "Sanity");
392
393 _marking_context->initialize_top_at_mark_start(r);
394 _regions[i] = r;
395 assert(!collection_set()->is_in(i), "New region should not be in collection set");
396
397 _affiliations[i] = ShenandoahAffiliation::FREE;
398 }
399
400 // Initialize to complete
401 _marking_context->mark_complete();
402 size_t young_cset_regions, old_cset_regions;
403
404 // We are initializing free set. We ignore cset region tallies.
405 size_t first_old, last_old, num_old;
406 _free_set->prepare_to_rebuild(young_cset_regions, old_cset_regions, first_old, last_old, num_old);
407 _free_set->finish_rebuild(young_cset_regions, old_cset_regions, num_old);
408 }
409
410 if (AlwaysPreTouch) {
411 // For NUMA, it is important to pre-touch the storage under bitmaps with worker threads,
412 // before initialize() below zeroes it with initializing thread. For any given region,
413 // we touch the region and the corresponding bitmaps from the same thread.
414 ShenandoahPushWorkerScope scope(workers(), _max_workers, false);
415
416 _pretouch_heap_page_size = heap_page_size;
417 _pretouch_bitmap_page_size = bitmap_page_size;
418
419 // OS memory managers may want to coalesce back-to-back pages. Make their jobs
420 // simpler by pre-touching continuous spaces (heap and bitmap) separately.
421
422 ShenandoahPretouchBitmapTask bcl(bitmap.base(), _bitmap_size, _pretouch_bitmap_page_size);
423 _workers->run_task(&bcl);
424
425 ShenandoahPretouchHeapTask hcl(_pretouch_heap_page_size);
426 _workers->run_task(&hcl);
427 }
428
429 //
430 // Initialize the rest of GC subsystems
431 //
432
433 _liveness_cache = NEW_C_HEAP_ARRAY(ShenandoahLiveData*, _max_workers, mtGC);
434 for (uint worker = 0; worker < _max_workers; worker++) {
435 _liveness_cache[worker] = NEW_C_HEAP_ARRAY(ShenandoahLiveData, _num_regions, mtGC);
436 Copy::fill_to_bytes(_liveness_cache[worker], _num_regions * sizeof(ShenandoahLiveData));
437 }
438
439 // There should probably be Shenandoah-specific options for these,
440 // just as there are G1-specific options.
441 {
442 ShenandoahSATBMarkQueueSet& satbqs = ShenandoahBarrierSet::satb_mark_queue_set();
443 satbqs.set_process_completed_buffers_threshold(20); // G1SATBProcessCompletedThreshold
444 satbqs.set_buffer_enqueue_threshold_percentage(60); // G1SATBBufferEnqueueingThresholdPercent
445 }
446
447 _monitoring_support = new ShenandoahMonitoringSupport(this);
448 _phase_timings = new ShenandoahPhaseTimings(max_workers());
449 ShenandoahCodeRoots::initialize();
450
451 if (ShenandoahPacing) {
452 _pacer = new ShenandoahPacer(this);
453 _pacer->setup_for_idle();
454 }
455
456 initialize_controller();
457
458 print_init_logger();
459
460 return JNI_OK;
461 }
462
463 void ShenandoahHeap::initialize_controller() {
464 _control_thread = new ShenandoahControlThread();
465 }
466
467 void ShenandoahHeap::print_init_logger() const {
468 ShenandoahInitLogger::print();
469 }
470
471 void ShenandoahHeap::initialize_mode() {
472 if (ShenandoahGCMode != nullptr) {
473 if (strcmp(ShenandoahGCMode, "satb") == 0) {
474 _gc_mode = new ShenandoahSATBMode();
475 } else if (strcmp(ShenandoahGCMode, "iu") == 0) {
476 _gc_mode = new ShenandoahIUMode();
477 } else if (strcmp(ShenandoahGCMode, "passive") == 0) {
478 _gc_mode = new ShenandoahPassiveMode();
479 } else if (strcmp(ShenandoahGCMode, "generational") == 0) {
480 _gc_mode = new ShenandoahGenerationalMode();
481 } else {
482 vm_exit_during_initialization("Unknown -XX:ShenandoahGCMode option");
483 }
484 } else {
485 vm_exit_during_initialization("Unknown -XX:ShenandoahGCMode option (null)");
486 }
487 _gc_mode->initialize_flags();
488 if (_gc_mode->is_diagnostic() && !UnlockDiagnosticVMOptions) {
489 vm_exit_during_initialization(
490 err_msg("GC mode \"%s\" is diagnostic, and must be enabled via -XX:+UnlockDiagnosticVMOptions.",
491 _gc_mode->name()));
492 }
493 if (_gc_mode->is_experimental() && !UnlockExperimentalVMOptions) {
494 vm_exit_during_initialization(
495 err_msg("GC mode \"%s\" is experimental, and must be enabled via -XX:+UnlockExperimentalVMOptions.",
496 _gc_mode->name()));
497 }
498 }
499
500 void ShenandoahHeap::initialize_heuristics() {
501 _global_generation = new ShenandoahGlobalGeneration(mode()->is_generational(), max_workers(), max_capacity(), max_capacity());
502 _global_generation->initialize_heuristics(mode());
503 }
504
505 #ifdef _MSC_VER
506 #pragma warning( push )
507 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
508 #endif
509
510 ShenandoahHeap::ShenandoahHeap(ShenandoahCollectorPolicy* policy) :
511 CollectedHeap(),
512 _gc_generation(nullptr),
513 _active_generation(nullptr),
514 _initial_size(0),
515 _committed(0),
516 _max_workers(MAX3(ConcGCThreads, ParallelGCThreads, 1U)),
517 _workers(nullptr),
518 _safepoint_workers(nullptr),
519 _heap_region_special(false),
520 _num_regions(0),
521 _regions(nullptr),
522 _affiliations(nullptr),
523 _gc_state_changed(false),
524 _gc_no_progress_count(0),
525 _cancel_requested_time(0),
526 _update_refs_iterator(this),
527 _global_generation(nullptr),
528 _control_thread(nullptr),
529 _young_generation(nullptr),
530 _old_generation(nullptr),
531 _shenandoah_policy(policy),
532 _gc_mode(nullptr),
533 _free_set(nullptr),
534 _pacer(nullptr),
535 _verifier(nullptr),
536 _phase_timings(nullptr),
537 _monitoring_support(nullptr),
538 _memory_pool(nullptr),
539 _stw_memory_manager("Shenandoah Pauses"),
540 _cycle_memory_manager("Shenandoah Cycles"),
541 _gc_timer(new ConcurrentGCTimer()),
542 _soft_ref_policy(),
543 _log_min_obj_alignment_in_bytes(LogMinObjAlignmentInBytes),
544 _marking_context(nullptr),
545 _bitmap_size(0),
546 _bitmap_regions_per_slice(0),
547 _bitmap_bytes_per_slice(0),
548 _bitmap_region_special(false),
549 _aux_bitmap_region_special(false),
550 _liveness_cache(nullptr),
551 _collection_set(nullptr)
552 {
553 // Initialize GC mode early, many subsequent initialization procedures depend on it
554 initialize_mode();
555 }
556
557 #ifdef _MSC_VER
558 #pragma warning( pop )
559 #endif
560
561 void ShenandoahHeap::print_on(outputStream* st) const {
562 st->print_cr("Shenandoah Heap");
563 st->print_cr(" " SIZE_FORMAT "%s max, " SIZE_FORMAT "%s soft max, " SIZE_FORMAT "%s committed, " SIZE_FORMAT "%s used",
564 byte_size_in_proper_unit(max_capacity()), proper_unit_for_byte_size(max_capacity()),
565 byte_size_in_proper_unit(soft_max_capacity()), proper_unit_for_byte_size(soft_max_capacity()),
566 byte_size_in_proper_unit(committed()), proper_unit_for_byte_size(committed()),
567 byte_size_in_proper_unit(used()), proper_unit_for_byte_size(used()));
568 st->print_cr(" " SIZE_FORMAT " x " SIZE_FORMAT"%s regions",
569 num_regions(),
570 byte_size_in_proper_unit(ShenandoahHeapRegion::region_size_bytes()),
571 proper_unit_for_byte_size(ShenandoahHeapRegion::region_size_bytes()));
572
573 st->print("Status: ");
574 if (has_forwarded_objects()) st->print("has forwarded objects, ");
575 if (!mode()->is_generational()) {
576 if (is_concurrent_mark_in_progress()) st->print("marking,");
577 } else {
578 if (is_concurrent_old_mark_in_progress()) st->print("old marking, ");
579 if (is_concurrent_young_mark_in_progress()) st->print("young marking, ");
580 }
581 if (is_evacuation_in_progress()) st->print("evacuating, ");
582 if (is_update_refs_in_progress()) st->print("updating refs, ");
583 if (is_degenerated_gc_in_progress()) st->print("degenerated gc, ");
584 if (is_full_gc_in_progress()) st->print("full gc, ");
585 if (is_full_gc_move_in_progress()) st->print("full gc move, ");
586 if (is_concurrent_weak_root_in_progress()) st->print("concurrent weak roots, ");
587 if (is_concurrent_strong_root_in_progress() &&
588 !is_concurrent_weak_root_in_progress()) st->print("concurrent strong roots, ");
589
590 if (cancelled_gc()) {
591 st->print("cancelled");
592 } else {
593 st->print("not cancelled");
594 }
595 st->cr();
596
597 st->print_cr("Reserved region:");
598 st->print_cr(" - [" PTR_FORMAT ", " PTR_FORMAT ") ",
599 p2i(reserved_region().start()),
600 p2i(reserved_region().end()));
601
602 ShenandoahCollectionSet* cset = collection_set();
603 st->print_cr("Collection set:");
604 if (cset != nullptr) {
605 st->print_cr(" - map (vanilla): " PTR_FORMAT, p2i(cset->map_address()));
606 st->print_cr(" - map (biased): " PTR_FORMAT, p2i(cset->biased_map_address()));
607 } else {
608 st->print_cr(" (null)");
609 }
610
611 st->cr();
612 MetaspaceUtils::print_on(st);
613
614 if (Verbose) {
615 st->cr();
616 print_heap_regions_on(st);
617 }
618 }
619
620 class ShenandoahInitWorkerGCLABClosure : public ThreadClosure {
621 public:
622 void do_thread(Thread* thread) {
623 assert(thread != nullptr, "Sanity");
624 assert(thread->is_Worker_thread(), "Only worker thread expected");
625 ShenandoahThreadLocalData::initialize_gclab(thread);
626 }
627 };
628
629 void ShenandoahHeap::post_initialize() {
630 CollectedHeap::post_initialize();
631
632 // Schedule periodic task to report on gc thread CPU utilization
633 _mmu_tracker.initialize();
634
635 MutexLocker ml(Threads_lock);
636
637 ShenandoahInitWorkerGCLABClosure init_gclabs;
638 _workers->threads_do(&init_gclabs);
639
640 // gclab can not be initialized early during VM startup, as it can not determinate its max_size.
641 // Now, we will let WorkerThreads to initialize gclab when new worker is created.
642 _workers->set_initialize_gclab();
643 if (_safepoint_workers != nullptr) {
644 _safepoint_workers->threads_do(&init_gclabs);
645 _safepoint_workers->set_initialize_gclab();
646 }
647
648 JFR_ONLY(ShenandoahJFRSupport::register_jfr_type_serializers());
649 }
650
651 ShenandoahHeuristics* ShenandoahHeap::heuristics() {
652 return _global_generation->heuristics();
653 }
654
655 size_t ShenandoahHeap::used() const {
656 return global_generation()->used();
657 }
658
659 size_t ShenandoahHeap::committed() const {
660 return Atomic::load(&_committed);
661 }
662
663 void ShenandoahHeap::increase_committed(size_t bytes) {
664 shenandoah_assert_heaplocked_or_safepoint();
665 _committed += bytes;
666 }
667
668 void ShenandoahHeap::decrease_committed(size_t bytes) {
669 shenandoah_assert_heaplocked_or_safepoint();
670 _committed -= bytes;
671 }
672
673 // For tracking usage based on allocations, it should be the case that:
674 // * The sum of regions::used == heap::used
675 // * The sum of a generation's regions::used == generation::used
676 // * The sum of a generation's humongous regions::free == generation::humongous_waste
677 // These invariants are checked by the verifier on GC safepoints.
