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
29 #include "cds/archiveHeapWriter.hpp"
30 #include "classfile/systemDictionary.hpp"
31 #include "code/codeCache.hpp"
32
33 #include "gc/shared/classUnloadingContext.hpp"
34 #include "gc/shared/fullGCForwarding.hpp"
35 #include "gc/shared/gcArguments.hpp"
36 #include "gc/shared/gcTimer.hpp"
37 #include "gc/shared/gcTraceTime.inline.hpp"
38 #include "gc/shared/locationPrinter.inline.hpp"
39 #include "gc/shared/memAllocator.hpp"
40 #include "gc/shared/plab.hpp"
41 #include "gc/shared/tlab_globals.hpp"
42
43 #include "gc/shenandoah/heuristics/shenandoahOldHeuristics.hpp"
44 #include "gc/shenandoah/heuristics/shenandoahYoungHeuristics.hpp"
45 #include "gc/shenandoah/shenandoahAllocRequest.hpp"
46 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
47 #include "gc/shenandoah/shenandoahCodeRoots.hpp"
48 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
49 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
50 #include "gc/shenandoah/shenandoahConcurrentMark.hpp"
51 #include "gc/shenandoah/shenandoahControlThread.hpp"
52 #include "gc/shenandoah/shenandoahClosures.inline.hpp"
53 #include "gc/shenandoah/shenandoahFreeSet.hpp"
54 #include "gc/shenandoah/shenandoahGenerationalEvacuationTask.hpp"
55 #include "gc/shenandoah/shenandoahGenerationalHeap.hpp"
56 #include "gc/shenandoah/shenandoahGlobalGeneration.hpp"
57 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
58 #include "gc/shenandoah/shenandoahHeapRegionClosures.hpp"
59 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
60 #include "gc/shenandoah/shenandoahHeapRegionSet.hpp"
61 #include "gc/shenandoah/shenandoahInitLogger.hpp"
62 #include "gc/shenandoah/shenandoahMarkingContext.inline.hpp"
63 #include "gc/shenandoah/shenandoahMemoryPool.hpp"
64 #include "gc/shenandoah/shenandoahMonitoringSupport.hpp"
65 #include "gc/shenandoah/shenandoahOldGeneration.hpp"
66 #include "gc/shenandoah/shenandoahPacer.inline.hpp"
67 #include "gc/shenandoah/shenandoahPadding.hpp"
68 #include "gc/shenandoah/shenandoahParallelCleaning.inline.hpp"
69 #include "gc/shenandoah/shenandoahPhaseTimings.hpp"
70 #include "gc/shenandoah/shenandoahReferenceProcessor.hpp"
71 #include "gc/shenandoah/shenandoahRootProcessor.inline.hpp"
72 #include "gc/shenandoah/shenandoahScanRemembered.inline.hpp"
73 #include "gc/shenandoah/shenandoahSTWMark.hpp"
74 #include "gc/shenandoah/shenandoahUncommitThread.hpp"
75 #include "gc/shenandoah/shenandoahUtils.hpp"
76 #include "gc/shenandoah/shenandoahVerifier.hpp"
77 #include "gc/shenandoah/shenandoahVMOperations.hpp"
78 #include "gc/shenandoah/shenandoahWorkGroup.hpp"
79 #include "gc/shenandoah/shenandoahWorkerPolicy.hpp"
80 #include "gc/shenandoah/shenandoahYoungGeneration.hpp"
81 #include "gc/shenandoah/mode/shenandoahGenerationalMode.hpp"
82 #include "gc/shenandoah/mode/shenandoahPassiveMode.hpp"
83 #include "gc/shenandoah/mode/shenandoahSATBMode.hpp"
84
85 #if INCLUDE_JFR
86 #include "gc/shenandoah/shenandoahJfrSupport.hpp"
87 #endif
88
89
90 #include "memory/allocation.hpp"
91 #include "memory/classLoaderMetaspace.hpp"
92 #include "memory/metaspaceUtils.hpp"
93 #include "memory/universe.hpp"
94 #include "nmt/mallocTracker.hpp"
95 #include "nmt/memTracker.hpp"
96 #include "oops/compressedOops.inline.hpp"
97 #include "prims/jvmtiTagMap.hpp"
98 #include "runtime/atomic.hpp"
99 #include "runtime/globals.hpp"
100 #include "runtime/interfaceSupport.inline.hpp"
101 #include "runtime/java.hpp"
102 #include "runtime/orderAccess.hpp"
103 #include "runtime/safepointMechanism.hpp"
104 #include "runtime/stackWatermarkSet.hpp"
105 #include "runtime/vmThread.hpp"
106 #include "utilities/globalDefinitions.hpp"
107 #include "utilities/events.hpp"
108 #include "utilities/powerOfTwo.hpp"
109
110 class ShenandoahPretouchHeapTask : public WorkerTask {
111 private:
112 ShenandoahRegionIterator _regions;
113 const size_t _page_size;
114 public:
115 ShenandoahPretouchHeapTask(size_t page_size) :
116 WorkerTask("Shenandoah Pretouch Heap"),
117 _page_size(page_size) {}
118
119 virtual void work(uint worker_id) {
120 ShenandoahHeapRegion* r = _regions.next();
121 while (r != nullptr) {
122 if (r->is_committed()) {
123 os::pretouch_memory(r->bottom(), r->end(), _page_size);
124 }
125 r = _regions.next();
126 }
127 }
128 };
129
130 class ShenandoahPretouchBitmapTask : public WorkerTask {
131 private:
132 ShenandoahRegionIterator _regions;
133 char* _bitmap_base;
134 const size_t _bitmap_size;
135 const size_t _page_size;
136 public:
137 ShenandoahPretouchBitmapTask(char* bitmap_base, size_t bitmap_size, size_t page_size) :
138 WorkerTask("Shenandoah Pretouch Bitmap"),
139 _bitmap_base(bitmap_base),
140 _bitmap_size(bitmap_size),
141 _page_size(page_size) {}
142
143 virtual void work(uint worker_id) {
144 ShenandoahHeapRegion* r = _regions.next();
145 while (r != nullptr) {
146 size_t start = r->index() * ShenandoahHeapRegion::region_size_bytes() / MarkBitMap::heap_map_factor();
147 size_t end = (r->index() + 1) * ShenandoahHeapRegion::region_size_bytes() / MarkBitMap::heap_map_factor();
148 assert (end <= _bitmap_size, "end is sane: " SIZE_FORMAT " < " SIZE_FORMAT, end, _bitmap_size);
149
150 if (r->is_committed()) {
151 os::pretouch_memory(_bitmap_base + start, _bitmap_base + end, _page_size);
152 }
153
154 r = _regions.next();
155 }
156 }
157 };
158
159 jint ShenandoahHeap::initialize() {
160 //
161 // Figure out heap sizing
162 //
163
164 size_t init_byte_size = InitialHeapSize;
165 size_t min_byte_size = MinHeapSize;
166 size_t max_byte_size = MaxHeapSize;
167 size_t heap_alignment = HeapAlignment;
168
169 size_t reg_size_bytes = ShenandoahHeapRegion::region_size_bytes();
170
171 Universe::check_alignment(max_byte_size, reg_size_bytes, "Shenandoah heap");
172 Universe::check_alignment(init_byte_size, reg_size_bytes, "Shenandoah heap");
173
174 _num_regions = ShenandoahHeapRegion::region_count();
175 assert(_num_regions == (max_byte_size / reg_size_bytes),
176 "Regions should cover entire heap exactly: " SIZE_FORMAT " != " SIZE_FORMAT "/" SIZE_FORMAT,
177 _num_regions, max_byte_size, reg_size_bytes);
178
179 size_t num_committed_regions = init_byte_size / reg_size_bytes;
180 num_committed_regions = MIN2(num_committed_regions, _num_regions);
181 assert(num_committed_regions <= _num_regions, "sanity");
182 _initial_size = num_committed_regions * reg_size_bytes;
183
184 size_t num_min_regions = min_byte_size / reg_size_bytes;
185 num_min_regions = MIN2(num_min_regions, _num_regions);
186 assert(num_min_regions <= _num_regions, "sanity");
187 _minimum_size = num_min_regions * reg_size_bytes;
188
189 // Default to max heap size.
190 _soft_max_size = _num_regions * reg_size_bytes;
191
192 _committed = _initial_size;
193
194 size_t heap_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
195 size_t bitmap_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
196 size_t region_page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
197
198 //
199 // Reserve and commit memory for heap
200 //
201
202 ReservedHeapSpace heap_rs = Universe::reserve_heap(max_byte_size, heap_alignment);
203 initialize_reserved_region(heap_rs);
204 _heap_region = MemRegion((HeapWord*)heap_rs.base(), heap_rs.size() / HeapWordSize);
205 _heap_region_special = heap_rs.special();
206
207 assert((((size_t) base()) & ShenandoahHeapRegion::region_size_bytes_mask()) == 0,
208 "Misaligned heap: " PTR_FORMAT, p2i(base()));
209 os::trace_page_sizes_for_requested_size("Heap",
210 max_byte_size, heap_alignment,
211 heap_rs.base(),
212 heap_rs.size(), heap_rs.page_size());
213
214 #if SHENANDOAH_OPTIMIZED_MARKTASK
215 // The optimized ShenandoahMarkTask takes some bits away from the full object bits.
216 // Fail if we ever attempt to address more than we can.
217 if ((uintptr_t)heap_rs.end() >= ShenandoahMarkTask::max_addressable()) {
218 FormatBuffer<512> buf("Shenandoah reserved [" PTR_FORMAT ", " PTR_FORMAT") for the heap, \n"
219 "but max object address is " PTR_FORMAT ". Try to reduce heap size, or try other \n"
220 "VM options that allocate heap at lower addresses (HeapBaseMinAddress, AllocateHeapAt, etc).",
221 p2i(heap_rs.base()), p2i(heap_rs.end()), ShenandoahMarkTask::max_addressable());
222 vm_exit_during_initialization("Fatal Error", buf);
223 }
224 #endif
225
226 ReservedSpace sh_rs = heap_rs.first_part(max_byte_size);
227 if (!_heap_region_special) {
228 os::commit_memory_or_exit(sh_rs.base(), _initial_size, heap_alignment, false,
229 "Cannot commit heap memory");
230 }
231
232 BarrierSet::set_barrier_set(new ShenandoahBarrierSet(this, _heap_region));
233
234 // Now we know the number of regions and heap sizes, initialize the heuristics.
235 initialize_heuristics();
236
237 assert(_heap_region.byte_size() == heap_rs.size(), "Need to know reserved size for card table");
238
239 //
240 // Worker threads must be initialized after the barrier is configured
241 //
242 _workers = new ShenandoahWorkerThreads("Shenandoah GC Threads", _max_workers);
243 if (_workers == nullptr) {
244 vm_exit_during_initialization("Failed necessary allocation.");
245 } else {
246 _workers->initialize_workers();
247 }
248
249 if (ParallelGCThreads > 1) {
250 _safepoint_workers = new ShenandoahWorkerThreads("Safepoint Cleanup Thread", ParallelGCThreads);
251 _safepoint_workers->initialize_workers();
252 }
253
254 //
255 // Reserve and commit memory for bitmap(s)
256 //
257
258 size_t bitmap_size_orig = ShenandoahMarkBitMap::compute_size(heap_rs.size());
259 _bitmap_size = align_up(bitmap_size_orig, bitmap_page_size);
260
261 size_t bitmap_bytes_per_region = reg_size_bytes / ShenandoahMarkBitMap::heap_map_factor();
262
263 guarantee(bitmap_bytes_per_region != 0,
264 "Bitmap bytes per region should not be zero");
265 guarantee(is_power_of_2(bitmap_bytes_per_region),
266 "Bitmap bytes per region should be power of two: " SIZE_FORMAT, bitmap_bytes_per_region);
267
268 if (bitmap_page_size > bitmap_bytes_per_region) {
269 _bitmap_regions_per_slice = bitmap_page_size / bitmap_bytes_per_region;
270 _bitmap_bytes_per_slice = bitmap_page_size;
271 } else {
272 _bitmap_regions_per_slice = 1;
273 _bitmap_bytes_per_slice = bitmap_bytes_per_region;
274 }
275
276 guarantee(_bitmap_regions_per_slice >= 1,
277 "Should have at least one region per slice: " SIZE_FORMAT,
278 _bitmap_regions_per_slice);
279
280 guarantee(((_bitmap_bytes_per_slice) % bitmap_page_size) == 0,
281 "Bitmap slices should be page-granular: bps = " SIZE_FORMAT ", page size = " SIZE_FORMAT,
282 _bitmap_bytes_per_slice, bitmap_page_size);
283
284 ReservedSpace bitmap(_bitmap_size, bitmap_page_size);
285 os::trace_page_sizes_for_requested_size("Mark Bitmap",
286 bitmap_size_orig, bitmap_page_size,
287 bitmap.base(),
288 bitmap.size(), bitmap.page_size());
289 MemTracker::record_virtual_memory_tag(bitmap.base(), mtGC);
290 _bitmap_region = MemRegion((HeapWord*) bitmap.base(), bitmap.size() / HeapWordSize);
291 _bitmap_region_special = bitmap.special();
292
293 size_t bitmap_init_commit = _bitmap_bytes_per_slice *
294 align_up(num_committed_regions, _bitmap_regions_per_slice) / _bitmap_regions_per_slice;
295 bitmap_init_commit = MIN2(_bitmap_size, bitmap_init_commit);
296 if (!_bitmap_region_special) {
297 os::commit_memory_or_exit((char *) _bitmap_region.start(), bitmap_init_commit, bitmap_page_size, false,
298 "Cannot commit bitmap memory");
299 }
300
301 _marking_context = new ShenandoahMarkingContext(_heap_region, _bitmap_region, _num_regions);
302
303 if (ShenandoahVerify) {
304 ReservedSpace verify_bitmap(_bitmap_size, bitmap_page_size);
305 os::trace_page_sizes_for_requested_size("Verify Bitmap",
306 bitmap_size_orig, bitmap_page_size,
307 verify_bitmap.base(),
308 verify_bitmap.size(), verify_bitmap.page_size());
309 if (!verify_bitmap.special()) {
310 os::commit_memory_or_exit(verify_bitmap.base(), verify_bitmap.size(), bitmap_page_size, false,
311 "Cannot commit verification bitmap memory");
312 }
313 MemTracker::record_virtual_memory_tag(verify_bitmap.base(), mtGC);
314 MemRegion verify_bitmap_region = MemRegion((HeapWord *) verify_bitmap.base(), verify_bitmap.size() / HeapWordSize);
315 _verification_bit_map.initialize(_heap_region, verify_bitmap_region);
316 _verifier = new ShenandoahVerifier(this, &_verification_bit_map);
317 }
318
319 // Reserve aux bitmap for use in object_iterate(). We don't commit it here.
