1 /*
2 * Copyright (c) 2023, 2026, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "cds/aotMappedHeap.hpp"
26 #include "cds/aotMappedHeapWriter.hpp"
27 #include "cds/aotReferenceObjSupport.hpp"
28 #include "cds/cdsConfig.hpp"
29 #include "cds/filemap.hpp"
30 #include "cds/heapShared.inline.hpp"
31 #include "cds/regeneratedClasses.hpp"
32 #include "classfile/javaClasses.hpp"
33 #include "classfile/modules.hpp"
34 #include "classfile/systemDictionary.hpp"
35 #include "gc/shared/collectedHeap.hpp"
36 #include "memory/allocation.inline.hpp"
37 #include "memory/iterator.inline.hpp"
38 #include "memory/oopFactory.hpp"
39 #include "memory/universe.hpp"
40 #include "oops/compressedOops.hpp"
41 #include "oops/objArrayOop.inline.hpp"
42 #include "oops/oop.inline.hpp"
43 #include "oops/oopHandle.inline.hpp"
44 #include "oops/typeArrayKlass.hpp"
45 #include "oops/typeArrayOop.hpp"
46 #include "runtime/java.hpp"
47 #include "runtime/mutexLocker.hpp"
48 #include "utilities/bitMap.inline.hpp"
49 #if INCLUDE_G1GC
50 #include "gc/g1/g1CollectedHeap.hpp"
51 #include "gc/g1/g1HeapRegion.hpp"
52 #endif
53
54 #if INCLUDE_CDS_JAVA_HEAP
55
56 GrowableArrayCHeap<u1, mtClassShared>* AOTMappedHeapWriter::_buffer = nullptr;
57
58 bool AOTMappedHeapWriter::_is_writing_deterministic_heap = false;
59 size_t AOTMappedHeapWriter::_buffer_used;
60
61 // Heap root segments
62 HeapRootSegments AOTMappedHeapWriter::_heap_root_segments;
63
64 address AOTMappedHeapWriter::_requested_bottom;
65 address AOTMappedHeapWriter::_requested_top;
66
67 static size_t _num_strings = 0;
68 static size_t _string_bytes = 0;
69 static size_t _num_packages = 0;
70 static size_t _num_protection_domains = 0;
71
72 GrowableArrayCHeap<AOTMappedHeapWriter::NativePointerInfo, mtClassShared>* AOTMappedHeapWriter::_native_pointers;
73 GrowableArrayCHeap<oop, mtClassShared>* AOTMappedHeapWriter::_source_objs;
74 GrowableArrayCHeap<AOTMappedHeapWriter::HeapObjOrder, mtClassShared>* AOTMappedHeapWriter::_source_objs_order;
75
76 AOTMappedHeapWriter::BufferOffsetToSourceObjectTable*
77 AOTMappedHeapWriter::_buffer_offset_to_source_obj_table = nullptr;
78
79 typedef HashTable<
80 size_t, // offset of a filler from AOTMappedHeapWriter::buffer_bottom()
81 size_t, // size of this filler (in bytes)
82 127, // prime number
83 AnyObj::C_HEAP,
84 mtClassShared> FillersTable;
85 static FillersTable* _fillers;
86 static int _num_native_ptrs = 0;
87
88 void AOTMappedHeapWriter::init() {
89 if (CDSConfig::is_dumping_heap()) {
90 Universe::heap()->collect(GCCause::_java_lang_system_gc);
91
92 _buffer_offset_to_source_obj_table = new (mtClassShared) BufferOffsetToSourceObjectTable(/*size (prime)*/36137, /*max size*/1 * M);
93 _fillers = new (mtClassShared) FillersTable();
94 _requested_bottom = nullptr;
95 _requested_top = nullptr;
96
97 _native_pointers = new GrowableArrayCHeap<NativePointerInfo, mtClassShared>(2048);
98 _source_objs = new GrowableArrayCHeap<oop, mtClassShared>(10000);
99
100 guarantee(MIN_GC_REGION_ALIGNMENT <= G1HeapRegion::min_region_size_in_words() * HeapWordSize, "must be");
101
102 if (CDSConfig::old_cds_flags_used()) {
103 // With the old CDS workflow, we can guatantee determninistic output: given
104 // the same classlist file, we can generate the same static CDS archive.
105 // To ensure determinism, we always use the same compressed oop encoding
106 // (zero-based, no shift). See set_requested_address_range().
107 _is_writing_deterministic_heap = true;
108 } else {
109 // Determninistic output is not supported by the new AOT workflow, so
110 // we don't force the (zero-based, no shift) encoding. This way, it is more
111 // likely that we can avoid oop relocation in the production run.