678 //
679 // Additional notes:
680 // * When a mutator's allocation request causes a region to be retired, the
681 // free memory left in that region is considered waste. It does not contribute
682 // to the usage, but it _does_ contribute to allocation rate.
683 // * The bottom of a PLAB must be aligned on card size. In some cases this will
684 // require padding in front of the PLAB (a filler object). Because this padding
685 // is included in the region's used memory we include the padding in the usage
686 // accounting as waste.
687 // * Mutator allocations are used to compute an allocation rate. They are also
688 // sent to the Pacer for those purposes.
689 // * There are three sources of waste:
690 // 1. The padding used to align a PLAB on card size
691 // 2. Region's free is less than minimum TLAB size and is retired
692 // 3. The unused portion of memory in the last region of a humongous object
693 void ShenandoahHeap::increase_used(const ShenandoahAllocRequest& req) {
694 size_t actual_bytes = req.actual_size() * HeapWordSize;
695 size_t wasted_bytes = req.waste() * HeapWordSize;
696 ShenandoahGeneration* generation = generation_for(req.affiliation());
697
698 if (req.is_gc_alloc()) {
699 assert(wasted_bytes == 0 || req.type() == ShenandoahAllocRequest::_alloc_plab, "Only PLABs have waste");
700 increase_used(generation, actual_bytes + wasted_bytes);
701 } else {
702 assert(req.is_mutator_alloc(), "Expected mutator alloc here");
703 // padding and actual size both count towards allocation counter
704 generation->increase_allocated(actual_bytes + wasted_bytes);
705
706 // only actual size counts toward usage for mutator allocations
707 increase_used(generation, actual_bytes);
708
709 // notify pacer of both actual size and waste
710 notify_mutator_alloc_words(req.actual_size(), req.waste());
711
712 if (wasted_bytes > 0 && ShenandoahHeapRegion::requires_humongous(req.actual_size())) {
713 increase_humongous_waste(generation,wasted_bytes);
714 }
715 }
716 }
717
718 void ShenandoahHeap::increase_humongous_waste(ShenandoahGeneration* generation, size_t bytes) {
719 generation->increase_humongous_waste(bytes);
720 if (!generation->is_global()) {
721 global_generation()->increase_humongous_waste(bytes);
722 }
723 }
724
725 void ShenandoahHeap::decrease_humongous_waste(ShenandoahGeneration* generation, size_t bytes) {
726 generation->decrease_humongous_waste(bytes);
727 if (!generation->is_global()) {
728 global_generation()->decrease_humongous_waste(bytes);
729 }
730 }
731
732 void ShenandoahHeap::increase_used(ShenandoahGeneration* generation, size_t bytes) {
733 generation->increase_used(bytes);
734 if (!generation->is_global()) {
735 global_generation()->increase_used(bytes);
736 }
737 }
738
739 void ShenandoahHeap::decrease_used(ShenandoahGeneration* generation, size_t bytes) {
740 generation->decrease_used(bytes);
741 if (!generation->is_global()) {
742 global_generation()->decrease_used(bytes);
743 }
744 }
745
746 void ShenandoahHeap::notify_mutator_alloc_words(size_t words, size_t waste) {
747 if (ShenandoahPacing) {
748 control_thread()->pacing_notify_alloc(words);
749 if (waste > 0) {
750 pacer()->claim_for_alloc<true>(waste);
751 }
752 }
753 }
754
755 size_t ShenandoahHeap::capacity() const {
756 return committed();
757 }
758
759 size_t ShenandoahHeap::max_capacity() const {
760 return _num_regions * ShenandoahHeapRegion::region_size_bytes();
761 }
762
763 size_t ShenandoahHeap::soft_max_capacity() const {
764 size_t v = Atomic::load(&_soft_max_size);
765 assert(min_capacity() <= v && v <= max_capacity(),
766 "Should be in bounds: " SIZE_FORMAT " <= " SIZE_FORMAT " <= " SIZE_FORMAT,
767 min_capacity(), v, max_capacity());
768 return v;
769 }
770
771 void ShenandoahHeap::set_soft_max_capacity(size_t v) {
772 assert(min_capacity() <= v && v <= max_capacity(),
773 "Should be in bounds: " SIZE_FORMAT " <= " SIZE_FORMAT " <= " SIZE_FORMAT,
774 min_capacity(), v, max_capacity());
775 Atomic::store(&_soft_max_size, v);
776 }
777
778 size_t ShenandoahHeap::min_capacity() const {
779 return _minimum_size;
780 }
781
782 size_t ShenandoahHeap::initial_capacity() const {
783 return _initial_size;
784 }
785
786 void ShenandoahHeap::maybe_uncommit(double shrink_before, size_t shrink_until) {
787 assert (ShenandoahUncommit, "should be enabled");
788
789 // Determine if there is work to do. This avoids taking heap lock if there is
790 // no work available, avoids spamming logs with superfluous logging messages,
791 // and minimises the amount of work while locks are taken.
792
793 if (committed() <= shrink_until) return;
794
795 bool has_work = false;
796 for (size_t i = 0; i < num_regions(); i++) {
797 ShenandoahHeapRegion* r = get_region(i);
798 if (r->is_empty_committed() && (r->empty_time() < shrink_before)) {
799 has_work = true;
800 break;
801 }
802 }
803
804 if (has_work) {
805 static const char* msg = "Concurrent uncommit";
806 ShenandoahConcurrentPhase gcPhase(msg, ShenandoahPhaseTimings::conc_uncommit, true /* log_heap_usage */);
807 EventMark em("%s", msg);
808
809 op_uncommit(shrink_before, shrink_until);
810 }
811 }
812
813 void ShenandoahHeap::op_uncommit(double shrink_before, size_t shrink_until) {
814 assert (ShenandoahUncommit, "should be enabled");
815
816 // Application allocates from the beginning of the heap, and GC allocates at
817 // the end of it. It is more efficient to uncommit from the end, so that applications
818 // could enjoy the near committed regions. GC allocations are much less frequent,
819 // and therefore can accept the committing costs.
820
821 size_t count = 0;
822 for (size_t i = num_regions(); i > 0; i--) { // care about size_t underflow
823 ShenandoahHeapRegion* r = get_region(i - 1);
824 if (r->is_empty_committed() && (r->empty_time() < shrink_before)) {
825 ShenandoahHeapLocker locker(lock());
826 if (r->is_empty_committed()) {
827 if (committed() < shrink_until + ShenandoahHeapRegion::region_size_bytes()) {
828 break;
829 }
830
831 r->make_uncommitted();
832 count++;
833 }
834 }
835 SpinPause(); // allow allocators to take the lock
836 }
837
838 if (count > 0) {
839 notify_heap_changed();
840 }
841 }
842
843 bool ShenandoahHeap::check_soft_max_changed() {
844 size_t new_soft_max = Atomic::load(&SoftMaxHeapSize);
845 size_t old_soft_max = soft_max_capacity();
846 if (new_soft_max != old_soft_max) {
847 new_soft_max = MAX2(min_capacity(), new_soft_max);
848 new_soft_max = MIN2(max_capacity(), new_soft_max);
849 if (new_soft_max != old_soft_max) {
850 log_info(gc)("Soft Max Heap Size: " SIZE_FORMAT "%s -> " SIZE_FORMAT "%s",
851 byte_size_in_proper_unit(old_soft_max), proper_unit_for_byte_size(old_soft_max),
852 byte_size_in_proper_unit(new_soft_max), proper_unit_for_byte_size(new_soft_max)
853 );
854 set_soft_max_capacity(new_soft_max);
855 return true;
856 }
857 }
858 return false;
859 }
860
861 void ShenandoahHeap::notify_heap_changed() {
862 // Update monitoring counters when we took a new region. This amortizes the
863 // update costs on slow path.
864 monitoring_support()->notify_heap_changed();
865 _heap_changed.try_set();
866 }
867
868 void ShenandoahHeap::set_forced_counters_update(bool value) {
869 monitoring_support()->set_forced_counters_update(value);
870 }
871
872 void ShenandoahHeap::handle_force_counters_update() {
873 monitoring_support()->handle_force_counters_update();
874 }
875
876 HeapWord* ShenandoahHeap::allocate_from_gclab_slow(Thread* thread, size_t size) {
877 // New object should fit the GCLAB size
878 size_t min_size = MAX2(size, PLAB::min_size());
879
880 // Figure out size of new GCLAB, looking back at heuristics. Expand aggressively.
881 size_t new_size = ShenandoahThreadLocalData::gclab_size(thread) * 2;
882
883 new_size = MIN2(new_size, PLAB::max_size());
884 new_size = MAX2(new_size, PLAB::min_size());
885
886 // Record new heuristic value even if we take any shortcut. This captures
887 // the case when moderately-sized objects always take a shortcut. At some point,
888 // heuristics should catch up with them.
889 log_debug(gc, free)("Set new GCLAB size: " SIZE_FORMAT, new_size);
890 ShenandoahThreadLocalData::set_gclab_size(thread, new_size);
891
892 if (new_size < size) {
893 // New size still does not fit the object. Fall back to shared allocation.
894 // This avoids retiring perfectly good GCLABs, when we encounter a large object.
895 log_debug(gc, free)("New gclab size (" SIZE_FORMAT ") is too small for " SIZE_FORMAT, new_size, size);
896 return nullptr;
897 }
898
899 // Retire current GCLAB, and allocate a new one.
900 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
901 gclab->retire();
902
903 size_t actual_size = 0;
904 HeapWord* gclab_buf = allocate_new_gclab(min_size, new_size, &actual_size);
905 if (gclab_buf == nullptr) {
906 return nullptr;
907 }
908
909 assert (size <= actual_size, "allocation should fit");
910
911 // ...and clear or zap just allocated TLAB, if needed.
912 if (ZeroTLAB) {
913 Copy::zero_to_words(gclab_buf, actual_size);
914 } else if (ZapTLAB) {
915 // Skip mangling the space corresponding to the object header to
916 // ensure that the returned space is not considered parsable by
917 // any concurrent GC thread.
918 size_t hdr_size = oopDesc::header_size();
919 Copy::fill_to_words(gclab_buf + hdr_size, actual_size - hdr_size, badHeapWordVal);
920 }
921 gclab->set_buf(gclab_buf, actual_size);
922 return gclab->allocate(size);
923 }
924
925 // Called from stubs in JIT code or interpreter
926 HeapWord* ShenandoahHeap::allocate_new_tlab(size_t min_size,
927 size_t requested_size,
928 size_t* actual_size) {
929 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_tlab(min_size, requested_size);
930 HeapWord* res = allocate_memory(req);
931 if (res != nullptr) {
932 *actual_size = req.actual_size();
933 } else {
934 *actual_size = 0;
935 }
936 return res;
937 }
938
939 HeapWord* ShenandoahHeap::allocate_new_gclab(size_t min_size,
940 size_t word_size,
941 size_t* actual_size) {
942 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_gclab(min_size, word_size);
943 HeapWord* res = allocate_memory(req);
944 if (res != nullptr) {
945 *actual_size = req.actual_size();
946 } else {
947 *actual_size = 0;
948 }
949 return res;
950 }
951
952 HeapWord* ShenandoahHeap::allocate_memory(ShenandoahAllocRequest& req) {
953 intptr_t pacer_epoch = 0;
954 bool in_new_region = false;
955 HeapWord* result = nullptr;
956
957 if (req.is_mutator_alloc()) {
958 if (ShenandoahPacing) {
959 pacer()->pace_for_alloc(req.size());
960 pacer_epoch = pacer()->epoch();
961 }
962
963 if (!ShenandoahAllocFailureALot || !should_inject_alloc_failure()) {
964 result = allocate_memory_under_lock(req, in_new_region);
965 }
966
967 // Check that gc overhead is not exceeded.
968 //
969 // Shenandoah will grind along for quite a while allocating one
970 // object at a time using shared (non-tlab) allocations. This check
971 // is testing that the GC overhead limit has not been exceeded.
972 // This will notify the collector to start a cycle, but will raise
973 // an OOME to the mutator if the last Full GCs have not made progress.
974 // gc_no_progress_count is incremented following each degen or full GC that fails to achieve is_good_progress().