320 size_t aux_bitmap_page_size = bitmap_page_size;
321
322 ReservedSpace aux_bitmap(_bitmap_size, aux_bitmap_page_size);
323 os::trace_page_sizes_for_requested_size("Aux Bitmap",
324 bitmap_size_orig, aux_bitmap_page_size,
325 aux_bitmap.base(),
326 aux_bitmap.size(), aux_bitmap.page_size());
327 MemTracker::record_virtual_memory_tag(aux_bitmap.base(), mtGC);
328 _aux_bitmap_region = MemRegion((HeapWord*) aux_bitmap.base(), aux_bitmap.size() / HeapWordSize);
329 _aux_bitmap_region_special = aux_bitmap.special();
330 _aux_bit_map.initialize(_heap_region, _aux_bitmap_region);
331
332 //
333 // Create regions and region sets
334 //
335 size_t region_align = align_up(sizeof(ShenandoahHeapRegion), SHENANDOAH_CACHE_LINE_SIZE);
336 size_t region_storage_size_orig = region_align * _num_regions;
337 size_t region_storage_size = align_up(region_storage_size_orig,
338 MAX2(region_page_size, os::vm_allocation_granularity()));
339
340 ReservedSpace region_storage(region_storage_size, region_page_size);
341 os::trace_page_sizes_for_requested_size("Region Storage",
342 region_storage_size_orig, region_page_size,
343 region_storage.base(),
344 region_storage.size(), region_storage.page_size());
345 MemTracker::record_virtual_memory_tag(region_storage.base(), mtGC);
346 if (!region_storage.special()) {
347 os::commit_memory_or_exit(region_storage.base(), region_storage_size, region_page_size, false,
348 "Cannot commit region memory");
349 }
350
351 // Try to fit the collection set bitmap at lower addresses. This optimizes code generation for cset checks.
352 // Go up until a sensible limit (subject to encoding constraints) and try to reserve the space there.
353 // If not successful, bite a bullet and allocate at whatever address.
354 {
355 const size_t cset_align = MAX2<size_t>(os::vm_page_size(), os::vm_allocation_granularity());
356 const size_t cset_size = align_up(((size_t) sh_rs.base() + sh_rs.size()) >> ShenandoahHeapRegion::region_size_bytes_shift(), cset_align);
357 const size_t cset_page_size = os::vm_page_size();
358
359 uintptr_t min = round_up_power_of_2(cset_align);
360 uintptr_t max = (1u << 30u);
361 ReservedSpace cset_rs;
362
363 for (uintptr_t addr = min; addr <= max; addr <<= 1u) {
364 char* req_addr = (char*)addr;
365 assert(is_aligned(req_addr, cset_align), "Should be aligned");
366 cset_rs = ReservedSpace(cset_size, cset_align, cset_page_size, req_addr);
367 if (cset_rs.is_reserved()) {
368 assert(cset_rs.base() == req_addr, "Allocated where requested: " PTR_FORMAT ", " PTR_FORMAT, p2i(cset_rs.base()), addr);
369 _collection_set = new ShenandoahCollectionSet(this, cset_rs, sh_rs.base());
370 break;
371 }
372 }
373
374 if (_collection_set == nullptr) {
375 cset_rs = ReservedSpace(cset_size, cset_align, os::vm_page_size());
376 _collection_set = new ShenandoahCollectionSet(this, cset_rs, sh_rs.base());
377 }
378 os::trace_page_sizes_for_requested_size("Collection Set",
379 cset_size, cset_page_size,
380 cset_rs.base(),
381 cset_rs.size(), cset_rs.page_size());
382 }
383
384 _regions = NEW_C_HEAP_ARRAY(ShenandoahHeapRegion*, _num_regions, mtGC);
385 _affiliations = NEW_C_HEAP_ARRAY(uint8_t, _num_regions, mtGC);
386 _free_set = new ShenandoahFreeSet(this, _num_regions);
387
388 {
389 ShenandoahHeapLocker locker(lock());
390
391 for (size_t i = 0; i < _num_regions; i++) {
392 HeapWord* start = (HeapWord*)sh_rs.base() + ShenandoahHeapRegion::region_size_words() * i;
393 bool is_committed = i < num_committed_regions;
394 void* loc = region_storage.base() + i * region_align;
395
396 ShenandoahHeapRegion* r = new (loc) ShenandoahHeapRegion(start, i, is_committed);
397 assert(is_aligned(r, SHENANDOAH_CACHE_LINE_SIZE), "Sanity");
398
399 _marking_context->initialize_top_at_mark_start(r);
400 _regions[i] = r;
401 assert(!collection_set()->is_in(i), "New region should not be in collection set");
402
403 _affiliations[i] = ShenandoahAffiliation::FREE;
404 }
405
406 // Initialize to complete
407 _marking_context->mark_complete();
408 size_t young_cset_regions, old_cset_regions;
409
410 // We are initializing free set. We ignore cset region tallies.
411 size_t first_old, last_old, num_old;
412 _free_set->prepare_to_rebuild(young_cset_regions, old_cset_regions, first_old, last_old, num_old);
413 _free_set->finish_rebuild(young_cset_regions, old_cset_regions, num_old);
414 }
415
416 if (AlwaysPreTouch) {
417 // For NUMA, it is important to pre-touch the storage under bitmaps with worker threads,
418 // before initialize() below zeroes it with initializing thread. For any given region,
419 // we touch the region and the corresponding bitmaps from the same thread.
420 ShenandoahPushWorkerScope scope(workers(), _max_workers, false);
421
422 _pretouch_heap_page_size = heap_page_size;
423 _pretouch_bitmap_page_size = bitmap_page_size;
424
425 // OS memory managers may want to coalesce back-to-back pages. Make their jobs
426 // simpler by pre-touching continuous spaces (heap and bitmap) separately.
427
428 ShenandoahPretouchBitmapTask bcl(bitmap.base(), _bitmap_size, _pretouch_bitmap_page_size);
429 _workers->run_task(&bcl);
430
431 ShenandoahPretouchHeapTask hcl(_pretouch_heap_page_size);
432 _workers->run_task(&hcl);
433 }
434
435 //
436 // Initialize the rest of GC subsystems
437 //
438
439 _liveness_cache = NEW_C_HEAP_ARRAY(ShenandoahLiveData*, _max_workers, mtGC);
440 for (uint worker = 0; worker < _max_workers; worker++) {
441 _liveness_cache[worker] = NEW_C_HEAP_ARRAY(ShenandoahLiveData, _num_regions, mtGC);
442 Copy::fill_to_bytes(_liveness_cache[worker], _num_regions * sizeof(ShenandoahLiveData));
443 }
444
445 // There should probably be Shenandoah-specific options for these,
446 // just as there are G1-specific options.
447 {
448 ShenandoahSATBMarkQueueSet& satbqs = ShenandoahBarrierSet::satb_mark_queue_set();
449 satbqs.set_process_completed_buffers_threshold(20); // G1SATBProcessCompletedThreshold
450 satbqs.set_buffer_enqueue_threshold_percentage(60); // G1SATBBufferEnqueueingThresholdPercent
451 }
452
453 _monitoring_support = new ShenandoahMonitoringSupport(this);
454 _phase_timings = new ShenandoahPhaseTimings(max_workers());
455 ShenandoahCodeRoots::initialize();
456
457 if (ShenandoahPacing) {
458 _pacer = new ShenandoahPacer(this);
459 _pacer->setup_for_idle();
460 }
461
462 initialize_controller();
463
464 if (ShenandoahUncommit) {
465 _uncommit_thread = new ShenandoahUncommitThread(this);
466 }
467
468 print_init_logger();
469
470 FullGCForwarding::initialize(_heap_region);
471
472 return JNI_OK;
473 }
474
475 void ShenandoahHeap::initialize_controller() {
476 _control_thread = new ShenandoahControlThread();
477 }
478
479 void ShenandoahHeap::print_init_logger() const {
480 ShenandoahInitLogger::print();
481 }
482
483 void ShenandoahHeap::initialize_mode() {
484 if (ShenandoahGCMode != nullptr) {
485 if (strcmp(ShenandoahGCMode, "satb") == 0) {
486 _gc_mode = new ShenandoahSATBMode();
487 } else if (strcmp(ShenandoahGCMode, "passive") == 0) {
488 _gc_mode = new ShenandoahPassiveMode();
489 } else if (strcmp(ShenandoahGCMode, "generational") == 0) {
490 _gc_mode = new ShenandoahGenerationalMode();
491 } else {
492 vm_exit_during_initialization("Unknown -XX:ShenandoahGCMode option");
493 }
494 } else {
495 vm_exit_during_initialization("Unknown -XX:ShenandoahGCMode option (null)");
496 }
497 _gc_mode->initialize_flags();
498 if (_gc_mode->is_diagnostic() && !UnlockDiagnosticVMOptions) {
499 vm_exit_during_initialization(
500 err_msg("GC mode \"%s\" is diagnostic, and must be enabled via -XX:+UnlockDiagnosticVMOptions.",
501 _gc_mode->name()));
502 }
503 if (_gc_mode->is_experimental() && !UnlockExperimentalVMOptions) {
504 vm_exit_during_initialization(
505 err_msg("GC mode \"%s\" is experimental, and must be enabled via -XX:+UnlockExperimentalVMOptions.",
506 _gc_mode->name()));
507 }
508 }
509
510 void ShenandoahHeap::initialize_heuristics() {
511 _global_generation = new ShenandoahGlobalGeneration(mode()->is_generational(), max_workers(), max_capacity(), max_capacity());
512 _global_generation->initialize_heuristics(mode());
513 }
514
515 #ifdef _MSC_VER
516 #pragma warning( push )
517 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
518 #endif
519
520 ShenandoahHeap::ShenandoahHeap(ShenandoahCollectorPolicy* policy) :
521 CollectedHeap(),
522 _gc_generation(nullptr),
523 _active_generation(nullptr),
524 _initial_size(0),
525 _committed(0),
526 _max_workers(MAX3(ConcGCThreads, ParallelGCThreads, 1U)),
527 _workers(nullptr),
528 _safepoint_workers(nullptr),
529 _heap_region_special(false),
530 _num_regions(0),
531 _regions(nullptr),
532 _affiliations(nullptr),
533 _gc_state_changed(false),
534 _gc_no_progress_count(0),
535 _cancel_requested_time(0),
536 _update_refs_iterator(this),
537 _global_generation(nullptr),
538 _control_thread(nullptr),
539 _uncommit_thread(nullptr),
540 _young_generation(nullptr),
541 _old_generation(nullptr),
542 _shenandoah_policy(policy),
543 _gc_mode(nullptr),
544 _free_set(nullptr),
545 _pacer(nullptr),
546 _verifier(nullptr),
547 _phase_timings(nullptr),
548 _mmu_tracker(),
549 _monitoring_support(nullptr),
550 _memory_pool(nullptr),
551 _stw_memory_manager("Shenandoah Pauses"),
552 _cycle_memory_manager("Shenandoah Cycles"),
553 _gc_timer(new ConcurrentGCTimer()),
554 _log_min_obj_alignment_in_bytes(LogMinObjAlignmentInBytes),
555 _marking_context(nullptr),
556 _bitmap_size(0),
557 _bitmap_regions_per_slice(0),
558 _bitmap_bytes_per_slice(0),
559 _bitmap_region_special(false),
560 _aux_bitmap_region_special(false),
561 _liveness_cache(nullptr),
562 _collection_set(nullptr)
563 {
564 // Initialize GC mode early, many subsequent initialization procedures depend on it
565 initialize_mode();
566 }
567
568 #ifdef _MSC_VER
569 #pragma warning( pop )
570 #endif
571
572 void ShenandoahHeap::print_on(outputStream* st) const {
573 st->print_cr("Shenandoah Heap");
574 st->print_cr(" " SIZE_FORMAT "%s max, " SIZE_FORMAT "%s soft max, " SIZE_FORMAT "%s committed, " SIZE_FORMAT "%s used",
575 byte_size_in_proper_unit(max_capacity()), proper_unit_for_byte_size(max_capacity()),
576 byte_size_in_proper_unit(soft_max_capacity()), proper_unit_for_byte_size(soft_max_capacity()),
577 byte_size_in_proper_unit(committed()), proper_unit_for_byte_size(committed()),
578 byte_size_in_proper_unit(used()), proper_unit_for_byte_size(used()));
579 st->print_cr(" " SIZE_FORMAT " x " SIZE_FORMAT"%s regions",
580 num_regions(),
581 byte_size_in_proper_unit(ShenandoahHeapRegion::region_size_bytes()),
582 proper_unit_for_byte_size(ShenandoahHeapRegion::region_size_bytes()));
583
584 st->print("Status: ");
585 if (has_forwarded_objects()) st->print("has forwarded objects, ");
586 if (!mode()->is_generational()) {
587 if (is_concurrent_mark_in_progress()) st->print("marking,");
588 } else {
589 if (is_concurrent_old_mark_in_progress()) st->print("old marking, ");
590 if (is_concurrent_young_mark_in_progress()) st->print("young marking, ");
591 }
592 if (is_evacuation_in_progress()) st->print("evacuating, ");
593 if (is_update_refs_in_progress()) st->print("updating refs, ");
594 if (is_degenerated_gc_in_progress()) st->print("degenerated gc, ");
595 if (is_full_gc_in_progress()) st->print("full gc, ");
596 if (is_full_gc_move_in_progress()) st->print("full gc move, ");
597 if (is_concurrent_weak_root_in_progress()) st->print("concurrent weak roots, ");
598 if (is_concurrent_strong_root_in_progress() &&
599 !is_concurrent_weak_root_in_progress()) st->print("concurrent strong roots, ");
600
601 if (cancelled_gc()) {
602 st->print("cancelled");
603 } else {
604 st->print("not cancelled");
605 }
606 st->cr();
607
608 st->print_cr("Reserved region:");
609 st->print_cr(" - [" PTR_FORMAT ", " PTR_FORMAT ") ",
610 p2i(reserved_region().start()),
611 p2i(reserved_region().end()));
612
613 ShenandoahCollectionSet* cset = collection_set();
614 st->print_cr("Collection set:");
615 if (cset != nullptr) {
616 st->print_cr(" - map (vanilla): " PTR_FORMAT, p2i(cset->map_address()));
617 st->print_cr(" - map (biased): " PTR_FORMAT, p2i(cset->biased_map_address()));
618 } else {
619 st->print_cr(" (null)");
620 }
621
622 st->cr();
623 MetaspaceUtils::print_on(st);
624
625 if (Verbose) {
626 st->cr();
627 print_heap_regions_on(st);
628 }
629 }
630
631 class ShenandoahInitWorkerGCLABClosure : public ThreadClosure {
632 public:
633 void do_thread(Thread* thread) {
634 assert(thread != nullptr, "Sanity");
635 assert(thread->is_Worker_thread(), "Only worker thread expected");
636 ShenandoahThreadLocalData::initialize_gclab(thread);
637 }
638 };
639
640 void ShenandoahHeap::post_initialize() {
641 CollectedHeap::post_initialize();
642 _mmu_tracker.initialize();
643
644 MutexLocker ml(Threads_lock);
645
646 ShenandoahInitWorkerGCLABClosure init_gclabs;
647 _workers->threads_do(&init_gclabs);
648
649 // gclab can not be initialized early during VM startup, as it can not determinate its max_size.