112 _is_writing_deterministic_heap = false;
113 }
114 }
115 }
116
117 // For AOTMappedHeapWriter::narrow_oop_{mode, base, shift}(), see comments
118 // in AOTMappedHeapWriter::set_requested_address_range(),
119 CompressedOops::Mode AOTMappedHeapWriter::narrow_oop_mode() {
120 if (is_writing_deterministic_heap()) {
121 return CompressedOops::UnscaledNarrowOop;
122 } else {
123 return CompressedOops::mode();
124 }
125 }
126
127 address AOTMappedHeapWriter::narrow_oop_base() {
128 if (is_writing_deterministic_heap()) {
129 return nullptr;
130 } else {
131 return CompressedOops::base();
132 }
133 }
134
135 int AOTMappedHeapWriter::narrow_oop_shift() {
136 if (is_writing_deterministic_heap()) {
137 return 0;
138 } else {
139 return CompressedOops::shift();
140 }
141 }
142
143 void AOTMappedHeapWriter::delete_tables_with_raw_oops() {
144 delete _source_objs;
145 _source_objs = nullptr;
146 }
147
148 void AOTMappedHeapWriter::add_source_obj(oop src_obj) {
149 _source_objs->append(src_obj);
150 }
151
152 void AOTMappedHeapWriter::write(GrowableArrayCHeap<oop, mtClassShared>* roots,
153 AOTMappedHeapInfo* heap_info) {
154 assert(CDSConfig::is_dumping_heap(), "sanity");
155 allocate_buffer();
156 copy_source_objs_to_buffer(roots);
157 set_requested_address_range(heap_info);
158 relocate_embedded_oops(roots, heap_info);
159 }
160
161 bool AOTMappedHeapWriter::is_too_large_to_archive(oop o) {
162 return is_too_large_to_archive(o->size());
163 }
164
165 bool AOTMappedHeapWriter::is_string_too_large_to_archive(oop string) {
166 typeArrayOop value = java_lang_String::value_no_keepalive(string);
167 return is_too_large_to_archive(value);
168 }
169
170 bool AOTMappedHeapWriter::is_too_large_to_archive(size_t size) {
171 assert(size > 0, "no zero-size object");
172 assert(size * HeapWordSize > size, "no overflow");
173 static_assert(MIN_GC_REGION_ALIGNMENT > 0, "must be positive");
174
175 size_t byte_size = size * HeapWordSize;
176 if (byte_size > size_t(MIN_GC_REGION_ALIGNMENT)) {
177 return true;
178 } else {
179 return false;
180 }
181 }
182
183 // Various lookup functions between source_obj, buffered_obj and requested_obj
184 bool AOTMappedHeapWriter::is_in_requested_range(oop o) {
185 assert(_requested_bottom != nullptr, "do not call before _requested_bottom is initialized");
186 address a = cast_from_oop<address>(o);
187 return (_requested_bottom <= a && a < _requested_top);
188 }
189
190 oop AOTMappedHeapWriter::requested_obj_from_buffer_offset(size_t offset) {
191 oop req_obj = cast_to_oop(_requested_bottom + offset);
192 assert(is_in_requested_range(req_obj), "must be");
193 return req_obj;
194 }
195
196 oop AOTMappedHeapWriter::source_obj_to_requested_obj(oop src_obj) {
197 assert(CDSConfig::is_dumping_heap(), "dump-time only");
198 HeapShared::CachedOopInfo* p = HeapShared::get_cached_oop_info(src_obj);
199 if (p != nullptr) {
200 return requested_obj_from_buffer_offset(p->buffer_offset());
201 } else {
202 return nullptr;
203 }
204 }
205
206 oop AOTMappedHeapWriter::buffered_addr_to_source_obj(address buffered_addr) {
207 OopHandle* oh = _buffer_offset_to_source_obj_table->get(buffered_address_to_offset(buffered_addr));
208 if (oh != nullptr) {
209 return oh->resolve();
210 } else {
211 return nullptr;
212 }
213 }
214
215 Klass* AOTMappedHeapWriter::real_klass_of_buffered_oop(address buffered_addr) {
216 oop p = buffered_addr_to_source_obj(buffered_addr);
217 if (p != nullptr) {
218 return p->klass();
219 } else if (get_filler_size_at(buffered_addr) > 0) {
220 return Universe::fillerArrayKlass();
221 } else {
222 // This is one of the root segments
223 return Universe::objectArrayKlass();
224 }
225 }
226
227 size_t AOTMappedHeapWriter::size_of_buffered_oop(address buffered_addr) {
228 oop p = buffered_addr_to_source_obj(buffered_addr);
229 if (p != nullptr) {
230 return p->size();
231 }
232
233 size_t nbytes = get_filler_size_at(buffered_addr);
234 if (nbytes > 0) {
235 assert((nbytes % BytesPerWord) == 0, "should be aligned");
236 return nbytes / BytesPerWord;
237 }
238
239 address hrs = buffer_bottom();
240 for (size_t seg_idx = 0; seg_idx < _heap_root_segments.count(); seg_idx++) {
241 nbytes = _heap_root_segments.size_in_bytes(seg_idx);
242 if (hrs == buffered_addr) {
243 assert((nbytes % BytesPerWord) == 0, "should be aligned");
244 return nbytes / BytesPerWord;
245 }
246 hrs += nbytes;
247 }
248
249 ShouldNotReachHere();
250 return 0;
251 }
252
253 address AOTMappedHeapWriter::buffered_addr_to_requested_addr(address buffered_addr) {
254 return _requested_bottom + buffered_address_to_offset(buffered_addr);
255 }
256
257 address AOTMappedHeapWriter::requested_address() {
258 assert(_buffer != nullptr, "must be initialized");
259 return _requested_bottom;
260 }
261
262 void AOTMappedHeapWriter::allocate_buffer() {
263 int initial_buffer_size = 100000;
264 _buffer = new GrowableArrayCHeap<u1, mtClassShared>(initial_buffer_size);
265 _buffer_used = 0;
266 ensure_buffer_space(1); // so that buffer_bottom() works
267 }
268
269 void AOTMappedHeapWriter::ensure_buffer_space(size_t min_bytes) {
270 // We usually have very small heaps. If we get a huge one it's probably caused by a bug.