975 if ((result == nullptr) && !req.is_lab_alloc() && (get_gc_no_progress_count() > ShenandoahNoProgressThreshold)) {
976 control_thread()->handle_alloc_failure(req, false);
977 req.set_actual_size(0);
978 return nullptr;
979 }
980
981 if (result == nullptr) {
982 // Block until control thread reacted, then retry allocation.
983 //
984 // It might happen that one of the threads requesting allocation would unblock
985 // way later after GC happened, only to fail the second allocation, because
986 // other threads have already depleted the free storage. In this case, a better
987 // strategy is to try again, until at least one full GC has completed.
988 //
989 // Stop retrying and return nullptr to cause OOMError exception if our allocation failed even after:
990 // a) We experienced a GC that had good progress, or
991 // b) We experienced at least one Full GC (whether or not it had good progress)
992
993 size_t original_count = shenandoah_policy()->full_gc_count();
994 while ((result == nullptr) && (original_count == shenandoah_policy()->full_gc_count())) {
995 control_thread()->handle_alloc_failure(req, true);
996 result = allocate_memory_under_lock(req, in_new_region);
997 }
998 if (result != nullptr) {
999 // If our allocation request has been satisifed after it initially failed, we count this as good gc progress
1000 notify_gc_progress();
1001 }
1002 if (log_develop_is_enabled(Debug, gc, alloc)) {
1003 ResourceMark rm;
1004 log_debug(gc, alloc)("Thread: %s, Result: " PTR_FORMAT ", Request: %s, Size: " SIZE_FORMAT
1005 ", Original: " SIZE_FORMAT ", Latest: " SIZE_FORMAT,
1006 Thread::current()->name(), p2i(result), req.type_string(), req.size(),
1007 original_count, get_gc_no_progress_count());
1008 }
1009 }
1010 } else {
1011 assert(req.is_gc_alloc(), "Can only accept GC allocs here");
1012 result = allocate_memory_under_lock(req, in_new_region);
1013 // Do not call handle_alloc_failure() here, because we cannot block.
1014 // The allocation failure would be handled by the LRB slowpath with handle_alloc_failure_evac().
1015 }
1016
1017 if (in_new_region) {
1018 notify_heap_changed();
1019 }
1020
1021 if (result == nullptr) {
1022 req.set_actual_size(0);
1023 }
1024
1025 // This is called regardless of the outcome of the allocation to account
1026 // for any waste created by retiring regions with this request.
1027 increase_used(req);
1028
1029 if (result != nullptr) {
1030 size_t requested = req.size();
1031 size_t actual = req.actual_size();
1032
1033 assert (req.is_lab_alloc() || (requested == actual),
1034 "Only LAB allocations are elastic: %s, requested = " SIZE_FORMAT ", actual = " SIZE_FORMAT,
1035 ShenandoahAllocRequest::alloc_type_to_string(req.type()), requested, actual);
1036
1037 if (req.is_mutator_alloc()) {
1038 // If we requested more than we were granted, give the rest back to pacer.
1039 // This only matters if we are in the same pacing epoch: do not try to unpace
1040 // over the budget for the other phase.
1041 if (ShenandoahPacing && (pacer_epoch > 0) && (requested > actual)) {
1042 pacer()->unpace_for_alloc(pacer_epoch, requested - actual);
1043 }
1044 }
1045 }
1046
1047 return result;
1048 }
1049
1050 HeapWord* ShenandoahHeap::allocate_memory_under_lock(ShenandoahAllocRequest& req, bool& in_new_region) {
1051 // If we are dealing with mutator allocation, then we may need to block for safepoint.
1052 // We cannot block for safepoint for GC allocations, because there is a high chance
1053 // we are already running at safepoint or from stack watermark machinery, and we cannot
1054 // block again.
1055 ShenandoahHeapLocker locker(lock(), req.is_mutator_alloc());
1056
1057 // Make sure the old generation has room for either evacuations or promotions before trying to allocate.
1058 if (req.is_old() && !old_generation()->can_allocate(req)) {
1059 return nullptr;
1060 }
1061
1062 // If TLAB request size is greater than available, allocate() will attempt to downsize request to fit within available
1063 // memory.
1064 HeapWord* result = _free_set->allocate(req, in_new_region);
1065
1066 // Record the plab configuration for this result and register the object.
1067 if (result != nullptr && req.is_old()) {
1068 old_generation()->configure_plab_for_current_thread(req);
1069 if (req.type() == ShenandoahAllocRequest::_alloc_shared_gc) {
1070 // Register the newly allocated object while we're holding the global lock since there's no synchronization
1071 // built in to the implementation of register_object(). There are potential races when multiple independent
1072 // threads are allocating objects, some of which might span the same card region. For example, consider
1073 // a card table's memory region within which three objects are being allocated by three different threads:
1074 //
1075 // objects being "concurrently" allocated:
1076 // [-----a------][-----b-----][--------------c------------------]
1077 // [---- card table memory range --------------]
1078 //
1079 // Before any objects are allocated, this card's memory range holds no objects. Note that allocation of object a
1080 // wants to set the starts-object, first-start, and last-start attributes of the preceding card region.
1081 // Allocation of object b wants to set the starts-object, first-start, and last-start attributes of this card region.
1082 // Allocation of object c also wants to set the starts-object, first-start, and last-start attributes of this
1083 // card region.
1084 //
1085 // The thread allocating b and the thread allocating c can "race" in various ways, resulting in confusion, such as
1086 // last-start representing object b while first-start represents object c. This is why we need to require all
1087 // register_object() invocations to be "mutually exclusive" with respect to each card's memory range.
1088 old_generation()->card_scan()->register_object(result);
1089 }
1090 }
1091
1092 return result;
1093 }
1094
1095 HeapWord* ShenandoahHeap::mem_allocate(size_t size,
1096 bool* gc_overhead_limit_was_exceeded) {
1097 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared(size);
1098 return allocate_memory(req);
1099 }
1100
1101 MetaWord* ShenandoahHeap::satisfy_failed_metadata_allocation(ClassLoaderData* loader_data,
1102 size_t size,
1103 Metaspace::MetadataType mdtype) {
1104 MetaWord* result;
1105
1106 // Inform metaspace OOM to GC heuristics if class unloading is possible.
1107 ShenandoahHeuristics* h = global_generation()->heuristics();
1108 if (h->can_unload_classes()) {
1109 h->record_metaspace_oom();
1110 }
1111
1112 // Expand and retry allocation
1113 result = loader_data->metaspace_non_null()->expand_and_allocate(size, mdtype);
1114 if (result != nullptr) {
1115 return result;
1116 }
1117
1118 // Start full GC
1119 collect(GCCause::_metadata_GC_clear_soft_refs);
1120
1121 // Retry allocation
1122 result = loader_data->metaspace_non_null()->allocate(size, mdtype);
1123 if (result != nullptr) {
1124 return result;
1125 }
1126
1127 // Expand and retry allocation
1128 result = loader_data->metaspace_non_null()->expand_and_allocate(size, mdtype);
1129 if (result != nullptr) {
1130 return result;
1131 }
1132
1133 // Out of memory
1134 return nullptr;
1135 }
1136
1137 class ShenandoahConcurrentEvacuateRegionObjectClosure : public ObjectClosure {
1138 private:
1139 ShenandoahHeap* const _heap;
1140 Thread* const _thread;
1141 public:
1142 ShenandoahConcurrentEvacuateRegionObjectClosure(ShenandoahHeap* heap) :
1143 _heap(heap), _thread(Thread::current()) {}
1144
1145 void do_object(oop p) {
1146 shenandoah_assert_marked(nullptr, p);
1147 if (!p->is_forwarded()) {
1148 _heap->evacuate_object(p, _thread);
1149 }
1150 }
1151 };
1152
1153 class ShenandoahEvacuationTask : public WorkerTask {
1154 private:
1155 ShenandoahHeap* const _sh;
1156 ShenandoahCollectionSet* const _cs;
1157 bool _concurrent;
1158 public:
1159 ShenandoahEvacuationTask(ShenandoahHeap* sh,
1160 ShenandoahCollectionSet* cs,
1161 bool concurrent) :
1162 WorkerTask("Shenandoah Evacuation"),
1163 _sh(sh),
1164 _cs(cs),
1165 _concurrent(concurrent)
1166 {}
1167
1168 void work(uint worker_id) {
1169 if (_concurrent) {
1170 ShenandoahConcurrentWorkerSession worker_session(worker_id);
1171 ShenandoahSuspendibleThreadSetJoiner stsj;
1172 ShenandoahEvacOOMScope oom_evac_scope;
1173 do_work();
1174 } else {
1175 ShenandoahParallelWorkerSession worker_session(worker_id);
1176 ShenandoahEvacOOMScope oom_evac_scope;
1177 do_work();
1178 }
1179 }
1180
1181 private:
1182 void do_work() {
1183 ShenandoahConcurrentEvacuateRegionObjectClosure cl(_sh);
1184 ShenandoahHeapRegion* r;
1185 while ((r =_cs->claim_next()) != nullptr) {
1186 assert(r->has_live(), "Region " SIZE_FORMAT " should have been reclaimed early", r->index());
1187 _sh->marked_object_iterate(r, &cl);
1188
1189 if (ShenandoahPacing) {
1190 _sh->pacer()->report_evac(r->used() >> LogHeapWordSize);
1191 }
1192
1193 if (_sh->check_cancelled_gc_and_yield(_concurrent)) {
1194 break;
1195 }
1196 }
1197 }
1198 };
1199
1200 void ShenandoahHeap::evacuate_collection_set(bool concurrent) {
1201 ShenandoahEvacuationTask task(this, _collection_set, concurrent);
1202 workers()->run_task(&task);
1203 }
1204
1205 oop ShenandoahHeap::evacuate_object(oop p, Thread* thread) {
1206 assert(thread == Thread::current(), "Expected thread parameter to be current thread.");
1207 if (ShenandoahThreadLocalData::is_oom_during_evac(thread)) {
1208 // This thread went through the OOM during evac protocol. It is safe to return
1209 // the forward pointer. It must not attempt to evacuate any other objects.
1210 return ShenandoahBarrierSet::resolve_forwarded(p);
1211 }
1212
1213 assert(ShenandoahThreadLocalData::is_evac_allowed(thread), "must be enclosed in oom-evac scope");
1214
1215 ShenandoahHeapRegion* r = heap_region_containing(p);
1216 assert(!r->is_humongous(), "never evacuate humongous objects");
1217
1218 ShenandoahAffiliation target_gen = r->affiliation();
1219 return try_evacuate_object(p, thread, r, target_gen);
1220 }
1221
1222 oop ShenandoahHeap::try_evacuate_object(oop p, Thread* thread, ShenandoahHeapRegion* from_region,
1223 ShenandoahAffiliation target_gen) {
1224 assert(target_gen == YOUNG_GENERATION, "Only expect evacuations to young in this mode");
1225 assert(from_region->is_young(), "Only expect evacuations from young in this mode");
1226 bool alloc_from_lab = true;
1227 HeapWord* copy = nullptr;
1228 size_t size = p->size();
1229
1230 #ifdef ASSERT
1231 if (ShenandoahOOMDuringEvacALot &&
1232 (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call
1233 copy = nullptr;
1234 } else {
1235 #endif
1236 if (UseTLAB) {
1237 copy = allocate_from_gclab(thread, size);
1238 }
1239 if (copy == nullptr) {
1240 // If we failed to allocate in LAB, we'll try a shared allocation.
1241 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size, target_gen);
1242 copy = allocate_memory(req);
1243 alloc_from_lab = false;
1244 }
1245 #ifdef ASSERT
1246 }
1247 #endif
1248
1249 if (copy == nullptr) {
1250 control_thread()->handle_alloc_failure_evac(size);
1251
1252 _oom_evac_handler.handle_out_of_memory_during_evacuation();
1253
1254 return ShenandoahBarrierSet::resolve_forwarded(p);
1255 }
1256
1257 // Copy the object:
1258 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(p), copy, size);
1259
1260 // Try to install the new forwarding pointer.
1261 oop copy_val = cast_to_oop(copy);
1262 oop result = ShenandoahForwarding::try_update_forwardee(p, copy_val);
1263 if (result == copy_val) {
1264 // Successfully evacuated. Our copy is now the public one!
1265 ContinuationGCSupport::relativize_stack_chunk(copy_val);
1266 shenandoah_assert_correct(nullptr, copy_val);
1267 return copy_val;
1268 } else {
1269 // Failed to evacuate. We need to deal with the object that is left behind. Since this
1270 // new allocation is certainly after TAMS, it will be considered live in the next cycle.