650 // Now, we will let WorkerThreads to initialize gclab when new worker is created.
651 _workers->set_initialize_gclab();
652 if (_safepoint_workers != nullptr) {
653 _safepoint_workers->threads_do(&init_gclabs);
654 _safepoint_workers->set_initialize_gclab();
655 }
656
657 JFR_ONLY(ShenandoahJFRSupport::register_jfr_type_serializers();)
658 }
659
660 ShenandoahHeuristics* ShenandoahHeap::heuristics() {
661 return _global_generation->heuristics();
662 }
663
664 size_t ShenandoahHeap::used() const {
665 return global_generation()->used();
666 }
667
668 size_t ShenandoahHeap::committed() const {
669 return Atomic::load(&_committed);
670 }
671
672 void ShenandoahHeap::increase_committed(size_t bytes) {
673 shenandoah_assert_heaplocked_or_safepoint();
674 _committed += bytes;
675 }
676
677 void ShenandoahHeap::decrease_committed(size_t bytes) {
678 shenandoah_assert_heaplocked_or_safepoint();
679 _committed -= bytes;
680 }
681
682 // For tracking usage based on allocations, it should be the case that:
683 // * The sum of regions::used == heap::used
684 // * The sum of a generation's regions::used == generation::used
685 // * The sum of a generation's humongous regions::free == generation::humongous_waste
686 // These invariants are checked by the verifier on GC safepoints.
687 //
688 // Additional notes:
689 // * When a mutator's allocation request causes a region to be retired, the
690 // free memory left in that region is considered waste. It does not contribute
691 // to the usage, but it _does_ contribute to allocation rate.
692 // * The bottom of a PLAB must be aligned on card size. In some cases this will
693 // require padding in front of the PLAB (a filler object). Because this padding
694 // is included in the region's used memory we include the padding in the usage
695 // accounting as waste.
696 // * Mutator allocations are used to compute an allocation rate. They are also
697 // sent to the Pacer for those purposes.
698 // * There are three sources of waste:
699 // 1. The padding used to align a PLAB on card size
700 // 2. Region's free is less than minimum TLAB size and is retired
701 // 3. The unused portion of memory in the last region of a humongous object
702 void ShenandoahHeap::increase_used(const ShenandoahAllocRequest& req) {
703 size_t actual_bytes = req.actual_size() * HeapWordSize;
704 size_t wasted_bytes = req.waste() * HeapWordSize;
705 ShenandoahGeneration* generation = generation_for(req.affiliation());
706
707 if (req.is_gc_alloc()) {
708 assert(wasted_bytes == 0 || req.type() == ShenandoahAllocRequest::_alloc_plab, "Only PLABs have waste");
709 increase_used(generation, actual_bytes + wasted_bytes);
710 } else {
711 assert(req.is_mutator_alloc(), "Expected mutator alloc here");
712 // padding and actual size both count towards allocation counter
713 generation->increase_allocated(actual_bytes + wasted_bytes);
714
715 // only actual size counts toward usage for mutator allocations
716 increase_used(generation, actual_bytes);
717
718 // notify pacer of both actual size and waste
719 notify_mutator_alloc_words(req.actual_size(), req.waste());
720
721 if (wasted_bytes > 0 && ShenandoahHeapRegion::requires_humongous(req.actual_size())) {
722 increase_humongous_waste(generation,wasted_bytes);
723 }
724 }
725 }
726
727 void ShenandoahHeap::increase_humongous_waste(ShenandoahGeneration* generation, size_t bytes) {
728 generation->increase_humongous_waste(bytes);
729 if (!generation->is_global()) {
730 global_generation()->increase_humongous_waste(bytes);
731 }
732 }
733
734 void ShenandoahHeap::decrease_humongous_waste(ShenandoahGeneration* generation, size_t bytes) {
735 generation->decrease_humongous_waste(bytes);
736 if (!generation->is_global()) {
737 global_generation()->decrease_humongous_waste(bytes);
738 }
739 }
740
741 void ShenandoahHeap::increase_used(ShenandoahGeneration* generation, size_t bytes) {
742 generation->increase_used(bytes);
743 if (!generation->is_global()) {
744 global_generation()->increase_used(bytes);
745 }
746 }
747
748 void ShenandoahHeap::decrease_used(ShenandoahGeneration* generation, size_t bytes) {
749 generation->decrease_used(bytes);
750 if (!generation->is_global()) {
751 global_generation()->decrease_used(bytes);
752 }
753 }
754
755 void ShenandoahHeap::notify_mutator_alloc_words(size_t words, size_t waste) {
756 if (ShenandoahPacing) {
757 control_thread()->pacing_notify_alloc(words);
758 if (waste > 0) {
759 pacer()->claim_for_alloc<true>(waste);
760 }
761 }
762 }
763
764 size_t ShenandoahHeap::capacity() const {
765 return committed();
766 }
767
768 size_t ShenandoahHeap::max_capacity() const {
769 return _num_regions * ShenandoahHeapRegion::region_size_bytes();
770 }
771
772 size_t ShenandoahHeap::soft_max_capacity() const {
773 size_t v = Atomic::load(&_soft_max_size);
774 assert(min_capacity() <= v && v <= max_capacity(),
775 "Should be in bounds: " SIZE_FORMAT " <= " SIZE_FORMAT " <= " SIZE_FORMAT,
776 min_capacity(), v, max_capacity());
777 return v;
778 }
779
780 void ShenandoahHeap::set_soft_max_capacity(size_t v) {
781 assert(min_capacity() <= v && v <= max_capacity(),
782 "Should be in bounds: " SIZE_FORMAT " <= " SIZE_FORMAT " <= " SIZE_FORMAT,
783 min_capacity(), v, max_capacity());
784 Atomic::store(&_soft_max_size, v);
785 }
786
787 size_t ShenandoahHeap::min_capacity() const {
788 return _minimum_size;
789 }
790
791 size_t ShenandoahHeap::initial_capacity() const {
792 return _initial_size;
793 }
794
795 bool ShenandoahHeap::is_in(const void* p) const {
796 if (is_in_reserved(p)) {
797 if (is_full_gc_move_in_progress()) {
798 // Full GC move is running, we do not have a consistent region
799 // information yet. But we know the pointer is in heap.
800 return true;
801 }
802 // Now check if we point to a live section in active region.
803 ShenandoahHeapRegion* r = heap_region_containing(p);
804 return (r->is_active() && p < r->top());
805 } else {
806 return false;
807 }
808 }
809
810 void ShenandoahHeap::notify_soft_max_changed() {
811 if (_uncommit_thread != nullptr) {
812 _uncommit_thread->notify_soft_max_changed();
813 }
814 }
815
816 void ShenandoahHeap::notify_explicit_gc_requested() {
817 if (_uncommit_thread != nullptr) {
818 _uncommit_thread->notify_explicit_gc_requested();
819 }
820 }
821
822 bool ShenandoahHeap::check_soft_max_changed() {
823 size_t new_soft_max = Atomic::load(&SoftMaxHeapSize);
824 size_t old_soft_max = soft_max_capacity();
825 if (new_soft_max != old_soft_max) {
826 new_soft_max = MAX2(min_capacity(), new_soft_max);
827 new_soft_max = MIN2(max_capacity(), new_soft_max);
828 if (new_soft_max != old_soft_max) {
829 log_info(gc)("Soft Max Heap Size: " SIZE_FORMAT "%s -> " SIZE_FORMAT "%s",
830 byte_size_in_proper_unit(old_soft_max), proper_unit_for_byte_size(old_soft_max),
831 byte_size_in_proper_unit(new_soft_max), proper_unit_for_byte_size(new_soft_max)
832 );
833 set_soft_max_capacity(new_soft_max);
834 return true;
835 }
836 }
837 return false;
838 }
839
840 void ShenandoahHeap::notify_heap_changed() {
841 // Update monitoring counters when we took a new region. This amortizes the
842 // update costs on slow path.
843 monitoring_support()->notify_heap_changed();
844 _heap_changed.try_set();
845 }
846
847 void ShenandoahHeap::set_forced_counters_update(bool value) {
848 monitoring_support()->set_forced_counters_update(value);
849 }
850
851 void ShenandoahHeap::handle_force_counters_update() {
852 monitoring_support()->handle_force_counters_update();
853 }
854
855 HeapWord* ShenandoahHeap::allocate_from_gclab_slow(Thread* thread, size_t size) {
856 // New object should fit the GCLAB size
857 size_t min_size = MAX2(size, PLAB::min_size());
858
859 // Figure out size of new GCLAB, looking back at heuristics. Expand aggressively.
860 size_t new_size = ShenandoahThreadLocalData::gclab_size(thread) * 2;
861
862 new_size = MIN2(new_size, PLAB::max_size());
863 new_size = MAX2(new_size, PLAB::min_size());
864
865 // Record new heuristic value even if we take any shortcut. This captures
866 // the case when moderately-sized objects always take a shortcut. At some point,
867 // heuristics should catch up with them.
868 log_debug(gc, free)("Set new GCLAB size: " SIZE_FORMAT, new_size);
869 ShenandoahThreadLocalData::set_gclab_size(thread, new_size);
870
871 if (new_size < size) {
872 // New size still does not fit the object. Fall back to shared allocation.
873 // This avoids retiring perfectly good GCLABs, when we encounter a large object.
874 log_debug(gc, free)("New gclab size (" SIZE_FORMAT ") is too small for " SIZE_FORMAT, new_size, size);
875 return nullptr;
876 }
877
878 // Retire current GCLAB, and allocate a new one.
879 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
880 gclab->retire();
881
882 size_t actual_size = 0;
883 HeapWord* gclab_buf = allocate_new_gclab(min_size, new_size, &actual_size);
884 if (gclab_buf == nullptr) {
885 return nullptr;
886 }
887
888 assert (size <= actual_size, "allocation should fit");
889
890 // ...and clear or zap just allocated TLAB, if needed.
891 if (ZeroTLAB) {
892 Copy::zero_to_words(gclab_buf, actual_size);
893 } else if (ZapTLAB) {
894 // Skip mangling the space corresponding to the object header to
895 // ensure that the returned space is not considered parsable by
896 // any concurrent GC thread.
897 size_t hdr_size = oopDesc::header_size();
898 Copy::fill_to_words(gclab_buf + hdr_size, actual_size - hdr_size, badHeapWordVal);
899 }
900 gclab->set_buf(gclab_buf, actual_size);
901 return gclab->allocate(size);
902 }
903
904 // Called from stubs in JIT code or interpreter
905 HeapWord* ShenandoahHeap::allocate_new_tlab(size_t min_size,
906 size_t requested_size,
907 size_t* actual_size) {
908 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_tlab(min_size, requested_size);
909 HeapWord* res = allocate_memory(req);
910 if (res != nullptr) {
911 *actual_size = req.actual_size();
912 } else {
913 *actual_size = 0;
914 }
915 return res;
916 }
917
918 HeapWord* ShenandoahHeap::allocate_new_gclab(size_t min_size,
919 size_t word_size,
920 size_t* actual_size) {
921 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_gclab(min_size, word_size);
922 HeapWord* res = allocate_memory(req);
923 if (res != nullptr) {
924 *actual_size = req.actual_size();
925 } else {
926 *actual_size = 0;
927 }
928 return res;
929 }
930
931 HeapWord* ShenandoahHeap::allocate_memory(ShenandoahAllocRequest& req) {
932 intptr_t pacer_epoch = 0;
933 bool in_new_region = false;
934 HeapWord* result = nullptr;
935
936 if (req.is_mutator_alloc()) {
937 if (ShenandoahPacing) {
938 pacer()->pace_for_alloc(req.size());
939 pacer_epoch = pacer()->epoch();
940 }
941
942 if (!ShenandoahAllocFailureALot || !should_inject_alloc_failure()) {
943 result = allocate_memory_under_lock(req, in_new_region);
944 }
945
946 // Check that gc overhead is not exceeded.
947 //
948 // Shenandoah will grind along for quite a while allocating one
949 // object at a time using shared (non-tlab) allocations. This check
950 // is testing that the GC overhead limit has not been exceeded.
951 // This will notify the collector to start a cycle, but will raise
952 // an OOME to the mutator if the last Full GCs have not made progress.
953 // gc_no_progress_count is incremented following each degen or full GC that fails to achieve is_good_progress().
954 if (result == nullptr && !req.is_lab_alloc() && get_gc_no_progress_count() > ShenandoahNoProgressThreshold) {
955 control_thread()->handle_alloc_failure(req, false);
956 req.set_actual_size(0);
957 return nullptr;
958 }
959
960 if (result == nullptr) {
961 // Block until control thread reacted, then retry allocation.
962 //
963 // It might happen that one of the threads requesting allocation would unblock
964 // way later after GC happened, only to fail the second allocation, because
965 // other threads have already depleted the free storage. In this case, a better
966 // strategy is to try again, until at least one full GC has completed.
967 //
968 // Stop retrying and return nullptr to cause OOMError exception if our allocation failed even after:
969 // a) We experienced a GC that had good progress, or
970 // b) We experienced at least one Full GC (whether or not it had good progress)
971
972 size_t original_count = shenandoah_policy()->full_gc_count();
973 while ((result == nullptr) && (original_count == shenandoah_policy()->full_gc_count())) {
974 control_thread()->handle_alloc_failure(req, true);
975 result = allocate_memory_under_lock(req, in_new_region);
976 }
977 if (result != nullptr) {
978 // If our allocation request has been satisifed after it initially failed, we count this as good gc progress
979 notify_gc_progress();
980 }
981 if (log_develop_is_enabled(Debug, gc, alloc)) {
982 ResourceMark rm;
983 log_debug(gc, alloc)("Thread: %s, Result: " PTR_FORMAT ", Request: %s, Size: " SIZE_FORMAT
984 ", Original: " SIZE_FORMAT ", Latest: " SIZE_FORMAT,
985 Thread::current()->name(), p2i(result), req.type_string(), req.size(),
986 original_count, get_gc_no_progress_count());
987 }
988 }
989 } else {
990 assert(req.is_gc_alloc(), "Can only accept GC allocs here");
991 result = allocate_memory_under_lock(req, in_new_region);
992 // Do not call handle_alloc_failure() here, because we cannot block.
993 // The allocation failure would be handled by the LRB slowpath with handle_alloc_failure_evac().
994 }
995
996 if (in_new_region) {
997 notify_heap_changed();
998 }
999
1000 if (result == nullptr) {
1001 req.set_actual_size(0);
1002 }
1003
1004 // This is called regardless of the outcome of the allocation to account
1005 // for any waste created by retiring regions with this request.
1006 increase_used(req);
1007
1008 if (result != nullptr) {
1009 size_t requested = req.size();
1010 size_t actual = req.actual_size();
1011
1012 assert (req.is_lab_alloc() || (requested == actual),
1013 "Only LAB allocations are elastic: %s, requested = " SIZE_FORMAT ", actual = " SIZE_FORMAT,
1014 ShenandoahAllocRequest::alloc_type_to_string(req.type()), requested, actual);
1015
1016 if (req.is_mutator_alloc()) {
1017 // If we requested more than we were granted, give the rest back to pacer.