1271 // But if it happens to contain references to evacuated regions, those references would
1272 // not get updated for this stale copy during this cycle, and we will crash while scanning
1273 // it the next cycle.
1274 if (alloc_from_lab) {
1275 // For LAB allocations, it is enough to rollback the allocation ptr. Either the next
1276 // object will overwrite this stale copy, or the filler object on LAB retirement will
1277 // do this.
1278 ShenandoahThreadLocalData::gclab(thread)->undo_allocation(copy, size);
1279 } else {
1280 // For non-LAB allocations, we have no way to retract the allocation, and
1281 // have to explicitly overwrite the copy with the filler object. With that overwrite,
1282 // we have to keep the fwdptr initialized and pointing to our (stale) copy.
1283 assert(size >= ShenandoahHeap::min_fill_size(), "previously allocated object known to be larger than min_size");
1284 fill_with_object(copy, size);
1285 shenandoah_assert_correct(nullptr, copy_val);
1286 // For non-LAB allocations, the object has already been registered
1287 }
1288 shenandoah_assert_correct(nullptr, result);
1289 return result;
1290 }
1291 }
1292
1293 void ShenandoahHeap::trash_cset_regions() {
1294 ShenandoahHeapLocker locker(lock());
1295
1296 ShenandoahCollectionSet* set = collection_set();
1297 ShenandoahHeapRegion* r;
1298 set->clear_current_index();
1299 while ((r = set->next()) != nullptr) {
1300 r->make_trash();
1301 }
1302 collection_set()->clear();
1303 }
1304
1305 void ShenandoahHeap::print_heap_regions_on(outputStream* st) const {
1306 st->print_cr("Heap Regions:");
1307 st->print_cr("Region state: EU=empty-uncommitted, EC=empty-committed, R=regular, H=humongous start, HP=pinned humongous start");
1308 st->print_cr(" HC=humongous continuation, CS=collection set, TR=trash, P=pinned, CSP=pinned collection set");
1309 st->print_cr("BTE=bottom/top/end, TAMS=top-at-mark-start");
1310 st->print_cr("UWM=update watermark, U=used");
1311 st->print_cr("T=TLAB allocs, G=GCLAB allocs");
1312 st->print_cr("S=shared allocs, L=live data");
1313 st->print_cr("CP=critical pins");
1314
1315 for (size_t i = 0; i < num_regions(); i++) {
1316 get_region(i)->print_on(st);
1317 }
1318 }
1319
1320 size_t ShenandoahHeap::trash_humongous_region_at(ShenandoahHeapRegion* start) {
1321 assert(start->is_humongous_start(), "reclaim regions starting with the first one");
1322
1323 oop humongous_obj = cast_to_oop(start->bottom());
1324 size_t size = humongous_obj->size();
1325 size_t required_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize);
1326 size_t index = start->index() + required_regions - 1;
1327
1328 assert(!start->has_live(), "liveness must be zero");
1329
1330 for(size_t i = 0; i < required_regions; i++) {
1331 // Reclaim from tail. Otherwise, assertion fails when printing region to trace log,
1332 // as it expects that every region belongs to a humongous region starting with a humongous start region.
1333 ShenandoahHeapRegion* region = get_region(index --);
1334
1335 assert(region->is_humongous(), "expect correct humongous start or continuation");
1336 assert(!region->is_cset(), "Humongous region should not be in collection set");
1337
1338 region->make_trash_immediate();
1339 }
1340 return required_regions;
1341 }
1342
1343 class ShenandoahCheckCleanGCLABClosure : public ThreadClosure {
1344 public:
1345 ShenandoahCheckCleanGCLABClosure() {}
1346 void do_thread(Thread* thread) {
1347 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
1348 assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name());
1349 assert(gclab->words_remaining() == 0, "GCLAB should not need retirement");
1350
1351 if (ShenandoahHeap::heap()->mode()->is_generational()) {
1352 PLAB* plab = ShenandoahThreadLocalData::plab(thread);
1353 assert(plab != nullptr, "PLAB should be initialized for %s", thread->name());
1354 assert(plab->words_remaining() == 0, "PLAB should not need retirement");
1355 }
1356 }
1357 };
1358
1359 class ShenandoahRetireGCLABClosure : public ThreadClosure {
1360 private:
1361 bool const _resize;
1362 public:
1363 ShenandoahRetireGCLABClosure(bool resize) : _resize(resize) {}
1364 void do_thread(Thread* thread) {
1365 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
1366 assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name());
1367 gclab->retire();
1368 if (_resize && ShenandoahThreadLocalData::gclab_size(thread) > 0) {
1369 ShenandoahThreadLocalData::set_gclab_size(thread, 0);
1370 }
1371
1372 if (ShenandoahHeap::heap()->mode()->is_generational()) {
1373 PLAB* plab = ShenandoahThreadLocalData::plab(thread);
1374 assert(plab != nullptr, "PLAB should be initialized for %s", thread->name());
1375
1376 // There are two reasons to retire all plabs between old-gen evacuation passes.
1377 // 1. We need to make the plab memory parsable by remembered-set scanning.
1378 // 2. We need to establish a trustworthy UpdateWaterMark value within each old-gen heap region
1379 ShenandoahGenerationalHeap::heap()->retire_plab(plab, thread);
1380 if (_resize && ShenandoahThreadLocalData::plab_size(thread) > 0) {
1381 ShenandoahThreadLocalData::set_plab_size(thread, 0);
1382 }
1383 }
1384 }
1385 };
1386
1387 void ShenandoahHeap::labs_make_parsable() {
1388 assert(UseTLAB, "Only call with UseTLAB");
1389
1390 ShenandoahRetireGCLABClosure cl(false);
1391
1392 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1393 ThreadLocalAllocBuffer& tlab = t->tlab();
1394 tlab.make_parsable();
1395 cl.do_thread(t);
1396 }
1397
1398 workers()->threads_do(&cl);
1399 }
1400
1401 void ShenandoahHeap::tlabs_retire(bool resize) {
1402 assert(UseTLAB, "Only call with UseTLAB");
1403 assert(!resize || ResizeTLAB, "Only call for resize when ResizeTLAB is enabled");
1404
1405 ThreadLocalAllocStats stats;
1406
1407 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1408 ThreadLocalAllocBuffer& tlab = t->tlab();
1409 tlab.retire(&stats);
1410 if (resize) {
1411 tlab.resize();
1412 }
1413 }
1414
1415 stats.publish();
1416
1417 #ifdef ASSERT
1418 ShenandoahCheckCleanGCLABClosure cl;
1419 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1420 cl.do_thread(t);
1421 }
1422 workers()->threads_do(&cl);
1423 #endif
1424 }
1425
1426 void ShenandoahHeap::gclabs_retire(bool resize) {
1427 assert(UseTLAB, "Only call with UseTLAB");
1428 assert(!resize || ResizeTLAB, "Only call for resize when ResizeTLAB is enabled");
1429
1430 ShenandoahRetireGCLABClosure cl(resize);
1431 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1432 cl.do_thread(t);
1433 }
1434 workers()->threads_do(&cl);
1435
1436 if (safepoint_workers() != nullptr) {
1437 safepoint_workers()->threads_do(&cl);
1438 }
1439 }
1440
1441 // Returns size in bytes
1442 size_t ShenandoahHeap::unsafe_max_tlab_alloc(Thread *thread) const {
1443 // Return the max allowed size, and let the allocation path
1444 // figure out the safe size for current allocation.
1445 return ShenandoahHeapRegion::max_tlab_size_bytes();
1446 }
1447
1448 size_t ShenandoahHeap::max_tlab_size() const {
1449 // Returns size in words
1450 return ShenandoahHeapRegion::max_tlab_size_words();
1451 }
1452
1453 void ShenandoahHeap::collect_as_vm_thread(GCCause::Cause cause) {
1454 // These requests are ignored because we can't easily have Shenandoah jump into
1455 // a synchronous (degenerated or full) cycle while it is in the middle of a concurrent
1456 // cycle. We _could_ cancel the concurrent cycle and then try to run a cycle directly
1457 // on the VM thread, but this would confuse the control thread mightily and doesn't
1458 // seem worth the trouble. Instead, we will have the caller thread run (and wait for) a
1459 // concurrent cycle in the prologue of the heap inspect/dump operation. This is how
1460 // other concurrent collectors in the JVM handle this scenario as well.
1461 assert(Thread::current()->is_VM_thread(), "Should be the VM thread");
1462 guarantee(cause == GCCause::_heap_dump || cause == GCCause::_heap_inspection, "Invalid cause");
1463 }
1464
1465 void ShenandoahHeap::collect(GCCause::Cause cause) {
1466 control_thread()->request_gc(cause);
1467 }
1468
1469 void ShenandoahHeap::do_full_collection(bool clear_all_soft_refs) {
1470 //assert(false, "Shouldn't need to do full collections");
1471 }
1472
1473 HeapWord* ShenandoahHeap::block_start(const void* addr) const {
1474 ShenandoahHeapRegion* r = heap_region_containing(addr);
1475 if (r != nullptr) {
1476 return r->block_start(addr);
1477 }
1478 return nullptr;
1479 }
1480
1481 bool ShenandoahHeap::block_is_obj(const HeapWord* addr) const {
1482 ShenandoahHeapRegion* r = heap_region_containing(addr);
1483 return r->block_is_obj(addr);
1484 }
1485
1486 bool ShenandoahHeap::print_location(outputStream* st, void* addr) const {
1487 return BlockLocationPrinter<ShenandoahHeap>::print_location(st, addr);
1488 }
1489
1490 void ShenandoahHeap::prepare_for_verify() {
1491 if (SafepointSynchronize::is_at_safepoint() && UseTLAB) {
1492 labs_make_parsable();
1493 }
1494 }
1495
1496 void ShenandoahHeap::gc_threads_do(ThreadClosure* tcl) const {
1497 if (_shenandoah_policy->is_at_shutdown()) {
1498 return;
1499 }
1500
1501 if (_control_thread != nullptr) {
1502 tcl->do_thread(_control_thread);
1503 }
1504
1505 workers()->threads_do(tcl);
1506 if (_safepoint_workers != nullptr) {
1507 _safepoint_workers->threads_do(tcl);
1508 }
1509 }
1510
1511 void ShenandoahHeap::print_tracing_info() const {
1512 LogTarget(Info, gc, stats) lt;
1513 if (lt.is_enabled()) {
1514 ResourceMark rm;
1515 LogStream ls(lt);
1516
1517 phase_timings()->print_global_on(&ls);
1518
1519 ls.cr();
1520 ls.cr();
1521
1522 shenandoah_policy()->print_gc_stats(&ls);
1523
1524 ls.cr();
1525 ls.cr();
1526 }
1527 }
1528
1529 void ShenandoahHeap::set_gc_generation(ShenandoahGeneration* generation) {
1530 shenandoah_assert_control_or_vm_thread_at_safepoint();
1531 _gc_generation = generation;
1532 }
1533
1534 // Active generation may only be set by the VM thread at a safepoint.
1535 void ShenandoahHeap::set_active_generation() {
1536 assert(Thread::current()->is_VM_thread(), "Only the VM Thread");
1537 assert(SafepointSynchronize::is_at_safepoint(), "Only at a safepoint!");
1538 assert(_gc_generation != nullptr, "Will set _active_generation to nullptr");
1539 _active_generation = _gc_generation;
1540 }
1541
1542 void ShenandoahHeap::on_cycle_start(GCCause::Cause cause, ShenandoahGeneration* generation) {
1543 shenandoah_policy()->record_collection_cause(cause);
1544
1545 const GCCause::Cause current = gc_cause();
1546 assert(current == GCCause::_no_gc, "Over-writing cause: %s, with: %s",
1547 GCCause::to_string(current), GCCause::to_string(cause));
1548 assert(_gc_generation == nullptr, "Over-writing _gc_generation");
1549
1550 set_gc_cause(cause);
1551 set_gc_generation(generation);
1552
1553 generation->heuristics()->record_cycle_start();
1554 }
1555
1556 void ShenandoahHeap::on_cycle_end(ShenandoahGeneration* generation) {
1557 assert(gc_cause() != GCCause::_no_gc, "cause wasn't set");
1558 assert(_gc_generation != nullptr, "_gc_generation wasn't set");
1559
1560 generation->heuristics()->record_cycle_end();
1561 if (mode()->is_generational() && generation->is_global()) {
1562 // If we just completed a GLOBAL GC, claim credit for completion of young-gen and old-gen GC as well
1563 young_generation()->heuristics()->record_cycle_end();
1564 old_generation()->heuristics()->record_cycle_end();
1565 }
1566
1567 set_gc_generation(nullptr);
1568 set_gc_cause(GCCause::_no_gc);
1569 }
1570
1571 void ShenandoahHeap::verify(VerifyOption vo) {
1572 if (ShenandoahSafepoint::is_at_shenandoah_safepoint()) {
1573 if (ShenandoahVerify) {
1574 verifier()->verify_generic(vo);
1575 } else {
1576 // TODO: Consider allocating verification bitmaps on demand,
1577 // and turn this on unconditionally.
1578 }
1579 }
1580 }
1581 size_t ShenandoahHeap::tlab_capacity(Thread *thr) const {
1582 return _free_set->capacity();
1583 }
1584
1585 class ObjectIterateScanRootClosure : public BasicOopIterateClosure {
1586 private:
1587 MarkBitMap* _bitmap;
1588 ShenandoahScanObjectStack* _oop_stack;
1589 ShenandoahHeap* const _heap;
1590 ShenandoahMarkingContext* const _marking_context;
1591
1592 template <class T>
1593 void do_oop_work(T* p) {
1594 T o = RawAccess<>::oop_load(p);
1595 if (!CompressedOops::is_null(o)) {
1596 oop obj = CompressedOops::decode_not_null(o);
1597 if (_heap->is_concurrent_weak_root_in_progress() && !_marking_context->is_marked(obj)) {
1598 // There may be dead oops in weak roots in concurrent root phase, do not touch them.
1599 return;
1600 }
1601 obj = ShenandoahBarrierSet::barrier_set()->load_reference_barrier(obj);
1602
1603 assert(oopDesc::is_oop(obj), "must be a valid oop");
1604 if (!_bitmap->is_marked(obj)) {
1605 _bitmap->mark(obj);
1606 _oop_stack->push(obj);
1607 }
1608 }
1609 }
1610 public:
1611 ObjectIterateScanRootClosure(MarkBitMap* bitmap, ShenandoahScanObjectStack* oop_stack) :
1612 _bitmap(bitmap), _oop_stack(oop_stack), _heap(ShenandoahHeap::heap()),
1613 _marking_context(_heap->marking_context()) {}
1614 void do_oop(oop* p) { do_oop_work(p); }
1615 void do_oop(narrowOop* p) { do_oop_work(p); }
1616 };
1617
1618 /*
1619 * This is public API, used in preparation of object_iterate().
1620 * Since we don't do linear scan of heap in object_iterate() (see comment below), we don't
1621 * need to make the heap parsable. For Shenandoah-internal linear heap scans that we can
1622 * control, we call SH::tlabs_retire, SH::gclabs_retire.
1623 */
1624 void ShenandoahHeap::ensure_parsability(bool retire_tlabs) {
1625 // No-op.
1626 }
1627
1628 /*
1629 * Iterates objects in the heap. This is public API, used for, e.g., heap dumping.
1630 *
1631 * We cannot safely iterate objects by doing a linear scan at random points in time. Linear
1632 * scanning needs to deal with dead objects, which may have dead Klass* pointers (e.g.
1633 * calling oopDesc::size() would crash) or dangling reference fields (crashes) etc. Linear
1634 * scanning therefore depends on having a valid marking bitmap to support it. However, we only
1635 * have a valid marking bitmap after successful marking. In particular, we *don't* have a valid
1636 * marking bitmap during marking, after aborted marking or during/after cleanup (when we just
1637 * wiped the bitmap in preparation for next marking).
1638 *
1639 * For all those reasons, we implement object iteration as a single marking traversal, reporting
1640 * objects as we mark+traverse through the heap, starting from GC roots. JVMTI IterateThroughHeap
1641 * is allowed to report dead objects, but is not required to do so.
1642 */
1643 void ShenandoahHeap::object_iterate(ObjectClosure* cl) {
1644 // Reset bitmap
1645 if (!prepare_aux_bitmap_for_iteration())
1646 return;
1647
1648 ShenandoahScanObjectStack oop_stack;
1649 ObjectIterateScanRootClosure oops(&_aux_bit_map, &oop_stack);
1650 // Seed the stack with root scan
1651 scan_roots_for_iteration(&oop_stack, &oops);
1652
1653 // Work through the oop stack to traverse heap
1654 while (! oop_stack.is_empty()) {
1655 oop obj = oop_stack.pop();
1656 assert(oopDesc::is_oop(obj), "must be a valid oop");
1657 cl->do_object(obj);
1658 obj->oop_iterate(&oops);
1659 }
1660
1661 assert(oop_stack.is_empty(), "should be empty");
1662 // Reclaim bitmap
1663 reclaim_aux_bitmap_for_iteration();
1664 }
1665
1666 bool ShenandoahHeap::prepare_aux_bitmap_for_iteration() {
1667 assert(SafepointSynchronize::is_at_safepoint(), "safe iteration is only available during safepoints");
1668
1669 if (!_aux_bitmap_region_special && !os::commit_memory((char*)_aux_bitmap_region.start(), _aux_bitmap_region.byte_size(), false)) {
1670 log_warning(gc)("Could not commit native memory for auxiliary marking bitmap for heap iteration");
1671 return false;
1672 }
1673 // Reset bitmap
1674 _aux_bit_map.clear();
1675 return true;
1676 }
1677
1678 void ShenandoahHeap::scan_roots_for_iteration(ShenandoahScanObjectStack* oop_stack, ObjectIterateScanRootClosure* oops) {
1679 // Process GC roots according to current GC cycle
1680 // This populates the work stack with initial objects
1681 // It is important to relinquish the associated locks before diving
1682 // into heap dumper
1683 uint n_workers = safepoint_workers() != nullptr ? safepoint_workers()->active_workers() : 1;
1684 ShenandoahHeapIterationRootScanner rp(n_workers);
1685 rp.roots_do(oops);
1686 }
1687
1688 void ShenandoahHeap::reclaim_aux_bitmap_for_iteration() {
1689 if (!_aux_bitmap_region_special && !os::uncommit_memory((char*)_aux_bitmap_region.start(), _aux_bitmap_region.byte_size())) {
1690 log_warning(gc)("Could not uncommit native memory for auxiliary marking bitmap for heap iteration");
1691 }
1692 }
1693
1694 // Closure for parallelly iterate objects
1695 class ShenandoahObjectIterateParScanClosure : public BasicOopIterateClosure {
1696 private:
1697 MarkBitMap* _bitmap;
1698 ShenandoahObjToScanQueue* _queue;
1699 ShenandoahHeap* const _heap;
1700 ShenandoahMarkingContext* const _marking_context;
1701
1702 template <class T>
1703 void do_oop_work(T* p) {
1704 T o = RawAccess<>::oop_load(p);
1705 if (!CompressedOops::is_null(o)) {
1706 oop obj = CompressedOops::decode_not_null(o);
1707 if (_heap->is_concurrent_weak_root_in_progress() && !_marking_context->is_marked(obj)) {
1708 // There may be dead oops in weak roots in concurrent root phase, do not touch them.
1709 return;
1710 }
1711 obj = ShenandoahBarrierSet::barrier_set()->load_reference_barrier(obj);
1712
1713 assert(oopDesc::is_oop(obj), "Must be a valid oop");
1714 if (_bitmap->par_mark(obj)) {
1715 _queue->push(ShenandoahMarkTask(obj));
1716 }
1717 }
1718 }
1719 public:
1720 ShenandoahObjectIterateParScanClosure(MarkBitMap* bitmap, ShenandoahObjToScanQueue* q) :
1721 _bitmap(bitmap), _queue(q), _heap(ShenandoahHeap::heap()),
1722 _marking_context(_heap->marking_context()) {}
1723 void do_oop(oop* p) { do_oop_work(p); }
1724 void do_oop(narrowOop* p) { do_oop_work(p); }
1725 };
1726
1727 // Object iterator for parallel heap iteraion.
1728 // The root scanning phase happenes in construction as a preparation of
1729 // parallel marking queues.
1730 // Every worker processes it's own marking queue. work-stealing is used
1731 // to balance workload.
1732 class ShenandoahParallelObjectIterator : public ParallelObjectIteratorImpl {
1733 private:
1734 uint _num_workers;
1735 bool _init_ready;
1736 MarkBitMap* _aux_bit_map;
1737 ShenandoahHeap* _heap;
1738 ShenandoahScanObjectStack _roots_stack; // global roots stack
1739 ShenandoahObjToScanQueueSet* _task_queues;
1740 public:
1741 ShenandoahParallelObjectIterator(uint num_workers, MarkBitMap* bitmap) :
1742 _num_workers(num_workers),
1743 _init_ready(false),
1744 _aux_bit_map(bitmap),
1745 _heap(ShenandoahHeap::heap()) {
1746 // Initialize bitmap
1747 _init_ready = _heap->prepare_aux_bitmap_for_iteration();
1748 if (!_init_ready) {
1749 return;
1750 }
1751
1752 ObjectIterateScanRootClosure oops(_aux_bit_map, &_roots_stack);
1753 _heap->scan_roots_for_iteration(&_roots_stack, &oops);
1754
1755 _init_ready = prepare_worker_queues();
1756 }
1757
1758 ~ShenandoahParallelObjectIterator() {
1759 // Reclaim bitmap
1760 _heap->reclaim_aux_bitmap_for_iteration();
1761 // Reclaim queue for workers
1762 if (_task_queues!= nullptr) {
1763 for (uint i = 0; i < _num_workers; ++i) {
1764 ShenandoahObjToScanQueue* q = _task_queues->queue(i);
1765 if (q != nullptr) {
1766 delete q;
1767 _task_queues->register_queue(i, nullptr);
1768 }
1769 }
1770 delete _task_queues;
1771 _task_queues = nullptr;
1772 }
1773 }
1774
1775 virtual void object_iterate(ObjectClosure* cl, uint worker_id) {
1776 if (_init_ready) {
1777 object_iterate_parallel(cl, worker_id, _task_queues);
1778 }
1779 }
1780
1781 private:
1782 // Divide global root_stack into worker queues
1783 bool prepare_worker_queues() {
1784 _task_queues = new ShenandoahObjToScanQueueSet((int) _num_workers);
1785 // Initialize queues for every workers
1786 for (uint i = 0; i < _num_workers; ++i) {
1787 ShenandoahObjToScanQueue* task_queue = new ShenandoahObjToScanQueue();
1788 _task_queues->register_queue(i, task_queue);
1789 }
1790 // Divide roots among the workers. Assume that object referencing distribution
1791 // is related with root kind, use round-robin to make every worker have same chance
1792 // to process every kind of roots
1793 size_t roots_num = _roots_stack.size();
1794 if (roots_num == 0) {
1795 // No work to do
1796 return false;
1797 }
1798
1799 for (uint j = 0; j < roots_num; j++) {
1800 uint stack_id = j % _num_workers;
1801 oop obj = _roots_stack.pop();
1802 _task_queues->queue(stack_id)->push(ShenandoahMarkTask(obj));
1803 }
1804 return true;
1805 }
1806
1807 void object_iterate_parallel(ObjectClosure* cl,
1808 uint worker_id,
1809 ShenandoahObjToScanQueueSet* queue_set) {
1810 assert(SafepointSynchronize::is_at_safepoint(), "safe iteration is only available during safepoints");
1811 assert(queue_set != nullptr, "task queue must not be null");
1812
1813 ShenandoahObjToScanQueue* q = queue_set->queue(worker_id);
1814 assert(q != nullptr, "object iterate queue must not be null");
1815
1816 ShenandoahMarkTask t;
1817 ShenandoahObjectIterateParScanClosure oops(_aux_bit_map, q);
1818
1819 // Work through the queue to traverse heap.
1820 // Steal when there is no task in queue.