1018 // This only matters if we are in the same pacing epoch: do not try to unpace
1019 // over the budget for the other phase.
1020 if (ShenandoahPacing && (pacer_epoch > 0) && (requested > actual)) {
1021 pacer()->unpace_for_alloc(pacer_epoch, requested - actual);
1022 }
1023 }
1024 }
1025
1026 return result;
1027 }
1028
1029 HeapWord* ShenandoahHeap::allocate_memory_under_lock(ShenandoahAllocRequest& req, bool& in_new_region) {
1030 // If we are dealing with mutator allocation, then we may need to block for safepoint.
1031 // We cannot block for safepoint for GC allocations, because there is a high chance
1032 // we are already running at safepoint or from stack watermark machinery, and we cannot
1033 // block again.
1034 ShenandoahHeapLocker locker(lock(), req.is_mutator_alloc());
1035
1036 // Make sure the old generation has room for either evacuations or promotions before trying to allocate.
1037 if (req.is_old() && !old_generation()->can_allocate(req)) {
1038 return nullptr;
1039 }
1040
1041 // If TLAB request size is greater than available, allocate() will attempt to downsize request to fit within available
1042 // memory.
1043 HeapWord* result = _free_set->allocate(req, in_new_region);
1044
1045 // Record the plab configuration for this result and register the object.
1046 if (result != nullptr && req.is_old()) {
1047 old_generation()->configure_plab_for_current_thread(req);
1048 if (req.type() == ShenandoahAllocRequest::_alloc_shared_gc) {
1049 // Register the newly allocated object while we're holding the global lock since there's no synchronization
1050 // built in to the implementation of register_object(). There are potential races when multiple independent
1051 // threads are allocating objects, some of which might span the same card region. For example, consider
1052 // a card table's memory region within which three objects are being allocated by three different threads:
1053 //
1054 // objects being "concurrently" allocated:
1055 // [-----a------][-----b-----][--------------c------------------]
1056 // [---- card table memory range --------------]
1057 //
1058 // Before any objects are allocated, this card's memory range holds no objects. Note that allocation of object a
1059 // wants to set the starts-object, first-start, and last-start attributes of the preceding card region.
1060 // Allocation of object b wants to set the starts-object, first-start, and last-start attributes of this card region.
1061 // Allocation of object c also wants to set the starts-object, first-start, and last-start attributes of this
1062 // card region.
1063 //
1064 // The thread allocating b and the thread allocating c can "race" in various ways, resulting in confusion, such as
1065 // last-start representing object b while first-start represents object c. This is why we need to require all
1066 // register_object() invocations to be "mutually exclusive" with respect to each card's memory range.
1067 old_generation()->card_scan()->register_object(result);
1068 }
1069 }
1070
1071 return result;
1072 }
1073
1074 HeapWord* ShenandoahHeap::mem_allocate(size_t size,
1075 bool* gc_overhead_limit_was_exceeded) {
1076 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared(size);
1077 return allocate_memory(req);
1078 }
1079
1080 MetaWord* ShenandoahHeap::satisfy_failed_metadata_allocation(ClassLoaderData* loader_data,
1081 size_t size,
1082 Metaspace::MetadataType mdtype) {
1083 MetaWord* result;
1084
1085 // Inform metaspace OOM to GC heuristics if class unloading is possible.
1086 ShenandoahHeuristics* h = global_generation()->heuristics();
1087 if (h->can_unload_classes()) {
1088 h->record_metaspace_oom();
1089 }
1090
1091 // Expand and retry allocation
1092 result = loader_data->metaspace_non_null()->expand_and_allocate(size, mdtype);
1093 if (result != nullptr) {
1094 return result;
1095 }
1096
1097 // Start full GC
1098 collect(GCCause::_metadata_GC_clear_soft_refs);
1099
1100 // Retry allocation
1101 result = loader_data->metaspace_non_null()->allocate(size, mdtype);
1102 if (result != nullptr) {
1103 return result;
1104 }
1105
1106 // Expand and retry allocation
1107 result = loader_data->metaspace_non_null()->expand_and_allocate(size, mdtype);
1108 if (result != nullptr) {
1109 return result;
1110 }
1111
1112 // Out of memory
1113 return nullptr;
1114 }
1115
1116 class ShenandoahConcurrentEvacuateRegionObjectClosure : public ObjectClosure {
1117 private:
1118 ShenandoahHeap* const _heap;
1119 Thread* const _thread;
1120 public:
1121 ShenandoahConcurrentEvacuateRegionObjectClosure(ShenandoahHeap* heap) :
1122 _heap(heap), _thread(Thread::current()) {}
1123
1124 void do_object(oop p) {
1125 shenandoah_assert_marked(nullptr, p);
1126 if (!p->is_forwarded()) {
1127 _heap->evacuate_object(p, _thread);
1128 }
1129 }
1130 };
1131
1132 class ShenandoahEvacuationTask : public WorkerTask {
1133 private:
1134 ShenandoahHeap* const _sh;
1135 ShenandoahCollectionSet* const _cs;
1136 bool _concurrent;
1137 public:
1138 ShenandoahEvacuationTask(ShenandoahHeap* sh,
1139 ShenandoahCollectionSet* cs,
1140 bool concurrent) :
1141 WorkerTask("Shenandoah Evacuation"),
1142 _sh(sh),
1143 _cs(cs),
1144 _concurrent(concurrent)
1145 {}
1146
1147 void work(uint worker_id) {
1148 if (_concurrent) {
1149 ShenandoahConcurrentWorkerSession worker_session(worker_id);
1150 ShenandoahSuspendibleThreadSetJoiner stsj;
1151 ShenandoahEvacOOMScope oom_evac_scope;
1152 do_work();
1153 } else {
1154 ShenandoahParallelWorkerSession worker_session(worker_id);
1155 ShenandoahEvacOOMScope oom_evac_scope;
1156 do_work();
1157 }
1158 }
1159
1160 private:
1161 void do_work() {
1162 ShenandoahConcurrentEvacuateRegionObjectClosure cl(_sh);
1163 ShenandoahHeapRegion* r;
1164 while ((r =_cs->claim_next()) != nullptr) {
1165 assert(r->has_live(), "Region " SIZE_FORMAT " should have been reclaimed early", r->index());
1166 _sh->marked_object_iterate(r, &cl);
1167
1168 if (ShenandoahPacing) {
1169 _sh->pacer()->report_evac(r->used() >> LogHeapWordSize);
1170 }
1171
1172 if (_sh->check_cancelled_gc_and_yield(_concurrent)) {
1173 break;
1174 }
1175 }
1176 }
1177 };
1178
1179 void ShenandoahHeap::evacuate_collection_set(bool concurrent) {
1180 ShenandoahEvacuationTask task(this, _collection_set, concurrent);
1181 workers()->run_task(&task);
1182 }
1183
1184 oop ShenandoahHeap::evacuate_object(oop p, Thread* thread) {
1185 assert(thread == Thread::current(), "Expected thread parameter to be current thread.");
1186 if (ShenandoahThreadLocalData::is_oom_during_evac(thread)) {
1187 // This thread went through the OOM during evac protocol. It is safe to return
1188 // the forward pointer. It must not attempt to evacuate any other objects.
1189 return ShenandoahBarrierSet::resolve_forwarded(p);
1190 }
1191
1192 assert(ShenandoahThreadLocalData::is_evac_allowed(thread), "must be enclosed in oom-evac scope");
1193
1194 ShenandoahHeapRegion* r = heap_region_containing(p);
1195 assert(!r->is_humongous(), "never evacuate humongous objects");
1196
1197 ShenandoahAffiliation target_gen = r->affiliation();
1198 return try_evacuate_object(p, thread, r, target_gen);
1199 }
1200
1201 oop ShenandoahHeap::try_evacuate_object(oop p, Thread* thread, ShenandoahHeapRegion* from_region,
1202 ShenandoahAffiliation target_gen) {
1203 assert(target_gen == YOUNG_GENERATION, "Only expect evacuations to young in this mode");
1204 assert(from_region->is_young(), "Only expect evacuations from young in this mode");
1205 bool alloc_from_lab = true;
1206 HeapWord* copy = nullptr;
1207 size_t size = ShenandoahForwarding::size(p);
1208
1209 #ifdef ASSERT
1210 if (ShenandoahOOMDuringEvacALot &&
1211 (os::random() & 1) == 0) { // Simulate OOM every ~2nd slow-path call
1212 copy = nullptr;
1213 } else {
1214 #endif
1215 if (UseTLAB) {
1216 copy = allocate_from_gclab(thread, size);
1217 }
1218 if (copy == nullptr) {
1219 // If we failed to allocate in LAB, we'll try a shared allocation.
1220 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared_gc(size, target_gen);
1221 copy = allocate_memory(req);
1222 alloc_from_lab = false;
1223 }
1224 #ifdef ASSERT
1225 }
1226 #endif
1227
1228 if (copy == nullptr) {
1229 control_thread()->handle_alloc_failure_evac(size);
1230
1231 _oom_evac_handler.handle_out_of_memory_during_evacuation();
1232
1233 return ShenandoahBarrierSet::resolve_forwarded(p);
1234 }
1235
1236 // Copy the object:
1237 Copy::aligned_disjoint_words(cast_from_oop<HeapWord*>(p), copy, size);
1238
1239 // Try to install the new forwarding pointer.
1240 oop copy_val = cast_to_oop(copy);
1241 oop result = ShenandoahForwarding::try_update_forwardee(p, copy_val);
1242 if (result == copy_val) {
1243 // Successfully evacuated. Our copy is now the public one!
1244 ContinuationGCSupport::relativize_stack_chunk(copy_val);
1245 shenandoah_assert_correct(nullptr, copy_val);
1246 return copy_val;
1247 } else {
1248 // Failed to evacuate. We need to deal with the object that is left behind. Since this
1249 // new allocation is certainly after TAMS, it will be considered live in the next cycle.
1250 // But if it happens to contain references to evacuated regions, those references would
1251 // not get updated for this stale copy during this cycle, and we will crash while scanning
1252 // it the next cycle.
1253 if (alloc_from_lab) {
1254 // For LAB allocations, it is enough to rollback the allocation ptr. Either the next
1255 // object will overwrite this stale copy, or the filler object on LAB retirement will
1256 // do this.
1257 ShenandoahThreadLocalData::gclab(thread)->undo_allocation(copy, size);
1258 } else {
1259 // For non-LAB allocations, we have no way to retract the allocation, and
1260 // have to explicitly overwrite the copy with the filler object. With that overwrite,
1261 // we have to keep the fwdptr initialized and pointing to our (stale) copy.
1262 assert(size >= ShenandoahHeap::min_fill_size(), "previously allocated object known to be larger than min_size");
1263 fill_with_object(copy, size);
1264 shenandoah_assert_correct(nullptr, copy_val);
1265 // For non-LAB allocations, the object has already been registered
1266 }
1267 shenandoah_assert_correct(nullptr, result);
1268 return result;
1269 }
1270 }
1271
1272 void ShenandoahHeap::trash_cset_regions() {
1273 ShenandoahHeapLocker locker(lock());
1274
1275 ShenandoahCollectionSet* set = collection_set();
1276 ShenandoahHeapRegion* r;
1277 set->clear_current_index();
1278 while ((r = set->next()) != nullptr) {
1279 r->make_trash();
1280 }
1281 collection_set()->clear();
1282 }
1283
1284 void ShenandoahHeap::print_heap_regions_on(outputStream* st) const {
1285 st->print_cr("Heap Regions:");
1286 st->print_cr("Region state: EU=empty-uncommitted, EC=empty-committed, R=regular, H=humongous start, HP=pinned humongous start");
1287 st->print_cr(" HC=humongous continuation, CS=collection set, TR=trash, P=pinned, CSP=pinned collection set");
1288 st->print_cr("BTE=bottom/top/end, TAMS=top-at-mark-start");
1289 st->print_cr("UWM=update watermark, U=used");
1290 st->print_cr("T=TLAB allocs, G=GCLAB allocs");
1291 st->print_cr("S=shared allocs, L=live data");
1292 st->print_cr("CP=critical pins");
1293
1294 for (size_t i = 0; i < num_regions(); i++) {
1295 get_region(i)->print_on(st);
1296 }
1297 }
1298
1299 size_t ShenandoahHeap::trash_humongous_region_at(ShenandoahHeapRegion* start) {
1300 assert(start->is_humongous_start(), "reclaim regions starting with the first one");
1301
1302 oop humongous_obj = cast_to_oop(start->bottom());
1303 size_t size = humongous_obj->size();
1304 size_t required_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize);
1305 size_t index = start->index() + required_regions - 1;
1306
1307 assert(!start->has_live(), "liveness must be zero");
1308
1309 for(size_t i = 0; i < required_regions; i++) {
1310 // Reclaim from tail. Otherwise, assertion fails when printing region to trace log,
1311 // as it expects that every region belongs to a humongous region starting with a humongous start region.
1312 ShenandoahHeapRegion* region = get_region(index --);
1313
1314 assert(region->is_humongous(), "expect correct humongous start or continuation");
1315 assert(!region->is_cset(), "Humongous region should not be in collection set");
1316
1317 region->make_trash_immediate();
1318 }
1319 return required_regions;
1320 }
1321
1322 class ShenandoahCheckCleanGCLABClosure : public ThreadClosure {
1323 public:
1324 ShenandoahCheckCleanGCLABClosure() {}
1325 void do_thread(Thread* thread) {
1326 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
1327 assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name());
1328 assert(gclab->words_remaining() == 0, "GCLAB should not need retirement");
1329
1330 if (ShenandoahHeap::heap()->mode()->is_generational()) {
1331 PLAB* plab = ShenandoahThreadLocalData::plab(thread);
1332 assert(plab != nullptr, "PLAB should be initialized for %s", thread->name());
1333 assert(plab->words_remaining() == 0, "PLAB should not need retirement");
1334 }
1335 }
1336 };
1337
1338 class ShenandoahRetireGCLABClosure : public ThreadClosure {
1339 private:
1340 bool const _resize;
1341 public:
1342 ShenandoahRetireGCLABClosure(bool resize) : _resize(resize) {}
1343 void do_thread(Thread* thread) {
1344 PLAB* gclab = ShenandoahThreadLocalData::gclab(thread);
1345 assert(gclab != nullptr, "GCLAB should be initialized for %s", thread->name());
1346 gclab->retire();
1347 if (_resize && ShenandoahThreadLocalData::gclab_size(thread) > 0) {
1348 ShenandoahThreadLocalData::set_gclab_size(thread, 0);
1349 }
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
1355 // There are two reasons to retire all plabs between old-gen evacuation passes.
1356 // 1. We need to make the plab memory parsable by remembered-set scanning.