1821 while (q->pop(t) || queue_set->steal(worker_id, t)) {
1822 oop obj = t.obj();
1823 assert(oopDesc::is_oop(obj), "must be a valid oop");
1824 cl->do_object(obj);
1825 obj->oop_iterate(&oops);
1826 }
1827 assert(q->is_empty(), "should be empty");
1828 }
1829 };
1830
1831 ParallelObjectIteratorImpl* ShenandoahHeap::parallel_object_iterator(uint workers) {
1832 return new ShenandoahParallelObjectIterator(workers, &_aux_bit_map);
1833 }
1834
1835 // Keep alive an object that was loaded with AS_NO_KEEPALIVE.
1836 void ShenandoahHeap::keep_alive(oop obj) {
1837 if (is_concurrent_mark_in_progress() && (obj != nullptr)) {
1838 ShenandoahBarrierSet::barrier_set()->enqueue(obj);
1839 }
1840 }
1841
1842 void ShenandoahHeap::heap_region_iterate(ShenandoahHeapRegionClosure* blk) const {
1843 for (size_t i = 0; i < num_regions(); i++) {
1844 ShenandoahHeapRegion* current = get_region(i);
1845 blk->heap_region_do(current);
1846 }
1847 }
1848
1849 class ShenandoahParallelHeapRegionTask : public WorkerTask {
1850 private:
1851 ShenandoahHeap* const _heap;
1852 ShenandoahHeapRegionClosure* const _blk;
1853 size_t const _stride;
1854
1855 shenandoah_padding(0);
1856 volatile size_t _index;
1857 shenandoah_padding(1);
1858
1859 public:
1860 ShenandoahParallelHeapRegionTask(ShenandoahHeapRegionClosure* blk, size_t stride) :
1861 WorkerTask("Shenandoah Parallel Region Operation"),
1862 _heap(ShenandoahHeap::heap()), _blk(blk), _stride(stride), _index(0) {}
1863
1864 void work(uint worker_id) {
1865 ShenandoahParallelWorkerSession worker_session(worker_id);
1866 size_t stride = _stride;
1867
1868 size_t max = _heap->num_regions();
1869 while (Atomic::load(&_index) < max) {
1870 size_t cur = Atomic::fetch_then_add(&_index, stride, memory_order_relaxed);
1871 size_t start = cur;
1872 size_t end = MIN2(cur + stride, max);
1873 if (start >= max) break;
1874
1875 for (size_t i = cur; i < end; i++) {
1876 ShenandoahHeapRegion* current = _heap->get_region(i);
1877 _blk->heap_region_do(current);
1878 }
1879 }
1880 }
1881 };
1882
1883 void ShenandoahHeap::parallel_heap_region_iterate(ShenandoahHeapRegionClosure* blk) const {
1884 assert(blk->is_thread_safe(), "Only thread-safe closures here");
1885 const uint active_workers = workers()->active_workers();
1886 const size_t n_regions = num_regions();
1887 size_t stride = ShenandoahParallelRegionStride;
1888 if (stride == 0 && active_workers > 1) {
1889 // Automatically derive the stride to balance the work between threads
1890 // evenly. Do not try to split work if below the reasonable threshold.
1891 constexpr size_t threshold = 4096;
1892 stride = n_regions <= threshold ?
1893 threshold :
1894 (n_regions + active_workers - 1) / active_workers;
1895 }
1896
1897 if (n_regions > stride && active_workers > 1) {
1898 ShenandoahParallelHeapRegionTask task(blk, stride);
1899 workers()->run_task(&task);
1900 } else {
1901 heap_region_iterate(blk);
1902 }
1903 }
1904
1905 class ShenandoahRendezvousClosure : public HandshakeClosure {
1906 public:
1907 inline ShenandoahRendezvousClosure() : HandshakeClosure("ShenandoahRendezvous") {}
1908 inline void do_thread(Thread* thread) {}
1909 };
1910
1911 void ShenandoahHeap::rendezvous_threads() {
1912 ShenandoahRendezvousClosure cl;
1913 Handshake::execute(&cl);
1914 }
1915
1916 void ShenandoahHeap::recycle_trash() {
1917 free_set()->recycle_trash();
1918 }
1919
1920 void ShenandoahHeap::do_class_unloading() {
1921 _unloader.unload();
1922 if (mode()->is_generational()) {
1923 old_generation()->set_parsable(false);
1924 }
1925 }
1926
1927 void ShenandoahHeap::stw_weak_refs(bool full_gc) {
1928 // Weak refs processing
1929 ShenandoahPhaseTimings::Phase phase = full_gc ? ShenandoahPhaseTimings::full_gc_weakrefs
1930 : ShenandoahPhaseTimings::degen_gc_weakrefs;
1931 ShenandoahTimingsTracker t(phase);
1932 ShenandoahGCWorkerPhase worker_phase(phase);
1933 shenandoah_assert_generations_reconciled();
1934 gc_generation()->ref_processor()->process_references(phase, workers(), false /* concurrent */);
1935 }
1936
1937 void ShenandoahHeap::prepare_update_heap_references(bool concurrent) {
1938 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at safepoint");
1939
1940 // Evacuation is over, no GCLABs are needed anymore. GCLABs are under URWM, so we need to
1941 // make them parsable for update code to work correctly. Plus, we can compute new sizes
1942 // for future GCLABs here.
1943 if (UseTLAB) {
1944 ShenandoahGCPhase phase(concurrent ?
1945 ShenandoahPhaseTimings::init_update_refs_manage_gclabs :
1946 ShenandoahPhaseTimings::degen_gc_init_update_refs_manage_gclabs);
1947 gclabs_retire(ResizeTLAB);
1948 }
1949
1950 _update_refs_iterator.reset();
1951 }
1952
1953 void ShenandoahHeap::propagate_gc_state_to_java_threads() {
1954 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Must be at Shenandoah safepoint");
1955 if (_gc_state_changed) {
1956 _gc_state_changed = false;
1957 char state = gc_state();
1958 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1959 ShenandoahThreadLocalData::set_gc_state(t, state);
1960 }
1961 }
1962 }
1963
1964 void ShenandoahHeap::set_gc_state(uint mask, bool value) {
1965 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Must be at Shenandoah safepoint");
1966 _gc_state.set_cond(mask, value);
1967 _gc_state_changed = true;
1968 }
1969
1970 void ShenandoahHeap::set_concurrent_young_mark_in_progress(bool in_progress) {
1971 uint mask;
1972 assert(!has_forwarded_objects(), "Young marking is not concurrent with evacuation");
1973 if (!in_progress && is_concurrent_old_mark_in_progress()) {
1974 assert(mode()->is_generational(), "Only generational GC has old marking");
1975 assert(_gc_state.is_set(MARKING), "concurrent_old_marking_in_progress implies MARKING");
1976 // If old-marking is in progress when we turn off YOUNG_MARKING, leave MARKING (and OLD_MARKING) on
1977 mask = YOUNG_MARKING;
1978 } else {
1979 mask = MARKING | YOUNG_MARKING;
1980 }
1981 set_gc_state(mask, in_progress);
1982 manage_satb_barrier(in_progress);
1983 }
1984
1985 void ShenandoahHeap::set_concurrent_old_mark_in_progress(bool in_progress) {
1986 #ifdef ASSERT
1987 // has_forwarded_objects() iff UPDATEREFS or EVACUATION
1988 bool has_forwarded = has_forwarded_objects();
1989 bool updating_or_evacuating = _gc_state.is_set(UPDATEREFS | EVACUATION);
1990 bool evacuating = _gc_state.is_set(EVACUATION);
1991 assert ((has_forwarded == updating_or_evacuating) || (evacuating && !has_forwarded && collection_set()->is_empty()),
1992 "Updating or evacuating iff has forwarded objects, or if evacuation phase is promoting in place without forwarding");
1993 #endif
1994 if (!in_progress && is_concurrent_young_mark_in_progress()) {
1995 // If young-marking is in progress when we turn off OLD_MARKING, leave MARKING (and YOUNG_MARKING) on
1996 assert(_gc_state.is_set(MARKING), "concurrent_young_marking_in_progress implies MARKING");
1997 set_gc_state(OLD_MARKING, in_progress);
1998 } else {
1999 set_gc_state(MARKING | OLD_MARKING, in_progress);
2000 }
2001 manage_satb_barrier(in_progress);
2002 }
2003
2004 bool ShenandoahHeap::is_prepare_for_old_mark_in_progress() const {
2005 return old_generation()->is_preparing_for_mark();
2006 }
2007
2008 void ShenandoahHeap::manage_satb_barrier(bool active) {
2009 if (is_concurrent_mark_in_progress()) {
2010 // Ignore request to deactivate barrier while concurrent mark is in progress.
2011 // Do not attempt to re-activate the barrier if it is already active.
2012 if (active && !ShenandoahBarrierSet::satb_mark_queue_set().is_active()) {
2013 ShenandoahBarrierSet::satb_mark_queue_set().set_active_all_threads(active, !active);
2014 }
2015 } else {
2016 // No concurrent marking is in progress so honor request to deactivate,
2017 // but only if the barrier is already active.
2018 if (!active && ShenandoahBarrierSet::satb_mark_queue_set().is_active()) {
2019 ShenandoahBarrierSet::satb_mark_queue_set().set_active_all_threads(active, !active);
2020 }
2021 }
2022 }
2023
2024 void ShenandoahHeap::set_evacuation_in_progress(bool in_progress) {
2025 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Only call this at safepoint");
2026 set_gc_state(EVACUATION, in_progress);
2027 }
2028
2029 void ShenandoahHeap::set_concurrent_strong_root_in_progress(bool in_progress) {
2030 if (in_progress) {
2031 _concurrent_strong_root_in_progress.set();
2032 } else {
2033 _concurrent_strong_root_in_progress.unset();
2034 }
2035 }
2036
2037 void ShenandoahHeap::set_concurrent_weak_root_in_progress(bool cond) {
2038 set_gc_state(WEAK_ROOTS, cond);
2039 }
2040
2041 GCTracer* ShenandoahHeap::tracer() {
2042 return shenandoah_policy()->tracer();
2043 }
2044
2045 size_t ShenandoahHeap::tlab_used(Thread* thread) const {
2046 return _free_set->used();
2047 }
2048
2049 bool ShenandoahHeap::try_cancel_gc() {
2050 jbyte prev = _cancelled_gc.cmpxchg(CANCELLED, CANCELLABLE);
2051 return prev == CANCELLABLE;
2052 }
2053
2054 void ShenandoahHeap::cancel_concurrent_mark() {
2055 if (mode()->is_generational()) {
2056 young_generation()->cancel_marking();
2057 old_generation()->cancel_marking();
2058 }
2059
2060 global_generation()->cancel_marking();
2061
2062 ShenandoahBarrierSet::satb_mark_queue_set().abandon_partial_marking();
2063 }
2064
2065 void ShenandoahHeap::cancel_gc(GCCause::Cause cause) {
2066 if (try_cancel_gc()) {
2067 FormatBuffer<> msg("Cancelling GC: %s", GCCause::to_string(cause));
2068 log_info(gc)("%s", msg.buffer());
2069 Events::log(Thread::current(), "%s", msg.buffer());
2070 _cancel_requested_time = os::elapsedTime();
2071 }
2072 }
2073
2074 uint ShenandoahHeap::max_workers() {
2075 return _max_workers;
2076 }
2077
2078 void ShenandoahHeap::stop() {
2079 // The shutdown sequence should be able to terminate when GC is running.
2080
2081 // Step 0. Notify policy to disable event recording.
2082 _shenandoah_policy->record_shutdown();
2083
2084 // Step 0a. Stop reporting on gc thread cpu utilization
2085 mmu_tracker()->stop();
2086
2087 // Step 1. Notify control thread that we are in shutdown.
2088 // Note that we cannot do that with stop(), because stop() is blocking and waits for the actual shutdown.
2089 // Doing stop() here would wait for the normal GC cycle to complete, never falling through to cancel below.
2090 control_thread()->prepare_for_graceful_shutdown();
2091
2092 // Step 2. Notify GC workers that we are cancelling GC.
2093 cancel_gc(GCCause::_shenandoah_stop_vm);
2094
2095 // Step 3. Wait until GC worker exits normally.
2096 control_thread()->stop();
2097 }
2098
2099 void ShenandoahHeap::stw_unload_classes(bool full_gc) {
2100 if (!unload_classes()) return;
2101 ClassUnloadingContext ctx(_workers->active_workers(),
2102 true /* unregister_nmethods_during_purge */,
2103 false /* lock_codeblob_free_separately */);
2104
2105 // Unload classes and purge SystemDictionary.