1357 // 2. We need to establish a trustworthy UpdateWaterMark value within each old-gen heap region
1358 ShenandoahGenerationalHeap::heap()->retire_plab(plab, thread);
1359 if (_resize && ShenandoahThreadLocalData::plab_size(thread) > 0) {
1360 ShenandoahThreadLocalData::set_plab_size(thread, 0);
1361 }
1362 }
1363 }
1364 };
1365
1366 void ShenandoahHeap::labs_make_parsable() {
1367 assert(UseTLAB, "Only call with UseTLAB");
1368
1369 ShenandoahRetireGCLABClosure cl(false);
1370
1371 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1372 ThreadLocalAllocBuffer& tlab = t->tlab();
1373 tlab.make_parsable();
1374 cl.do_thread(t);
1375 }
1376
1377 workers()->threads_do(&cl);
1378 }
1379
1380 void ShenandoahHeap::tlabs_retire(bool resize) {
1381 assert(UseTLAB, "Only call with UseTLAB");
1382 assert(!resize || ResizeTLAB, "Only call for resize when ResizeTLAB is enabled");
1383
1384 ThreadLocalAllocStats stats;
1385
1386 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1387 ThreadLocalAllocBuffer& tlab = t->tlab();
1388 tlab.retire(&stats);
1389 if (resize) {
1390 tlab.resize();
1391 }
1392 }
1393
1394 stats.publish();
1395
1396 #ifdef ASSERT
1397 ShenandoahCheckCleanGCLABClosure cl;
1398 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1399 cl.do_thread(t);
1400 }
1401 workers()->threads_do(&cl);
1402 #endif
1403 }
1404
1405 void ShenandoahHeap::gclabs_retire(bool resize) {
1406 assert(UseTLAB, "Only call with UseTLAB");
1407 assert(!resize || ResizeTLAB, "Only call for resize when ResizeTLAB is enabled");
1408
1409 ShenandoahRetireGCLABClosure cl(resize);
1410 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1411 cl.do_thread(t);
1412 }
1413 workers()->threads_do(&cl);
1414
1415 if (safepoint_workers() != nullptr) {
1416 safepoint_workers()->threads_do(&cl);
1417 }
1418 }
1419
1420 // Returns size in bytes
1421 size_t ShenandoahHeap::unsafe_max_tlab_alloc(Thread *thread) const {
1422 // Return the max allowed size, and let the allocation path
1423 // figure out the safe size for current allocation.
1424 return ShenandoahHeapRegion::max_tlab_size_bytes();
1425 }
1426
1427 size_t ShenandoahHeap::max_tlab_size() const {
1428 // Returns size in words
1429 return ShenandoahHeapRegion::max_tlab_size_words();
1430 }
1431
1432 void ShenandoahHeap::collect(GCCause::Cause cause) {
1433 control_thread()->request_gc(cause);
1434 }
1435
1436 void ShenandoahHeap::do_full_collection(bool clear_all_soft_refs) {
1437 //assert(false, "Shouldn't need to do full collections");
1438 }
1439
1440 HeapWord* ShenandoahHeap::block_start(const void* addr) const {
1441 ShenandoahHeapRegion* r = heap_region_containing(addr);
1442 if (r != nullptr) {
1443 return r->block_start(addr);
1444 }
1445 return nullptr;
1446 }
1447
1448 bool ShenandoahHeap::block_is_obj(const HeapWord* addr) const {
1449 ShenandoahHeapRegion* r = heap_region_containing(addr);
1450 return r->block_is_obj(addr);
1451 }
1452
1453 bool ShenandoahHeap::print_location(outputStream* st, void* addr) const {
1454 return BlockLocationPrinter<ShenandoahHeap>::print_location(st, addr);
1455 }
1456
1457 void ShenandoahHeap::prepare_for_verify() {
1458 if (SafepointSynchronize::is_at_safepoint() && UseTLAB) {
1459 labs_make_parsable();
1460 }
1461 }
1462
1463 void ShenandoahHeap::gc_threads_do(ThreadClosure* tcl) const {
1464 if (_shenandoah_policy->is_at_shutdown()) {
1465 return;
1466 }
1467
1468 if (_control_thread != nullptr) {
1469 tcl->do_thread(_control_thread);
1470 }
1471
1472 if (_uncommit_thread != nullptr) {
1473 tcl->do_thread(_uncommit_thread);
1474 }
1475
1476 workers()->threads_do(tcl);
1477 if (_safepoint_workers != nullptr) {
1478 _safepoint_workers->threads_do(tcl);
1479 }
1480 }
1481
1482 void ShenandoahHeap::print_tracing_info() const {
1483 LogTarget(Info, gc, stats) lt;
1484 if (lt.is_enabled()) {
1485 ResourceMark rm;
1486 LogStream ls(lt);
1487
1488 phase_timings()->print_global_on(&ls);
1489
1490 ls.cr();
1491 ls.cr();
1492
1493 shenandoah_policy()->print_gc_stats(&ls);
1494
1495 ls.cr();
1496 ls.cr();
1497 }
1498 }
1499
1500 void ShenandoahHeap::set_gc_generation(ShenandoahGeneration* generation) {
1501 shenandoah_assert_control_or_vm_thread_at_safepoint();
1502 _gc_generation = generation;
1503 }
1504
1505 // Active generation may only be set by the VM thread at a safepoint.
1506 void ShenandoahHeap::set_active_generation() {
1507 assert(Thread::current()->is_VM_thread(), "Only the VM Thread");
1508 assert(SafepointSynchronize::is_at_safepoint(), "Only at a safepoint!");
1509 assert(_gc_generation != nullptr, "Will set _active_generation to nullptr");
1510 _active_generation = _gc_generation;
1511 }
1512
1513 void ShenandoahHeap::on_cycle_start(GCCause::Cause cause, ShenandoahGeneration* generation) {
1514 shenandoah_policy()->record_collection_cause(cause);
1515
1516 assert(gc_cause() == GCCause::_no_gc, "Over-writing cause");
1517 assert(_gc_generation == nullptr, "Over-writing _gc_generation");
1518
1519 set_gc_cause(cause);
1520 set_gc_generation(generation);
1521
1522 generation->heuristics()->record_cycle_start();
1523 }
1524
1525 void ShenandoahHeap::on_cycle_end(ShenandoahGeneration* generation) {
1526 assert(gc_cause() != GCCause::_no_gc, "cause wasn't set");
1527 assert(_gc_generation != nullptr, "_gc_generation wasn't set");
1528
1529 generation->heuristics()->record_cycle_end();
1530 if (mode()->is_generational() && generation->is_global()) {
1531 // If we just completed a GLOBAL GC, claim credit for completion of young-gen and old-gen GC as well
1532 young_generation()->heuristics()->record_cycle_end();
1533 old_generation()->heuristics()->record_cycle_end();
1534 }
1535
1536 set_gc_generation(nullptr);
1537 set_gc_cause(GCCause::_no_gc);
1538 }
1539
1540 void ShenandoahHeap::verify(VerifyOption vo) {
1541 if (ShenandoahSafepoint::is_at_shenandoah_safepoint()) {
1542 if (ShenandoahVerify) {
1543 verifier()->verify_generic(vo);
1544 } else {
1545 // TODO: Consider allocating verification bitmaps on demand,
1546 // and turn this on unconditionally.
1547 }
1548 }
1549 }
1550 size_t ShenandoahHeap::tlab_capacity(Thread *thr) const {
1551 return _free_set->capacity();
1552 }
1553
1554 class ObjectIterateScanRootClosure : public BasicOopIterateClosure {
1555 private:
1556 MarkBitMap* _bitmap;
1557 ShenandoahScanObjectStack* _oop_stack;
1558 ShenandoahHeap* const _heap;
1559 ShenandoahMarkingContext* const _marking_context;
1560
1561 template <class T>
1562 void do_oop_work(T* p) {
1563 T o = RawAccess<>::oop_load(p);
1564 if (!CompressedOops::is_null(o)) {
1565 oop obj = CompressedOops::decode_not_null(o);
1566 if (_heap->is_concurrent_weak_root_in_progress() && !_marking_context->is_marked(obj)) {
1567 // There may be dead oops in weak roots in concurrent root phase, do not touch them.
1568 return;
1569 }
1570 obj = ShenandoahBarrierSet::barrier_set()->load_reference_barrier(obj);
1571
1572 assert(oopDesc::is_oop(obj), "must be a valid oop");
1573 if (!_bitmap->is_marked(obj)) {
1574 _bitmap->mark(obj);
1575 _oop_stack->push(obj);
1576 }
1577 }
1578 }
1579 public:
1580 ObjectIterateScanRootClosure(MarkBitMap* bitmap, ShenandoahScanObjectStack* oop_stack) :
1581 _bitmap(bitmap), _oop_stack(oop_stack), _heap(ShenandoahHeap::heap()),
1582 _marking_context(_heap->marking_context()) {}
1583 void do_oop(oop* p) { do_oop_work(p); }
1584 void do_oop(narrowOop* p) { do_oop_work(p); }
1585 };
1586
1587 /*
1588 * This is public API, used in preparation of object_iterate().
1589 * Since we don't do linear scan of heap in object_iterate() (see comment below), we don't
1590 * need to make the heap parsable. For Shenandoah-internal linear heap scans that we can
1591 * control, we call SH::tlabs_retire, SH::gclabs_retire.
1592 */
1593 void ShenandoahHeap::ensure_parsability(bool retire_tlabs) {
1594 // No-op.
1595 }
1596
1597 /*
1598 * Iterates objects in the heap. This is public API, used for, e.g., heap dumping.
1599 *
1600 * We cannot safely iterate objects by doing a linear scan at random points in time. Linear
1601 * scanning needs to deal with dead objects, which may have dead Klass* pointers (e.g.
1602 * calling oopDesc::size() would crash) or dangling reference fields (crashes) etc. Linear
1603 * scanning therefore depends on having a valid marking bitmap to support it. However, we only
1604 * have a valid marking bitmap after successful marking. In particular, we *don't* have a valid
1605 * marking bitmap during marking, after aborted marking or during/after cleanup (when we just
1606 * wiped the bitmap in preparation for next marking).
1607 *
1608 * For all those reasons, we implement object iteration as a single marking traversal, reporting
1609 * objects as we mark+traverse through the heap, starting from GC roots. JVMTI IterateThroughHeap
1610 * is allowed to report dead objects, but is not required to do so.
1611 */
1612 void ShenandoahHeap::object_iterate(ObjectClosure* cl) {
1613 // Reset bitmap
1614 if (!prepare_aux_bitmap_for_iteration())
1615 return;
1616
1617 ShenandoahScanObjectStack oop_stack;
1618 ObjectIterateScanRootClosure oops(&_aux_bit_map, &oop_stack);
1619 // Seed the stack with root scan
1620 scan_roots_for_iteration(&oop_stack, &oops);
1621
1622 // Work through the oop stack to traverse heap
1623 while (! oop_stack.is_empty()) {
1624 oop obj = oop_stack.pop();
1625 assert(oopDesc::is_oop(obj), "must be a valid oop");
1626 cl->do_object(obj);
1627 obj->oop_iterate(&oops);
1628 }
1629
1630 assert(oop_stack.is_empty(), "should be empty");
1631 // Reclaim bitmap
1632 reclaim_aux_bitmap_for_iteration();
1633 }
1634
1635 bool ShenandoahHeap::prepare_aux_bitmap_for_iteration() {
1636 assert(SafepointSynchronize::is_at_safepoint(), "safe iteration is only available during safepoints");
1637
1638 if (!_aux_bitmap_region_special && !os::commit_memory((char*)_aux_bitmap_region.start(), _aux_bitmap_region.byte_size(), false)) {
1639 log_warning(gc)("Could not commit native memory for auxiliary marking bitmap for heap iteration");
1640 return false;
1641 }
1642 // Reset bitmap
1643 _aux_bit_map.clear();
1644 return true;
1645 }
1646
1647 void ShenandoahHeap::scan_roots_for_iteration(ShenandoahScanObjectStack* oop_stack, ObjectIterateScanRootClosure* oops) {
1648 // Process GC roots according to current GC cycle
1649 // This populates the work stack with initial objects
1650 // It is important to relinquish the associated locks before diving
1651 // into heap dumper
1652 uint n_workers = safepoint_workers() != nullptr ? safepoint_workers()->active_workers() : 1;
1653 ShenandoahHeapIterationRootScanner rp(n_workers);
1654 rp.roots_do(oops);
1655 }
1656
1657 void ShenandoahHeap::reclaim_aux_bitmap_for_iteration() {
1658 if (!_aux_bitmap_region_special && !os::uncommit_memory((char*)_aux_bitmap_region.start(), _aux_bitmap_region.byte_size())) {
1659 log_warning(gc)("Could not uncommit native memory for auxiliary marking bitmap for heap iteration");
1660 }
1661 }
1662
1663 // Closure for parallelly iterate objects
1664 class ShenandoahObjectIterateParScanClosure : public BasicOopIterateClosure {
1665 private:
1666 MarkBitMap* _bitmap;
1667 ShenandoahObjToScanQueue* _queue;
1668 ShenandoahHeap* const _heap;
1669 ShenandoahMarkingContext* const _marking_context;
1670
1671 template <class T>
1672 void do_oop_work(T* p) {
1673 T o = RawAccess<>::oop_load(p);
1674 if (!CompressedOops::is_null(o)) {
1675 oop obj = CompressedOops::decode_not_null(o);
1676 if (_heap->is_concurrent_weak_root_in_progress() && !_marking_context->is_marked(obj)) {
1677 // There may be dead oops in weak roots in concurrent root phase, do not touch them.
1678 return;
1679 }
1680 obj = ShenandoahBarrierSet::barrier_set()->load_reference_barrier(obj);
1681
1682 assert(oopDesc::is_oop(obj), "Must be a valid oop");
1683 if (_bitmap->par_mark(obj)) {
1684 _queue->push(ShenandoahMarkTask(obj));
1685 }
1686 }
1687 }
1688 public:
1689 ShenandoahObjectIterateParScanClosure(MarkBitMap* bitmap, ShenandoahObjToScanQueue* q) :
1690 _bitmap(bitmap), _queue(q), _heap(ShenandoahHeap::heap()),
1691 _marking_context(_heap->marking_context()) {}
1692 void do_oop(oop* p) { do_oop_work(p); }
1693 void do_oop(narrowOop* p) { do_oop_work(p); }
1694 };
1695
1696 // Object iterator for parallel heap iteraion.
1697 // The root scanning phase happenes in construction as a preparation of
1698 // parallel marking queues.
1699 // Every worker processes it's own marking queue. work-stealing is used
1700 // to balance workload.
1701 class ShenandoahParallelObjectIterator : public ParallelObjectIteratorImpl {
1702 private:
1703 uint _num_workers;
1704 bool _init_ready;
1705 MarkBitMap* _aux_bit_map;
1706 ShenandoahHeap* _heap;
1707 ShenandoahScanObjectStack _roots_stack; // global roots stack
1708 ShenandoahObjToScanQueueSet* _task_queues;
1709 public:
1710 ShenandoahParallelObjectIterator(uint num_workers, MarkBitMap* bitmap) :
1711 _num_workers(num_workers),
1712 _init_ready(false),
1713 _aux_bit_map(bitmap),
1714 _heap(ShenandoahHeap::heap()) {
1715 // Initialize bitmap
1716 _init_ready = _heap->prepare_aux_bitmap_for_iteration();
1717 if (!_init_ready) {
1718 return;
1719 }
1720
1721 ObjectIterateScanRootClosure oops(_aux_bit_map, &_roots_stack);
1722 _heap->scan_roots_for_iteration(&_roots_stack, &oops);
1723
1724 _init_ready = prepare_worker_queues();
1725 }
1726
1727 ~ShenandoahParallelObjectIterator() {
1728 // Reclaim bitmap
1729 _heap->reclaim_aux_bitmap_for_iteration();
1730 // Reclaim queue for workers
1731 if (_task_queues!= nullptr) {
1732 for (uint i = 0; i < _num_workers; ++i) {
1733 ShenandoahObjToScanQueue* q = _task_queues->queue(i);
1734 if (q != nullptr) {
1735 delete q;
1736 _task_queues->register_queue(i, nullptr);
1737 }
1738 }
1739 delete _task_queues;
1740 _task_queues = nullptr;
1741 }
1742 }
1743
1744 virtual void object_iterate(ObjectClosure* cl, uint worker_id) {
1745 if (_init_ready) {
1746 object_iterate_parallel(cl, worker_id, _task_queues);
1747 }
1748 }
1749
1750 private:
1751 // Divide global root_stack into worker queues
1752 bool prepare_worker_queues() {
1753 _task_queues = new ShenandoahObjToScanQueueSet((int) _num_workers);
1754 // Initialize queues for every workers
1755 for (uint i = 0; i < _num_workers; ++i) {
1756 ShenandoahObjToScanQueue* task_queue = new ShenandoahObjToScanQueue();
1757 _task_queues->register_queue(i, task_queue);
1758 }
1759 // Divide roots among the workers. Assume that object referencing distribution
1760 // is related with root kind, use round-robin to make every worker have same chance
1761 // to process every kind of roots
1762 size_t roots_num = _roots_stack.size();
1763 if (roots_num == 0) {
1764 // No work to do
1765 return false;
1766 }
1767
1768 for (uint j = 0; j < roots_num; j++) {
1769 uint stack_id = j % _num_workers;
1770 oop obj = _roots_stack.pop();
1771 _task_queues->queue(stack_id)->push(ShenandoahMarkTask(obj));
1772 }
1773 return true;
1774 }
1775
1776 void object_iterate_parallel(ObjectClosure* cl,
1777 uint worker_id,
1778 ShenandoahObjToScanQueueSet* queue_set) {
1779 assert(SafepointSynchronize::is_at_safepoint(), "safe iteration is only available during safepoints");
1780 assert(queue_set != nullptr, "task queue must not be null");
1781
1782 ShenandoahObjToScanQueue* q = queue_set->queue(worker_id);
1783 assert(q != nullptr, "object iterate queue must not be null");
1784
1785 ShenandoahMarkTask t;
1786 ShenandoahObjectIterateParScanClosure oops(_aux_bit_map, q);
1787
1788 // Work through the queue to traverse heap.
1789 // Steal when there is no task in queue.
1790 while (q->pop(t) || queue_set->steal(worker_id, t)) {
1791 oop obj = t.obj();
1792 assert(oopDesc::is_oop(obj), "must be a valid oop");
1793 cl->do_object(obj);
1794 obj->oop_iterate(&oops);
1795 }
1796 assert(q->is_empty(), "should be empty");
1797 }
1798 };
1799
1800 ParallelObjectIteratorImpl* ShenandoahHeap::parallel_object_iterator(uint workers) {
1801 return new ShenandoahParallelObjectIterator(workers, &_aux_bit_map);
1802 }
1803
1804 // Keep alive an object that was loaded with AS_NO_KEEPALIVE.
1805 void ShenandoahHeap::keep_alive(oop obj) {
1806 if (is_concurrent_mark_in_progress() && (obj != nullptr)) {
1807 ShenandoahBarrierSet::barrier_set()->enqueue(obj);
1808 }
1809 }
1810
1811 void ShenandoahHeap::heap_region_iterate(ShenandoahHeapRegionClosure* blk) const {
1812 for (size_t i = 0; i < num_regions(); i++) {
1813 ShenandoahHeapRegion* current = get_region(i);
1814 blk->heap_region_do(current);
1815 }
1816 }
1817
1818 class ShenandoahParallelHeapRegionTask : public WorkerTask {
1819 private:
1820 ShenandoahHeap* const _heap;
1821 ShenandoahHeapRegionClosure* const _blk;
1822 size_t const _stride;
1823
1824 shenandoah_padding(0);
1825 volatile size_t _index;
1826 shenandoah_padding(1);
1827
1828 public:
1829 ShenandoahParallelHeapRegionTask(ShenandoahHeapRegionClosure* blk, size_t stride) :
1830 WorkerTask("Shenandoah Parallel Region Operation"),
1831 _heap(ShenandoahHeap::heap()), _blk(blk), _stride(stride), _index(0) {}
1832
1833 void work(uint worker_id) {
1834 ShenandoahParallelWorkerSession worker_session(worker_id);
1835 size_t stride = _stride;
1836
1837 size_t max = _heap->num_regions();
1838 while (Atomic::load(&_index) < max) {
1839 size_t cur = Atomic::fetch_then_add(&_index, stride, memory_order_relaxed);
1840 size_t start = cur;
1841 size_t end = MIN2(cur + stride, max);
1842 if (start >= max) break;
1843
1844 for (size_t i = cur; i < end; i++) {
1845 ShenandoahHeapRegion* current = _heap->get_region(i);
1846 _blk->heap_region_do(current);
1847 }
1848 }
1849 }
1850 };
1851
1852 void ShenandoahHeap::parallel_heap_region_iterate(ShenandoahHeapRegionClosure* blk) const {
1853 assert(blk->is_thread_safe(), "Only thread-safe closures here");
1854 const uint active_workers = workers()->active_workers();
1855 const size_t n_regions = num_regions();
1856 size_t stride = ShenandoahParallelRegionStride;
1857 if (stride == 0 && active_workers > 1) {
1858 // Automatically derive the stride to balance the work between threads
1859 // evenly. Do not try to split work if below the reasonable threshold.
1860 constexpr size_t threshold = 4096;
1861 stride = n_regions <= threshold ?
1862 threshold :
1863 (n_regions + active_workers - 1) / active_workers;
1864 }
1865
1866 if (n_regions > stride && active_workers > 1) {
1867 ShenandoahParallelHeapRegionTask task(blk, stride);
1868 workers()->run_task(&task);
1869 } else {
1870 heap_region_iterate(blk);
1871 }
1872 }
1873
1874 class ShenandoahRendezvousClosure : public HandshakeClosure {
1875 public:
1876 inline ShenandoahRendezvousClosure(const char* name) : HandshakeClosure(name) {}
1877 inline void do_thread(Thread* thread) {}
1878 };
1879
1880 void ShenandoahHeap::rendezvous_threads(const char* name) {
1881 ShenandoahRendezvousClosure cl(name);
1882 Handshake::execute(&cl);
1883 }
1884
1885 void ShenandoahHeap::recycle_trash() {
1886 free_set()->recycle_trash();
1887 }
1888
1889 void ShenandoahHeap::do_class_unloading() {
1890 _unloader.unload();
1891 if (mode()->is_generational()) {
1892 old_generation()->set_parsable(false);
1893 }
1894 }
1895
1896 void ShenandoahHeap::stw_weak_refs(bool full_gc) {
1897 // Weak refs processing
1898 ShenandoahPhaseTimings::Phase phase = full_gc ? ShenandoahPhaseTimings::full_gc_weakrefs
1899 : ShenandoahPhaseTimings::degen_gc_weakrefs;
1900 ShenandoahTimingsTracker t(phase);
1901 ShenandoahGCWorkerPhase worker_phase(phase);
1902 shenandoah_assert_generations_reconciled();
1903 gc_generation()->ref_processor()->process_references(phase, workers(), false /* concurrent */);
1904 }
1905
1906 void ShenandoahHeap::prepare_update_heap_references(bool concurrent) {
1907 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "must be at safepoint");
1908
1909 // Evacuation is over, no GCLABs are needed anymore. GCLABs are under URWM, so we need to
1910 // make them parsable for update code to work correctly. Plus, we can compute new sizes
1911 // for future GCLABs here.
1912 if (UseTLAB) {
1913 ShenandoahGCPhase phase(concurrent ?
1914 ShenandoahPhaseTimings::init_update_refs_manage_gclabs :
1915 ShenandoahPhaseTimings::degen_gc_init_update_refs_manage_gclabs);
1916 gclabs_retire(ResizeTLAB);
1917 }
1918
1919 _update_refs_iterator.reset();
1920 }
1921
1922 void ShenandoahHeap::propagate_gc_state_to_java_threads() {
1923 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Must be at Shenandoah safepoint");
1924 if (_gc_state_changed) {
1925 _gc_state_changed = false;
1926 char state = gc_state();
1927 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *t = jtiwh.next(); ) {
1928 ShenandoahThreadLocalData::set_gc_state(t, state);
1929 }
1930 }
1931 }
1932
1933 void ShenandoahHeap::set_gc_state(uint mask, bool value) {
1934 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Must be at Shenandoah safepoint");
1935 _gc_state.set_cond(mask, value);
1936 _gc_state_changed = true;
1937 }
1938
1939 void ShenandoahHeap::set_concurrent_young_mark_in_progress(bool in_progress) {
1940 uint mask;
1941 assert(!has_forwarded_objects(), "Young marking is not concurrent with evacuation");
1942 if (!in_progress && is_concurrent_old_mark_in_progress()) {
1943 assert(mode()->is_generational(), "Only generational GC has old marking");
1944 assert(_gc_state.is_set(MARKING), "concurrent_old_marking_in_progress implies MARKING");
1945 // If old-marking is in progress when we turn off YOUNG_MARKING, leave MARKING (and OLD_MARKING) on
1946 mask = YOUNG_MARKING;
1947 } else {
1948 mask = MARKING | YOUNG_MARKING;
1949 }
1950 set_gc_state(mask, in_progress);
1951 manage_satb_barrier(in_progress);
1952 }
1953
1954 void ShenandoahHeap::set_concurrent_old_mark_in_progress(bool in_progress) {
1955 #ifdef ASSERT
1956 // has_forwarded_objects() iff UPDATEREFS or EVACUATION
1957 bool has_forwarded = has_forwarded_objects();
1958 bool updating_or_evacuating = _gc_state.is_set(UPDATEREFS | EVACUATION);
1959 bool evacuating = _gc_state.is_set(EVACUATION);
1960 assert ((has_forwarded == updating_or_evacuating) || (evacuating && !has_forwarded && collection_set()->is_empty()),
1961 "Updating or evacuating iff has forwarded objects, or if evacuation phase is promoting in place without forwarding");
1962 #endif
1963 if (!in_progress && is_concurrent_young_mark_in_progress()) {
1964 // If young-marking is in progress when we turn off OLD_MARKING, leave MARKING (and YOUNG_MARKING) on
1965 assert(_gc_state.is_set(MARKING), "concurrent_young_marking_in_progress implies MARKING");
1966 set_gc_state(OLD_MARKING, in_progress);
1967 } else {
1968 set_gc_state(MARKING | OLD_MARKING, in_progress);
1969 }
1970 manage_satb_barrier(in_progress);
1971 }
1972
1973 bool ShenandoahHeap::is_prepare_for_old_mark_in_progress() const {
1974 return old_generation()->is_preparing_for_mark();
1975 }
1976
1977 void ShenandoahHeap::manage_satb_barrier(bool active) {
1978 if (is_concurrent_mark_in_progress()) {
1979 // Ignore request to deactivate barrier while concurrent mark is in progress.
1980 // Do not attempt to re-activate the barrier if it is already active.
1981 if (active && !ShenandoahBarrierSet::satb_mark_queue_set().is_active()) {
1982 ShenandoahBarrierSet::satb_mark_queue_set().set_active_all_threads(active, !active);
1983 }
1984 } else {
1985 // No concurrent marking is in progress so honor request to deactivate,
1986 // but only if the barrier is already active.
1987 if (!active && ShenandoahBarrierSet::satb_mark_queue_set().is_active()) {
1988 ShenandoahBarrierSet::satb_mark_queue_set().set_active_all_threads(active, !active);
1989 }
1990 }
1991 }
1992
1993 void ShenandoahHeap::set_evacuation_in_progress(bool in_progress) {
1994 assert(ShenandoahSafepoint::is_at_shenandoah_safepoint(), "Only call this at safepoint");
1995 set_gc_state(EVACUATION, in_progress);
1996 }
1997
1998 void ShenandoahHeap::set_concurrent_strong_root_in_progress(bool in_progress) {
1999 if (in_progress) {
2000 _concurrent_strong_root_in_progress.set();
2001 } else {
2002 _concurrent_strong_root_in_progress.unset();
2003 }
2004 }
2005
2006 void ShenandoahHeap::set_concurrent_weak_root_in_progress(bool cond) {
2007 set_gc_state(WEAK_ROOTS, cond);
2008 }
2009
2010 GCTracer* ShenandoahHeap::tracer() {
2011 return shenandoah_policy()->tracer();
2012 }
2013
2014 size_t ShenandoahHeap::tlab_used(Thread* thread) const {
2015 return _free_set->used();
2016 }
2017
2018 bool ShenandoahHeap::try_cancel_gc() {
2019 jbyte prev = _cancelled_gc.cmpxchg(CANCELLED, CANCELLABLE);
2020 return prev == CANCELLABLE;
2021 }
2022
2023 void ShenandoahHeap::cancel_concurrent_mark() {
2024 if (mode()->is_generational()) {
2025 young_generation()->cancel_marking();
2026 old_generation()->cancel_marking();
2027 }
2028
2029 global_generation()->cancel_marking();
2030
2031 ShenandoahBarrierSet::satb_mark_queue_set().abandon_partial_marking();
2032 }
2033
2034 void ShenandoahHeap::cancel_gc(GCCause::Cause cause) {
2035 if (try_cancel_gc()) {
2036 FormatBuffer<> msg("Cancelling GC: %s", GCCause::to_string(cause));
2037 log_info(gc)("%s", msg.buffer());
2038 Events::log(Thread::current(), "%s", msg.buffer());
2039 _cancel_requested_time = os::elapsedTime();
2040 }
2041 }
2042
2043 uint ShenandoahHeap::max_workers() {
2044 return _max_workers;
2045 }
2046
2047 void ShenandoahHeap::stop() {
2048 // The shutdown sequence should be able to terminate when GC is running.
2049
2050 // Step 0. Notify policy to disable event recording and prevent visiting gc threads during shutdown
2051 _shenandoah_policy->record_shutdown();
2052
2053 // Step 1. Notify control thread that we are in shutdown.