2106 {
2107 ShenandoahPhaseTimings::Phase phase = full_gc ?
2108 ShenandoahPhaseTimings::full_gc_purge_class_unload :
2109 ShenandoahPhaseTimings::degen_gc_purge_class_unload;
2110 ShenandoahIsAliveSelector is_alive;
2111 {
2112 CodeCache::UnlinkingScope scope(is_alive.is_alive_closure());
2113 ShenandoahGCPhase gc_phase(phase);
2114 ShenandoahGCWorkerPhase worker_phase(phase);
2115 bool unloading_occurred = SystemDictionary::do_unloading(gc_timer());
2116
2117 uint num_workers = _workers->active_workers();
2118 ShenandoahClassUnloadingTask unlink_task(phase, num_workers, unloading_occurred);
2119 _workers->run_task(&unlink_task);
2120 }
2121 // Release unloaded nmethods's memory.
2122 ClassUnloadingContext::context()->purge_and_free_nmethods();
2123 }
2124
2125 {
2126 ShenandoahGCPhase phase(full_gc ?
2127 ShenandoahPhaseTimings::full_gc_purge_cldg :
2128 ShenandoahPhaseTimings::degen_gc_purge_cldg);
2129 ClassLoaderDataGraph::purge(true /* at_safepoint */);
2130 }
2131 // Resize and verify metaspace
2132 MetaspaceGC::compute_new_size();
2133 DEBUG_ONLY(MetaspaceUtils::verify();)
2134 }
2135
2136 // Weak roots are either pre-evacuated (final mark) or updated (final updaterefs),
2137 // so they should not have forwarded oops.
2138 // However, we do need to "null" dead oops in the roots, if can not be done
2139 // in concurrent cycles.
2140 void ShenandoahHeap::stw_process_weak_roots(bool full_gc) {
2141 uint num_workers = _workers->active_workers();
2142 ShenandoahPhaseTimings::Phase timing_phase = full_gc ?
2143 ShenandoahPhaseTimings::full_gc_purge_weak_par :
2144 ShenandoahPhaseTimings::degen_gc_purge_weak_par;
2145 ShenandoahGCPhase phase(timing_phase);
2146 ShenandoahGCWorkerPhase worker_phase(timing_phase);
2147 // Cleanup weak roots
2148 if (has_forwarded_objects()) {
2149 ShenandoahForwardedIsAliveClosure is_alive;
2150 ShenandoahUpdateRefsClosure keep_alive;
2151 ShenandoahParallelWeakRootsCleaningTask<ShenandoahForwardedIsAliveClosure, ShenandoahUpdateRefsClosure>
2152 cleaning_task(timing_phase, &is_alive, &keep_alive, num_workers);
2153 _workers->run_task(&cleaning_task);
2154 } else {
2155 ShenandoahIsAliveClosure is_alive;
2156 #ifdef ASSERT
2157 ShenandoahAssertNotForwardedClosure verify_cl;
2158 ShenandoahParallelWeakRootsCleaningTask<ShenandoahIsAliveClosure, ShenandoahAssertNotForwardedClosure>
2159 cleaning_task(timing_phase, &is_alive, &verify_cl, num_workers);
2160 #else
2161 ShenandoahParallelWeakRootsCleaningTask<ShenandoahIsAliveClosure, DoNothingClosure>
2162 cleaning_task(timing_phase, &is_alive, &do_nothing_cl, num_workers);
2163 #endif
2164 _workers->run_task(&cleaning_task);
2165 }
2166 }
2167
2168 void ShenandoahHeap::parallel_cleaning(bool full_gc) {
2169 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint");
2170 assert(is_stw_gc_in_progress(), "Only for Degenerated and Full GC");
2171 ShenandoahGCPhase phase(full_gc ?
2172 ShenandoahPhaseTimings::full_gc_purge :
2173 ShenandoahPhaseTimings::degen_gc_purge);
2174 stw_weak_refs(full_gc);
2175 stw_process_weak_roots(full_gc);
2176 stw_unload_classes(full_gc);
2177 }
2178
2179 void ShenandoahHeap::set_has_forwarded_objects(bool cond) {
2180 set_gc_state(HAS_FORWARDED, cond);
2181 }
2182
2183 void ShenandoahHeap::set_unload_classes(bool uc) {
2184 _unload_classes.set_cond(uc);
2185 }
2186
2187 bool ShenandoahHeap::unload_classes() const {
2188 return _unload_classes.is_set();
2189 }
2190
2191 address ShenandoahHeap::in_cset_fast_test_addr() {
2192 ShenandoahHeap* heap = ShenandoahHeap::heap();
2193 assert(heap->collection_set() != nullptr, "Sanity");
2194 return (address) heap->collection_set()->biased_map_address();
2195 }
2196
2197 void ShenandoahHeap::reset_bytes_allocated_since_gc_start() {
2198 if (mode()->is_generational()) {
2199 young_generation()->reset_bytes_allocated_since_gc_start();
2200 old_generation()->reset_bytes_allocated_since_gc_start();
2201 }
2202
2203 global_generation()->reset_bytes_allocated_since_gc_start();
2204 }
2205
2206 void ShenandoahHeap::set_degenerated_gc_in_progress(bool in_progress) {
2207 _degenerated_gc_in_progress.set_cond(in_progress);
2208 }
2209
2210 void ShenandoahHeap::set_full_gc_in_progress(bool in_progress) {
2211 _full_gc_in_progress.set_cond(in_progress);
2212 }
2213
2214 void ShenandoahHeap::set_full_gc_move_in_progress(bool in_progress) {
2215 assert (is_full_gc_in_progress(), "should be");
2216 _full_gc_move_in_progress.set_cond(in_progress);
2217 }
2218
2219 void ShenandoahHeap::set_update_refs_in_progress(bool in_progress) {
2220 set_gc_state(UPDATEREFS, in_progress);
2221 }
2222
2223 void ShenandoahHeap::register_nmethod(nmethod* nm) {
2224 ShenandoahCodeRoots::register_nmethod(nm);
2225 }
2226
2227 void ShenandoahHeap::unregister_nmethod(nmethod* nm) {
2228 ShenandoahCodeRoots::unregister_nmethod(nm);
2229 }
2230
2231 void ShenandoahHeap::pin_object(JavaThread* thr, oop o) {
2232 heap_region_containing(o)->record_pin();
2233 }
2234
2235 void ShenandoahHeap::unpin_object(JavaThread* thr, oop o) {
2236 ShenandoahHeapRegion* r = heap_region_containing(o);
2237 assert(r != nullptr, "Sanity");
2238 assert(r->pin_count() > 0, "Region " SIZE_FORMAT " should have non-zero pins", r->index());
2239 r->record_unpin();
2240 }
2241
2242 void ShenandoahHeap::sync_pinned_region_status() {
2243 ShenandoahHeapLocker locker(lock());
2244
2245 for (size_t i = 0; i < num_regions(); i++) {
2246 ShenandoahHeapRegion *r = get_region(i);
2247 if (r->is_active()) {
2248 if (r->is_pinned()) {
2249 if (r->pin_count() == 0) {
2250 r->make_unpinned();
2251 }
2252 } else {
2253 if (r->pin_count() > 0) {
2254 r->make_pinned();
2255 }
2256 }
2257 }
2258 }
2259
2260 assert_pinned_region_status();
2261 }
2262
2263 #ifdef ASSERT
2264 void ShenandoahHeap::assert_pinned_region_status() {
2265 for (size_t i = 0; i < num_regions(); i++) {
2266 ShenandoahHeapRegion* r = get_region(i);
2267 shenandoah_assert_generations_reconciled();
2268 if (gc_generation()->contains(r)) {
2269 assert((r->is_pinned() && r->pin_count() > 0) || (!r->is_pinned() && r->pin_count() == 0),
2270 "Region " SIZE_FORMAT " pinning status is inconsistent", i);
2271 }
2272 }
2273 }
2274 #endif
2275
2276 ConcurrentGCTimer* ShenandoahHeap::gc_timer() const {
2277 return _gc_timer;
2278 }
2279
2280 void ShenandoahHeap::prepare_concurrent_roots() {
2281 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint");
2282 assert(!is_stw_gc_in_progress(), "Only concurrent GC");
2283 set_concurrent_strong_root_in_progress(!collection_set()->is_empty());
2284 set_concurrent_weak_root_in_progress(true);
2285 if (unload_classes()) {
2286 _unloader.prepare();
2287 }
2288 }
2289
2290 void ShenandoahHeap::finish_concurrent_roots() {
2291 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint");
2292 assert(!is_stw_gc_in_progress(), "Only concurrent GC");
2293 if (unload_classes()) {
2294 _unloader.finish();
2295 }
2296 }
2297
2298 #ifdef ASSERT
2299 void ShenandoahHeap::assert_gc_workers(uint nworkers) {
2300 assert(nworkers > 0 && nworkers <= max_workers(), "Sanity");
2301
2302 if (ShenandoahSafepoint::is_at_shenandoah_safepoint()) {
2303 // Use ParallelGCThreads inside safepoints
2304 assert(nworkers == ParallelGCThreads, "Use ParallelGCThreads (%u) within safepoint, not %u",
2305 ParallelGCThreads, nworkers);
2306 } else {
2307 // Use ConcGCThreads outside safepoints
2308 assert(nworkers == ConcGCThreads, "Use ConcGCThreads (%u) outside safepoints, %u",
2309 ConcGCThreads, nworkers);
2310 }
2311 }
2312 #endif
2313
2314 ShenandoahVerifier* ShenandoahHeap::verifier() {
2315 guarantee(ShenandoahVerify, "Should be enabled");
2316 assert (_verifier != nullptr, "sanity");
2317 return _verifier;
2318 }
2319
2320 template<bool CONCURRENT>
2321 class ShenandoahUpdateHeapRefsTask : public WorkerTask {
2322 private:
2323 ShenandoahHeap* _heap;
2324 ShenandoahRegionIterator* _regions;
2325 public:
2326 explicit ShenandoahUpdateHeapRefsTask(ShenandoahRegionIterator* regions) :
2327 WorkerTask("Shenandoah Update References"),
2328 _heap(ShenandoahHeap::heap()),
2329 _regions(regions) {
2330 }
2331
2332 void work(uint worker_id) {
2333 if (CONCURRENT) {
2334 ShenandoahConcurrentWorkerSession worker_session(worker_id);
2335 ShenandoahSuspendibleThreadSetJoiner stsj;
2336 do_work<ShenandoahConcUpdateRefsClosure>(worker_id);
2337 } else {
2338 ShenandoahParallelWorkerSession worker_session(worker_id);
2339 do_work<ShenandoahSTWUpdateRefsClosure>(worker_id);
2340 }
2341 }
2342
2343 private:
2344 template<class T>
2345 void do_work(uint worker_id) {
2346 if (CONCURRENT && (worker_id == 0)) {
2347 // We ask the first worker to replenish the Mutator free set by moving regions previously reserved to hold the
2348 // results of evacuation. These reserves are no longer necessary because evacuation has completed.
2349 size_t cset_regions = _heap->collection_set()->count();
2350
2351 // Now that evacuation is done, we can reassign any regions that had been reserved to hold the results of evacuation
2352 // to the mutator free set. At the end of GC, we will have cset_regions newly evacuated fully empty regions from
2353 // which we will be able to replenish the Collector free set and the OldCollector free set in preparation for the
2354 // next GC cycle.