2054 // Note that we cannot do that with stop(), because stop() is blocking and waits for the actual shutdown.
2055 // Doing stop() here would wait for the normal GC cycle to complete, never falling through to cancel below.
2056 control_thread()->prepare_for_graceful_shutdown();
2057
2058 // Step 2. Notify GC workers that we are cancelling GC.
2059 cancel_gc(GCCause::_shenandoah_stop_vm);
2060
2061 // Step 3. Wait until GC worker exits normally.
2062 control_thread()->stop();
2063
2064 // Stop 4. Shutdown uncommit thread.
2065 if (_uncommit_thread != nullptr) {
2066 _uncommit_thread->stop();
2067 }
2068 }
2069
2070 void ShenandoahHeap::stw_unload_classes(bool full_gc) {
2071 if (!unload_classes()) return;
2072 ClassUnloadingContext ctx(_workers->active_workers(),
2073 true /* unregister_nmethods_during_purge */,
2074 false /* lock_nmethod_free_separately */);
2075
2076 // Unload classes and purge SystemDictionary.
2077 {
2078 ShenandoahPhaseTimings::Phase phase = full_gc ?
2079 ShenandoahPhaseTimings::full_gc_purge_class_unload :
2080 ShenandoahPhaseTimings::degen_gc_purge_class_unload;
2081 ShenandoahIsAliveSelector is_alive;
2082 {
2083 CodeCache::UnlinkingScope scope(is_alive.is_alive_closure());
2084 ShenandoahGCPhase gc_phase(phase);
2085 ShenandoahGCWorkerPhase worker_phase(phase);
2086 bool unloading_occurred = SystemDictionary::do_unloading(gc_timer());
2087
2088 uint num_workers = _workers->active_workers();
2089 ShenandoahClassUnloadingTask unlink_task(phase, num_workers, unloading_occurred);
2090 _workers->run_task(&unlink_task);
2091 }
2092 // Release unloaded nmethods's memory.
2093 ClassUnloadingContext::context()->purge_and_free_nmethods();
2094 }
2095
2096 {
2097 ShenandoahGCPhase phase(full_gc ?
2098 ShenandoahPhaseTimings::full_gc_purge_cldg :
2099 ShenandoahPhaseTimings::degen_gc_purge_cldg);
2100 ClassLoaderDataGraph::purge(true /* at_safepoint */);
2101 }
2102 // Resize and verify metaspace
2103 MetaspaceGC::compute_new_size();
2104 DEBUG_ONLY(MetaspaceUtils::verify();)
2105 }
2106
2107 // Weak roots are either pre-evacuated (final mark) or updated (final updaterefs),
2108 // so they should not have forwarded oops.
2109 // However, we do need to "null" dead oops in the roots, if can not be done
2110 // in concurrent cycles.
2111 void ShenandoahHeap::stw_process_weak_roots(bool full_gc) {
2112 uint num_workers = _workers->active_workers();
2113 ShenandoahPhaseTimings::Phase timing_phase = full_gc ?
2114 ShenandoahPhaseTimings::full_gc_purge_weak_par :
2115 ShenandoahPhaseTimings::degen_gc_purge_weak_par;
2116 ShenandoahGCPhase phase(timing_phase);
2117 ShenandoahGCWorkerPhase worker_phase(timing_phase);
2118 // Cleanup weak roots
2119 if (has_forwarded_objects()) {
2120 ShenandoahForwardedIsAliveClosure is_alive;
2121 ShenandoahNonConcUpdateRefsClosure keep_alive;
2122 ShenandoahParallelWeakRootsCleaningTask<ShenandoahForwardedIsAliveClosure, ShenandoahNonConcUpdateRefsClosure>
2123 cleaning_task(timing_phase, &is_alive, &keep_alive, num_workers);
2124 _workers->run_task(&cleaning_task);
2125 } else {
2126 ShenandoahIsAliveClosure is_alive;
2127 #ifdef ASSERT
2128 ShenandoahAssertNotForwardedClosure verify_cl;
2129 ShenandoahParallelWeakRootsCleaningTask<ShenandoahIsAliveClosure, ShenandoahAssertNotForwardedClosure>
2130 cleaning_task(timing_phase, &is_alive, &verify_cl, num_workers);
2131 #else
2132 ShenandoahParallelWeakRootsCleaningTask<ShenandoahIsAliveClosure, DoNothingClosure>
2133 cleaning_task(timing_phase, &is_alive, &do_nothing_cl, num_workers);
2134 #endif
2135 _workers->run_task(&cleaning_task);
2136 }
2137 }
2138
2139 void ShenandoahHeap::parallel_cleaning(bool full_gc) {
2140 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint");
2141 assert(is_stw_gc_in_progress(), "Only for Degenerated and Full GC");
2142 ShenandoahGCPhase phase(full_gc ?
2143 ShenandoahPhaseTimings::full_gc_purge :
2144 ShenandoahPhaseTimings::degen_gc_purge);
2145 stw_weak_refs(full_gc);
2146 stw_process_weak_roots(full_gc);
2147 stw_unload_classes(full_gc);
2148 }
2149
2150 void ShenandoahHeap::set_has_forwarded_objects(bool cond) {
2151 set_gc_state(HAS_FORWARDED, cond);
2152 }
2153
2154 void ShenandoahHeap::set_unload_classes(bool uc) {
2155 _unload_classes.set_cond(uc);
2156 }
2157
2158 bool ShenandoahHeap::unload_classes() const {
2159 return _unload_classes.is_set();
2160 }
2161
2162 address ShenandoahHeap::in_cset_fast_test_addr() {
2163 ShenandoahHeap* heap = ShenandoahHeap::heap();
2164 assert(heap->collection_set() != nullptr, "Sanity");
2165 return (address) heap->collection_set()->biased_map_address();
2166 }
2167
2168 void ShenandoahHeap::reset_bytes_allocated_since_gc_start() {
2169 if (mode()->is_generational()) {
2170 young_generation()->reset_bytes_allocated_since_gc_start();
2171 old_generation()->reset_bytes_allocated_since_gc_start();
2172 }
2173
2174 global_generation()->reset_bytes_allocated_since_gc_start();
2175 }
2176
2177 void ShenandoahHeap::set_degenerated_gc_in_progress(bool in_progress) {
2178 _degenerated_gc_in_progress.set_cond(in_progress);
2179 }
2180
2181 void ShenandoahHeap::set_full_gc_in_progress(bool in_progress) {
2182 _full_gc_in_progress.set_cond(in_progress);
2183 }
2184
2185 void ShenandoahHeap::set_full_gc_move_in_progress(bool in_progress) {
2186 assert (is_full_gc_in_progress(), "should be");
2187 _full_gc_move_in_progress.set_cond(in_progress);
2188 }
2189
2190 void ShenandoahHeap::set_update_refs_in_progress(bool in_progress) {
2191 set_gc_state(UPDATEREFS, in_progress);
2192 }
2193
2194 void ShenandoahHeap::register_nmethod(nmethod* nm) {
2195 ShenandoahCodeRoots::register_nmethod(nm);
2196 }
2197
2198 void ShenandoahHeap::unregister_nmethod(nmethod* nm) {
2199 ShenandoahCodeRoots::unregister_nmethod(nm);
2200 }
2201
2202 void ShenandoahHeap::pin_object(JavaThread* thr, oop o) {
2203 heap_region_containing(o)->record_pin();
2204 }
2205
2206 void ShenandoahHeap::unpin_object(JavaThread* thr, oop o) {
2207 ShenandoahHeapRegion* r = heap_region_containing(o);
2208 assert(r != nullptr, "Sanity");
2209 assert(r->pin_count() > 0, "Region " SIZE_FORMAT " should have non-zero pins", r->index());
2210 r->record_unpin();
2211 }
2212
2213 void ShenandoahHeap::sync_pinned_region_status() {
2214 ShenandoahHeapLocker locker(lock());
2215
2216 for (size_t i = 0; i < num_regions(); i++) {
2217 ShenandoahHeapRegion *r = get_region(i);
2218 if (r->is_active()) {
2219 if (r->is_pinned()) {
2220 if (r->pin_count() == 0) {
2221 r->make_unpinned();
2222 }
2223 } else {
2224 if (r->pin_count() > 0) {
2225 r->make_pinned();
2226 }
2227 }
2228 }
2229 }
2230
2231 assert_pinned_region_status();
2232 }
2233
2234 #ifdef ASSERT
2235 void ShenandoahHeap::assert_pinned_region_status() {
2236 for (size_t i = 0; i < num_regions(); i++) {
2237 ShenandoahHeapRegion* r = get_region(i);
2238 shenandoah_assert_generations_reconciled();
2239 if (gc_generation()->contains(r)) {
2240 assert((r->is_pinned() && r->pin_count() > 0) || (!r->is_pinned() && r->pin_count() == 0),
2241 "Region " SIZE_FORMAT " pinning status is inconsistent", i);
2242 }
2243 }
2244 }
2245 #endif
2246
2247 ConcurrentGCTimer* ShenandoahHeap::gc_timer() const {
2248 return _gc_timer;
2249 }
2250
2251 void ShenandoahHeap::prepare_concurrent_roots() {
2252 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint");
2253 assert(!is_stw_gc_in_progress(), "Only concurrent GC");
2254 set_concurrent_strong_root_in_progress(!collection_set()->is_empty());
2255 set_concurrent_weak_root_in_progress(true);
2256 if (unload_classes()) {
2257 _unloader.prepare();
2258 }
2259 }
2260
2261 void ShenandoahHeap::finish_concurrent_roots() {
2262 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint");
2263 assert(!is_stw_gc_in_progress(), "Only concurrent GC");
2264 if (unload_classes()) {
2265 _unloader.finish();
2266 }
2267 }
2268
2269 #ifdef ASSERT
2270 void ShenandoahHeap::assert_gc_workers(uint nworkers) {
2271 assert(nworkers > 0 && nworkers <= max_workers(), "Sanity");
2272
2273 if (ShenandoahSafepoint::is_at_shenandoah_safepoint()) {
2274 // Use ParallelGCThreads inside safepoints
2275 assert(nworkers == ParallelGCThreads, "Use ParallelGCThreads (%u) within safepoint, not %u",
2276 ParallelGCThreads, nworkers);
2277 } else {
2278 // Use ConcGCThreads outside safepoints
2279 assert(nworkers == ConcGCThreads, "Use ConcGCThreads (%u) outside safepoints, %u",
2280 ConcGCThreads, nworkers);
2281 }
2282 }
2283 #endif
2284
2285 ShenandoahVerifier* ShenandoahHeap::verifier() {
2286 guarantee(ShenandoahVerify, "Should be enabled");
2287 assert (_verifier != nullptr, "sanity");
2288 return _verifier;
2289 }
2290
2291 template<bool CONCURRENT>
2292 class ShenandoahUpdateHeapRefsTask : public WorkerTask {
2293 private:
2294 ShenandoahHeap* _heap;
2295 ShenandoahRegionIterator* _regions;
2296 public:
2297 explicit ShenandoahUpdateHeapRefsTask(ShenandoahRegionIterator* regions) :
2298 WorkerTask("Shenandoah Update References"),
2299 _heap(ShenandoahHeap::heap()),
2300 _regions(regions) {
2301 }
2302
2303 void work(uint worker_id) {
2304 if (CONCURRENT) {
2305 ShenandoahConcurrentWorkerSession worker_session(worker_id);
2306 ShenandoahSuspendibleThreadSetJoiner stsj;
2307 do_work<ShenandoahConcUpdateRefsClosure>(worker_id);
2308 } else {
2309 ShenandoahParallelWorkerSession worker_session(worker_id);
2310 do_work<ShenandoahNonConcUpdateRefsClosure>(worker_id);
2311 }
2312 }
2313
2314 private:
2315 template<class T>
2316 void do_work(uint worker_id) {
2317 if (CONCURRENT && (worker_id == 0)) {
2318 // We ask the first worker to replenish the Mutator free set by moving regions previously reserved to hold the
2319 // results of evacuation. These reserves are no longer necessary because evacuation has completed.
2320 size_t cset_regions = _heap->collection_set()->count();
2321
2322 // Now that evacuation is done, we can reassign any regions that had been reserved to hold the results of evacuation
2323 // to the mutator free set. At the end of GC, we will have cset_regions newly evacuated fully empty regions from
2324 // which we will be able to replenish the Collector free set and the OldCollector free set in preparation for the
2325 // next GC cycle.
2326 _heap->free_set()->move_regions_from_collector_to_mutator(cset_regions);
2327 }
2328 // If !CONCURRENT, there's no value in expanding Mutator free set
2329 T cl;
2330 ShenandoahHeapRegion* r = _regions->next();
2331 while (r != nullptr) {
2332 HeapWord* update_watermark = r->get_update_watermark();
2333 assert (update_watermark >= r->bottom(), "sanity");
2334 if (r->is_active() && !r->is_cset()) {
2335 _heap->marked_object_oop_iterate(r, &cl, update_watermark);
2336 if (ShenandoahPacing) {
2337 _heap->pacer()->report_updaterefs(pointer_delta(update_watermark, r->bottom()));
2338 }
2339 }
2340 if (_heap->check_cancelled_gc_and_yield(CONCURRENT)) {
2341 return;
2342 }
2343 r = _regions->next();
2344 }
2345 }
2346 };
2347
2348 void ShenandoahHeap::update_heap_references(bool concurrent) {
2349 assert(!is_full_gc_in_progress(), "Only for concurrent and degenerated GC");
2350
2351 if (concurrent) {
2352 ShenandoahUpdateHeapRefsTask<true> task(&_update_refs_iterator);
2353 workers()->run_task(&task);
2354 } else {
2355 ShenandoahUpdateHeapRefsTask<false> task(&_update_refs_iterator);
2356 workers()->run_task(&task);
2357 }
2358 }
2359
2360 void ShenandoahHeap::update_heap_region_states(bool concurrent) {
2361 assert(SafepointSynchronize::is_at_safepoint(), "Must be at a safepoint");
2362 assert(!is_full_gc_in_progress(), "Only for concurrent and degenerated GC");
2363
2364 {
2365 ShenandoahGCPhase phase(concurrent ?
2366 ShenandoahPhaseTimings::final_update_refs_update_region_states :
2367 ShenandoahPhaseTimings::degen_gc_final_update_refs_update_region_states);
2368
2369 final_update_refs_update_region_states();
2370
2371 assert_pinned_region_status();
2372 }
2373
2374 {
2375 ShenandoahGCPhase phase(concurrent ?
2376 ShenandoahPhaseTimings::final_update_refs_trash_cset :
2377 ShenandoahPhaseTimings::degen_gc_final_update_refs_trash_cset);
2378 trash_cset_regions();
2379 }
2380 }
2381
2382 void ShenandoahHeap::final_update_refs_update_region_states() {
2383 ShenandoahSynchronizePinnedRegionStates cl;
2384 parallel_heap_region_iterate(&cl);
2385 }
2386
2387 void ShenandoahHeap::rebuild_free_set(bool concurrent) {
2388 ShenandoahGCPhase phase(concurrent ?
2389 ShenandoahPhaseTimings::final_update_refs_rebuild_freeset :
2390 ShenandoahPhaseTimings::degen_gc_final_update_refs_rebuild_freeset);
2391 ShenandoahHeapLocker locker(lock());
2392 size_t young_cset_regions, old_cset_regions;
2393 size_t first_old_region, last_old_region, old_region_count;
2394 _free_set->prepare_to_rebuild(young_cset_regions, old_cset_regions, first_old_region, last_old_region, old_region_count);
2395 // If there are no old regions, first_old_region will be greater than last_old_region
2396 assert((first_old_region > last_old_region) ||
2397 ((last_old_region + 1 - first_old_region >= old_region_count) &&
2398 get_region(first_old_region)->is_old() && get_region(last_old_region)->is_old()),
2399 "sanity: old_region_count: " SIZE_FORMAT ", first_old_region: " SIZE_FORMAT ", last_old_region: " SIZE_FORMAT,
2400 old_region_count, first_old_region, last_old_region);
2401
2402 if (mode()->is_generational()) {
2403 #ifdef ASSERT
2404 if (ShenandoahVerify) {
2405 verifier()->verify_before_rebuilding_free_set();
2406 }
2407 #endif
2408
2409 // The computation of bytes_of_allocation_runway_before_gc_trigger is quite conservative so consider all of this
2410 // available for transfer to old. Note that transfer of humongous regions does not impact available.