2355 _heap->free_set()->move_regions_from_collector_to_mutator(cset_regions);
2356 }
2357 // If !CONCURRENT, there's no value in expanding Mutator free set
2358 T cl;
2359 ShenandoahHeapRegion* r = _regions->next();
2360 while (r != nullptr) {
2361 HeapWord* update_watermark = r->get_update_watermark();
2362 assert (update_watermark >= r->bottom(), "sanity");
2363 if (r->is_active() && !r->is_cset()) {
2364 _heap->marked_object_oop_iterate(r, &cl, update_watermark);
2365 if (ShenandoahPacing) {
2366 _heap->pacer()->report_updaterefs(pointer_delta(update_watermark, r->bottom()));
2367 }
2368 }
2369 if (_heap->check_cancelled_gc_and_yield(CONCURRENT)) {
2370 return;
2371 }
2372 r = _regions->next();
2373 }
2374 }
2375 };
2376
2377 void ShenandoahHeap::update_heap_references(bool concurrent) {
2378 assert(!is_full_gc_in_progress(), "Only for concurrent and degenerated GC");
2379
2380 if (concurrent) {
2381 ShenandoahUpdateHeapRefsTask<true> task(&_update_refs_iterator);
2382 workers()->run_task(&task);
2383 } else {
2384 ShenandoahUpdateHeapRefsTask<false> task(&_update_refs_iterator);
2385 workers()->run_task(&task);
2386 }
2387 }
2388
2389 void ShenandoahHeap::update_heap_region_states(bool concurrent) {
2390 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint");
2391 assert(!is_full_gc_in_progress(), "Only for concurrent and degenerated GC");
2392
2393 {
2394 ShenandoahGCPhase phase(concurrent ?
2395 ShenandoahPhaseTimings::final_update_refs_update_region_states :
2396 ShenandoahPhaseTimings::degen_gc_final_update_refs_update_region_states);
2397
2398 final_update_refs_update_region_states();
2399
2400 assert_pinned_region_status();
2401 }
2402
2403 {
2404 ShenandoahGCPhase phase(concurrent ?
2405 ShenandoahPhaseTimings::final_update_refs_trash_cset :
2406 ShenandoahPhaseTimings::degen_gc_final_update_refs_trash_cset);
2407 trash_cset_regions();
2408 }
2409 }
2410
2411 void ShenandoahHeap::final_update_refs_update_region_states() {
2412 ShenandoahSynchronizePinnedRegionStates cl;
2413 parallel_heap_region_iterate(&cl);
2414 }
2415
2416 void ShenandoahHeap::rebuild_free_set(bool concurrent) {
2417 ShenandoahGCPhase phase(concurrent ?
2418 ShenandoahPhaseTimings::final_update_refs_rebuild_freeset :
2419 ShenandoahPhaseTimings::degen_gc_final_update_refs_rebuild_freeset);
2420 ShenandoahHeapLocker locker(lock());
2421 size_t young_cset_regions, old_cset_regions;
2422 size_t first_old_region, last_old_region, old_region_count;
2423 _free_set->prepare_to_rebuild(young_cset_regions, old_cset_regions, first_old_region, last_old_region, old_region_count);
2424 // If there are no old regions, first_old_region will be greater than last_old_region
2425 assert((first_old_region > last_old_region) ||
2426 ((last_old_region + 1 - first_old_region >= old_region_count) &&
2427 get_region(first_old_region)->is_old() && get_region(last_old_region)->is_old()),
2428 "sanity: old_region_count: " SIZE_FORMAT ", first_old_region: " SIZE_FORMAT ", last_old_region: " SIZE_FORMAT,
2429 old_region_count, first_old_region, last_old_region);
2430
2431 if (mode()->is_generational()) {
2432 #ifdef ASSERT
2433 if (ShenandoahVerify) {
2434 verifier()->verify_before_rebuilding_free_set();
2435 }
2436 #endif
2437
2438 // The computation of bytes_of_allocation_runway_before_gc_trigger is quite conservative so consider all of this
2439 // available for transfer to old. Note that transfer of humongous regions does not impact available.
2440 ShenandoahGenerationalHeap* gen_heap = ShenandoahGenerationalHeap::heap();
2441 size_t allocation_runway = gen_heap->young_generation()->heuristics()->bytes_of_allocation_runway_before_gc_trigger(young_cset_regions);
2442 gen_heap->compute_old_generation_balance(allocation_runway, old_cset_regions);
2443
2444 // Total old_available may have been expanded to hold anticipated promotions. We trigger if the fragmented available
2445 // memory represents more than 16 regions worth of data. Note that fragmentation may increase when we promote regular
2446 // regions in place when many of these regular regions have an abundant amount of available memory within them. Fragmentation
2447 // will decrease as promote-by-copy consumes the available memory within these partially consumed regions.
2448 //
2449 // We consider old-gen to have excessive fragmentation if more than 12.5% of old-gen is free memory that resides
2450 // within partially consumed regions of memory.
2451 }
2452 // Rebuild free set based on adjusted generation sizes.
2453 _free_set->finish_rebuild(young_cset_regions, old_cset_regions, old_region_count);
2454
2455 if (mode()->is_generational()) {
2456 ShenandoahGenerationalHeap* gen_heap = ShenandoahGenerationalHeap::heap();
2457 ShenandoahOldGeneration* old_gen = gen_heap->old_generation();
2458 old_gen->heuristics()->evaluate_triggers(first_old_region, last_old_region, old_region_count, num_regions());
2459 }
2460 }
2461
2462 void ShenandoahHeap::print_extended_on(outputStream *st) const {
2463 print_on(st);
2464 st->cr();
2465 print_heap_regions_on(st);
2466 }
2467
2468 bool ShenandoahHeap::is_bitmap_slice_committed(ShenandoahHeapRegion* r, bool skip_self) {
2469 size_t slice = r->index() / _bitmap_regions_per_slice;
2470
2471 size_t regions_from = _bitmap_regions_per_slice * slice;
2472 size_t regions_to = MIN2(num_regions(), _bitmap_regions_per_slice * (slice + 1));
2473 for (size_t g = regions_from; g < regions_to; g++) {
2474 assert (g / _bitmap_regions_per_slice == slice, "same slice");
2475 if (skip_self && g == r->index()) continue;
2476 if (get_region(g)->is_committed()) {
2477 return true;
2478 }
2479 }
2480 return false;
2481 }
2482
2483 bool ShenandoahHeap::commit_bitmap_slice(ShenandoahHeapRegion* r) {
2484 shenandoah_assert_heaplocked();
2485
2486 // Bitmaps in special regions do not need commits
2487 if (_bitmap_region_special) {
2488 return true;
2489 }
2490
2491 if (is_bitmap_slice_committed(r, true)) {
2492 // Some other region from the group is already committed, meaning the bitmap
2493 // slice is already committed, we exit right away.
2494 return true;
2495 }
2496
2497 // Commit the bitmap slice:
2498 size_t slice = r->index() / _bitmap_regions_per_slice;
2499 size_t off = _bitmap_bytes_per_slice * slice;
2500 size_t len = _bitmap_bytes_per_slice;
2501 char* start = (char*) _bitmap_region.start() + off;
2502
2503 if (!os::commit_memory(start, len, false)) {
2504 return false;
2505 }
2506
2507 if (AlwaysPreTouch) {
2508 os::pretouch_memory(start, start + len, _pretouch_bitmap_page_size);
2509 }
2510
2511 return true;
2512 }
2513
2514 bool ShenandoahHeap::uncommit_bitmap_slice(ShenandoahHeapRegion *r) {
2515 shenandoah_assert_heaplocked();
2516
2517 // Bitmaps in special regions do not need uncommits
2518 if (_bitmap_region_special) {
2519 return true;
2520 }
2521
2522 if (is_bitmap_slice_committed(r, true)) {
2523 // Some other region from the group is still committed, meaning the bitmap
2524 // slice is should stay committed, exit right away.
2525 return true;
2526 }
2527
2528 // Uncommit the bitmap slice:
2529 size_t slice = r->index() / _bitmap_regions_per_slice;
2530 size_t off = _bitmap_bytes_per_slice * slice;
2531 size_t len = _bitmap_bytes_per_slice;
2532 if (!os::uncommit_memory((char*)_bitmap_region.start() + off, len)) {
2533 return false;
2534 }
2535 return true;
2536 }
2537
2538 void ShenandoahHeap::safepoint_synchronize_begin() {
2539 SuspendibleThreadSet::synchronize();
2540 }
2541
2542 void ShenandoahHeap::safepoint_synchronize_end() {
2543 SuspendibleThreadSet::desynchronize();
2544 }
2545
2546 void ShenandoahHeap::try_inject_alloc_failure() {
2547 if (ShenandoahAllocFailureALot && !cancelled_gc() && ((os::random() % 1000) > 950)) {
2548 _inject_alloc_failure.set();
2549 os::naked_short_sleep(1);
2550 if (cancelled_gc()) {
2551 log_info(gc)("Allocation failure was successfully injected");
2552 }
2553 }
2554 }
2555
2556 bool ShenandoahHeap::should_inject_alloc_failure() {
2557 return _inject_alloc_failure.is_set() && _inject_alloc_failure.try_unset();
2558 }
2559
2560 void ShenandoahHeap::initialize_serviceability() {
2561 _memory_pool = new ShenandoahMemoryPool(this);
2562 _cycle_memory_manager.add_pool(_memory_pool);
2563 _stw_memory_manager.add_pool(_memory_pool);
2564 }
2565
2566 GrowableArray<GCMemoryManager*> ShenandoahHeap::memory_managers() {
2567 GrowableArray<GCMemoryManager*> memory_managers(2);
2568 memory_managers.append(&_cycle_memory_manager);
2569 memory_managers.append(&_stw_memory_manager);
2570 return memory_managers;
2571 }
2572
2573 GrowableArray<MemoryPool*> ShenandoahHeap::memory_pools() {
2574 GrowableArray<MemoryPool*> memory_pools(1);
2575 memory_pools.append(_memory_pool);
2576 return memory_pools;
2577 }
2578
2579 MemoryUsage ShenandoahHeap::memory_usage() {
2580 return MemoryUsage(_initial_size, used(), committed(), max_capacity());
2581 }
2582
2583 ShenandoahRegionIterator::ShenandoahRegionIterator() :
2584 _heap(ShenandoahHeap::heap()),
2585 _index(0) {}
2586
2587 ShenandoahRegionIterator::ShenandoahRegionIterator(ShenandoahHeap* heap) :
2588 _heap(heap),
2589 _index(0) {}
2590
2591 void ShenandoahRegionIterator::reset() {
2592 _index = 0;
2593 }
2594
2595 bool ShenandoahRegionIterator::has_next() const {
2596 return _index < _heap->num_regions();
2597 }
2598
2599 char ShenandoahHeap::gc_state() const {
2600 return _gc_state.raw_value();
2601 }
2602
2603 ShenandoahLiveData* ShenandoahHeap::get_liveness_cache(uint worker_id) {
2604 #ifdef ASSERT
2605 assert(_liveness_cache != nullptr, "sanity");
2606 assert(worker_id < _max_workers, "sanity");
2607 for (uint i = 0; i < num_regions(); i++) {
2608 assert(_liveness_cache[worker_id][i] == 0, "liveness cache should be empty");
2609 }
2610 #endif
2611 return _liveness_cache[worker_id];
2612 }
2613
2614 void ShenandoahHeap::flush_liveness_cache(uint worker_id) {
2615 assert(worker_id < _max_workers, "sanity");
2616 assert(_liveness_cache != nullptr, "sanity");
2617 ShenandoahLiveData* ld = _liveness_cache[worker_id];
2618 for (uint i = 0; i < num_regions(); i++) {
2619 ShenandoahLiveData live = ld[i];
2620 if (live > 0) {
2621 ShenandoahHeapRegion* r = get_region(i);
2622 r->increase_live_data_gc_words(live);
2623 ld[i] = 0;
2624 }
2625 }
2626 }
2627
2628 bool ShenandoahHeap::requires_barriers(stackChunkOop obj) const {
2629 if (is_idle()) return false;
2630
2631 // Objects allocated after marking start are implicitly alive, don't need any barriers during
2632 // marking phase.
2633 if (is_concurrent_mark_in_progress() &&
2634 !marking_context()->allocated_after_mark_start(obj)) {
2635 return true;
2636 }
2637
2638 // Can not guarantee obj is deeply good.
2639 if (has_forwarded_objects()) {
2640 return true;
2641 }
2642
2643 return false;
2644 }
2645
2646 ShenandoahGeneration* ShenandoahHeap::generation_for(ShenandoahAffiliation affiliation) const {
2647 if (!mode()->is_generational()) {
2648 return global_generation();
2649 } else if (affiliation == YOUNG_GENERATION) {
2650 return young_generation();
2651 } else if (affiliation == OLD_GENERATION) {
2652 return old_generation();
2653 }
2654
2655 ShouldNotReachHere();
2656 return nullptr;
2657 }
2658
2659 void ShenandoahHeap::log_heap_status(const char* msg) const {
2660 if (mode()->is_generational()) {
2661 young_generation()->log_status(msg);
2662 old_generation()->log_status(msg);
2663 } else {
2664 global_generation()->log_status(msg);
2665 }
2666 }