2411 ShenandoahGenerationalHeap* gen_heap = ShenandoahGenerationalHeap::heap();
2412 size_t allocation_runway = gen_heap->young_generation()->heuristics()->bytes_of_allocation_runway_before_gc_trigger(young_cset_regions);
2413 gen_heap->compute_old_generation_balance(allocation_runway, old_cset_regions);
2414
2415 // Total old_available may have been expanded to hold anticipated promotions. We trigger if the fragmented available
2416 // memory represents more than 16 regions worth of data. Note that fragmentation may increase when we promote regular
2417 // regions in place when many of these regular regions have an abundant amount of available memory within them. Fragmentation
2418 // will decrease as promote-by-copy consumes the available memory within these partially consumed regions.
2419 //
2420 // We consider old-gen to have excessive fragmentation if more than 12.5% of old-gen is free memory that resides
2421 // within partially consumed regions of memory.
2422 }
2423 // Rebuild free set based on adjusted generation sizes.
2424 _free_set->finish_rebuild(young_cset_regions, old_cset_regions, old_region_count);
2425
2426 if (mode()->is_generational()) {
2427 ShenandoahGenerationalHeap* gen_heap = ShenandoahGenerationalHeap::heap();
2428 ShenandoahOldGeneration* old_gen = gen_heap->old_generation();
2429 old_gen->heuristics()->evaluate_triggers(first_old_region, last_old_region, old_region_count, num_regions());
2430 }
2431 }
2432
2433 void ShenandoahHeap::print_extended_on(outputStream *st) const {
2434 print_on(st);
2435 st->cr();
2436 print_heap_regions_on(st);
2437 }
2438
2439 bool ShenandoahHeap::is_bitmap_slice_committed(ShenandoahHeapRegion* r, bool skip_self) {
2440 size_t slice = r->index() / _bitmap_regions_per_slice;
2441
2442 size_t regions_from = _bitmap_regions_per_slice * slice;
2443 size_t regions_to = MIN2(num_regions(), _bitmap_regions_per_slice * (slice + 1));
2444 for (size_t g = regions_from; g < regions_to; g++) {
2445 assert (g / _bitmap_regions_per_slice == slice, "same slice");
2446 if (skip_self && g == r->index()) continue;
2447 if (get_region(g)->is_committed()) {
2448 return true;
2449 }
2450 }
2451 return false;
2452 }
2453
2454 bool ShenandoahHeap::commit_bitmap_slice(ShenandoahHeapRegion* r) {
2455 shenandoah_assert_heaplocked();
2456
2457 // Bitmaps in special regions do not need commits
2458 if (_bitmap_region_special) {
2459 return true;
2460 }
2461
2462 if (is_bitmap_slice_committed(r, true)) {
2463 // Some other region from the group is already committed, meaning the bitmap
2464 // slice is already committed, we exit right away.
2465 return true;
2466 }
2467
2468 // Commit the bitmap slice:
2469 size_t slice = r->index() / _bitmap_regions_per_slice;
2470 size_t off = _bitmap_bytes_per_slice * slice;
2471 size_t len = _bitmap_bytes_per_slice;
2472 char* start = (char*) _bitmap_region.start() + off;
2473
2474 if (!os::commit_memory(start, len, false)) {
2475 return false;
2476 }
2477
2478 if (AlwaysPreTouch) {
2479 os::pretouch_memory(start, start + len, _pretouch_bitmap_page_size);
2480 }
2481
2482 return true;
2483 }
2484
2485 bool ShenandoahHeap::uncommit_bitmap_slice(ShenandoahHeapRegion *r) {
2486 shenandoah_assert_heaplocked();
2487
2488 // Bitmaps in special regions do not need uncommits
2489 if (_bitmap_region_special) {
2490 return true;
2491 }
2492
2493 if (is_bitmap_slice_committed(r, true)) {
2494 // Some other region from the group is still committed, meaning the bitmap
2495 // slice should stay committed, exit right away.
2496 return true;
2497 }
2498
2499 // Uncommit the bitmap slice:
2500 size_t slice = r->index() / _bitmap_regions_per_slice;
2501 size_t off = _bitmap_bytes_per_slice * slice;
2502 size_t len = _bitmap_bytes_per_slice;
2503 if (!os::uncommit_memory((char*)_bitmap_region.start() + off, len)) {
2504 return false;
2505 }
2506 return true;
2507 }
2508
2509 void ShenandoahHeap::forbid_uncommit() {
2510 if (_uncommit_thread != nullptr) {
2511 _uncommit_thread->forbid_uncommit();
2512 }
2513 }
2514
2515 void ShenandoahHeap::allow_uncommit() {
2516 if (_uncommit_thread != nullptr) {
2517 _uncommit_thread->allow_uncommit();
2518 }
2519 }
2520
2521 #ifdef ASSERT
2522 bool ShenandoahHeap::is_uncommit_in_progress() {
2523 if (_uncommit_thread != nullptr) {
2524 return _uncommit_thread->is_uncommit_in_progress();
2525 }
2526 return false;
2527 }
2528 #endif
2529
2530 void ShenandoahHeap::safepoint_synchronize_begin() {
2531 StackWatermarkSet::safepoint_synchronize_begin();
2532 SuspendibleThreadSet::synchronize();
2533 }
2534
2535 void ShenandoahHeap::safepoint_synchronize_end() {
2536 SuspendibleThreadSet::desynchronize();
2537 }
2538
2539 void ShenandoahHeap::try_inject_alloc_failure() {
2540 if (ShenandoahAllocFailureALot && !cancelled_gc() && ((os::random() % 1000) > 950)) {
2541 _inject_alloc_failure.set();
2542 os::naked_short_sleep(1);
2543 if (cancelled_gc()) {
2544 log_info(gc)("Allocation failure was successfully injected");
2545 }
2546 }
2547 }
2548
2549 bool ShenandoahHeap::should_inject_alloc_failure() {
2550 return _inject_alloc_failure.is_set() && _inject_alloc_failure.try_unset();
2551 }
2552
2553 void ShenandoahHeap::initialize_serviceability() {
2554 _memory_pool = new ShenandoahMemoryPool(this);
2555 _cycle_memory_manager.add_pool(_memory_pool);
2556 _stw_memory_manager.add_pool(_memory_pool);
2557 }
2558
2559 GrowableArray<GCMemoryManager*> ShenandoahHeap::memory_managers() {
2560 GrowableArray<GCMemoryManager*> memory_managers(2);
2561 memory_managers.append(&_cycle_memory_manager);
2562 memory_managers.append(&_stw_memory_manager);
2563 return memory_managers;
2564 }
2565
2566 GrowableArray<MemoryPool*> ShenandoahHeap::memory_pools() {
2567 GrowableArray<MemoryPool*> memory_pools(1);
2568 memory_pools.append(_memory_pool);
2569 return memory_pools;
2570 }
2571
2572 MemoryUsage ShenandoahHeap::memory_usage() {
2573 return MemoryUsage(_initial_size, used(), committed(), max_capacity());
2574 }
2575
2576 ShenandoahRegionIterator::ShenandoahRegionIterator() :
2577 _heap(ShenandoahHeap::heap()),
2578 _index(0) {}
2579
2580 ShenandoahRegionIterator::ShenandoahRegionIterator(ShenandoahHeap* heap) :
2581 _heap(heap),
2582 _index(0) {}
2583
2584 void ShenandoahRegionIterator::reset() {
2585 _index = 0;
2586 }
2587
2588 bool ShenandoahRegionIterator::has_next() const {
2589 return _index < _heap->num_regions();
2590 }
2591
2592 char ShenandoahHeap::gc_state() const {
2593 return _gc_state.raw_value();
2594 }
2595
2596 ShenandoahLiveData* ShenandoahHeap::get_liveness_cache(uint worker_id) {
2597 #ifdef ASSERT
2598 assert(_liveness_cache != nullptr, "sanity");
2599 assert(worker_id < _max_workers, "sanity");
2600 for (uint i = 0; i < num_regions(); i++) {
2601 assert(_liveness_cache[worker_id][i] == 0, "liveness cache should be empty");
2602 }
2603 #endif
2604 return _liveness_cache[worker_id];
2605 }
2606
2607 void ShenandoahHeap::flush_liveness_cache(uint worker_id) {
2608 assert(worker_id < _max_workers, "sanity");
2609 assert(_liveness_cache != nullptr, "sanity");
2610 ShenandoahLiveData* ld = _liveness_cache[worker_id];
2611 for (uint i = 0; i < num_regions(); i++) {
2612 ShenandoahLiveData live = ld[i];
2613 if (live > 0) {
2614 ShenandoahHeapRegion* r = get_region(i);
2615 r->increase_live_data_gc_words(live);
2616 ld[i] = 0;
2617 }
2618 }
2619 }
2620
2621 bool ShenandoahHeap::requires_barriers(stackChunkOop obj) const {
2622 if (is_idle()) return false;
2623
2624 // Objects allocated after marking start are implicitly alive, don't need any barriers during
2625 // marking phase.
2626 if (is_concurrent_mark_in_progress() &&
2627 !marking_context()->allocated_after_mark_start(obj)) {
2628 return true;
2629 }
2630
2631 // Can not guarantee obj is deeply good.
2632 if (has_forwarded_objects()) {
2633 return true;
2634 }
2635
2636 return false;
2637 }
2638
2639 HeapWord* ShenandoahHeap::allocate_loaded_archive_space(size_t size) {
2640 #if INCLUDE_CDS_JAVA_HEAP
2641 // CDS wants a continuous memory range to load a bunch of objects.
2642 // This effectively bypasses normal allocation paths, and requires
2643 // a bit of massaging to unbreak GC invariants.
2644
2645 ShenandoahAllocRequest req = ShenandoahAllocRequest::for_shared(size);
2646
2647 // Easy case: a single regular region, no further adjustments needed.
2648 if (!ShenandoahHeapRegion::requires_humongous(size)) {
2649 return allocate_memory(req);
2650 }
2651
2652 // Hard case: the requested size would cause a humongous allocation.
2653 // We need to make sure it looks like regular allocation to the rest of GC.
2654
2655 // CDS code would guarantee no objects straddle multiple regions, as long as
2656 // regions are as large as MIN_GC_REGION_ALIGNMENT. It is impractical at this
2657 // point to deal with case when Shenandoah runs with smaller regions.
2658 // TODO: This check can be dropped once MIN_GC_REGION_ALIGNMENT agrees more with Shenandoah.
2659 if (ShenandoahHeapRegion::region_size_bytes() < ArchiveHeapWriter::MIN_GC_REGION_ALIGNMENT) {
2660 return nullptr;
2661 }
2662
2663 HeapWord* mem = allocate_memory(req);
2664 size_t start_idx = heap_region_index_containing(mem);
2665 size_t num_regions = ShenandoahHeapRegion::required_regions(size * HeapWordSize);
2666
2667 // Flip humongous -> regular.
2668 {
2669 ShenandoahHeapLocker locker(lock(), false);
2670 for (size_t c = start_idx; c < start_idx + num_regions; c++) {
2671 get_region(c)->make_regular_bypass();
2672 }
2673 }
2674
2675 return mem;
2676 #else
2677 assert(false, "Archive heap loader should not be available, should not be here");
2678 return nullptr;
2679 #endif // INCLUDE_CDS_JAVA_HEAP
2680 }
2681
2682 void ShenandoahHeap::complete_loaded_archive_space(MemRegion archive_space) {
2683 // Nothing to do here, except checking that heap looks fine.
2684 #ifdef ASSERT
2685 HeapWord* start = archive_space.start();
2686 HeapWord* end = archive_space.end();
2687
2688 // No unclaimed space between the objects.
2689 // Objects are properly allocated in correct regions.
2690 HeapWord* cur = start;
2691 while (cur < end) {
2692 oop oop = cast_to_oop(cur);
2693 shenandoah_assert_in_correct_region(nullptr, oop);
2694 cur += oop->size();
2695 }
2696
2697 // No unclaimed tail at the end of archive space.
2698 assert(cur == end,
2699 "Archive space should be fully used: " PTR_FORMAT " " PTR_FORMAT,
2700 p2i(cur), p2i(end));
2701
2702 // Region bounds are good.
2703 ShenandoahHeapRegion* begin_reg = heap_region_containing(start);
2704 ShenandoahHeapRegion* end_reg = heap_region_containing(end);
2705 assert(begin_reg->is_regular(), "Must be");
2706 assert(end_reg->is_regular(), "Must be");
2707 assert(begin_reg->bottom() == start,
2708 "Must agree: archive-space-start: " PTR_FORMAT ", begin-region-bottom: " PTR_FORMAT,
2709 p2i(start), p2i(begin_reg->bottom()));
2710 assert(end_reg->top() == end,
2711 "Must agree: archive-space-end: " PTR_FORMAT ", end-region-top: " PTR_FORMAT,
2712 p2i(end), p2i(end_reg->top()));
2713 #endif
2714 }
2715
2716 ShenandoahGeneration* ShenandoahHeap::generation_for(ShenandoahAffiliation affiliation) const {
2717 if (!mode()->is_generational()) {
2718 return global_generation();
2719 } else if (affiliation == YOUNG_GENERATION) {
2720 return young_generation();
2721 } else if (affiliation == OLD_GENERATION) {
2722 return old_generation();
2723 }
2724
2725 ShouldNotReachHere();
2726 return nullptr;
2727 }
2728
2729 void ShenandoahHeap::log_heap_status(const char* msg) const {
2730 if (mode()->is_generational()) {
2731 young_generation()->log_status(msg);
2732 old_generation()->log_status(msg);
2733 } else {
2734 global_generation()->log_status(msg);
2735 }
2736 }
2737