271 guarantee(min_bytes <= max_jint, "we dont support archiving more than 2G of objects");
272 _buffer->at_grow(to_array_index(min_bytes));
273 }
274
275 objArrayOop AOTMappedHeapWriter::allocate_root_segment(size_t offset, int element_count) {
276 HeapWord* mem = offset_to_buffered_address<HeapWord *>(offset);
277 memset(mem, 0, objArrayOopDesc::object_size(element_count));
278
279 // The initialization code is copied from MemAllocator::finish and ObjArrayAllocator::initialize.
280 if (UseCompactObjectHeaders) {
281 oopDesc::release_set_mark(mem, Universe::objectArrayKlass()->prototype_header());
282 } else {
283 oopDesc::set_mark(mem, markWord::prototype());
284 oopDesc::release_set_klass(mem, Universe::objectArrayKlass());
285 }
286 arrayOopDesc::set_length(mem, element_count);
287 return objArrayOop(cast_to_oop(mem));
288 }
289
290 void AOTMappedHeapWriter::root_segment_at_put(objArrayOop segment, int index, oop root) {
291 // Do not use arrayOop->obj_at_put(i, o) as arrayOop is outside the real heap!
292 if (UseCompressedOops) {
293 *segment->obj_at_addr<narrowOop>(index) = CompressedOops::encode(root);
294 } else {
295 *segment->obj_at_addr<oop>(index) = root;
296 }
297 }
298
299 void AOTMappedHeapWriter::copy_roots_to_buffer(GrowableArrayCHeap<oop, mtClassShared>* roots) {
300 // Depending on the number of classes we are archiving, a single roots array may be
301 // larger than MIN_GC_REGION_ALIGNMENT. Roots are allocated first in the buffer, which
302 // allows us to chop the large array into a series of "segments". Current layout
303 // starts with zero or more segments exactly fitting MIN_GC_REGION_ALIGNMENT, and end
304 // with a single segment that may be smaller than MIN_GC_REGION_ALIGNMENT.
305 // This is simple and efficient. We do not need filler objects anywhere between the segments,
306 // or immediately after the last segment. This allows starting the object dump immediately
307 // after the roots.
308
309 assert((_buffer_used % MIN_GC_REGION_ALIGNMENT) == 0,
310 "Pre-condition: Roots start at aligned boundary: %zu", _buffer_used);
311
312 int max_elem_count = ((MIN_GC_REGION_ALIGNMENT - arrayOopDesc::header_size_in_bytes()) / heapOopSize);
313 assert(objArrayOopDesc::object_size(max_elem_count)*HeapWordSize == MIN_GC_REGION_ALIGNMENT,
314 "Should match exactly");
315
316 HeapRootSegments segments(_buffer_used,
317 roots->length(),
318 MIN_GC_REGION_ALIGNMENT,
319 max_elem_count);
320
321 int root_index = 0;
322 for (size_t seg_idx = 0; seg_idx < segments.count(); seg_idx++) {
323 int size_elems = segments.size_in_elems(seg_idx);
324 size_t size_bytes = segments.size_in_bytes(seg_idx);
325
326 size_t oop_offset = _buffer_used;
327 _buffer_used = oop_offset + size_bytes;
328 ensure_buffer_space(_buffer_used);
329
330 assert((oop_offset % MIN_GC_REGION_ALIGNMENT) == 0,
331 "Roots segment %zu start is not aligned: %zu",
332 segments.count(), oop_offset);
333
334 objArrayOop seg_oop = allocate_root_segment(oop_offset, size_elems);
335 for (int i = 0; i < size_elems; i++) {
336 root_segment_at_put(seg_oop, i, roots->at(root_index++));
337 }
338
339 log_info(aot, heap)("archived obj root segment [%d] = %zu bytes, obj = " PTR_FORMAT,
340 size_elems, size_bytes, p2i(seg_oop));
341 }
342
343 assert(root_index == roots->length(), "Post-condition: All roots are handled");
344
345 _heap_root_segments = segments;
346 }
347
348 // The goal is to sort the objects in increasing order of:
349 // - objects that have only oop pointers
350 // - objects that have both native and oop pointers
351 // - objects that have only native pointers
352 // - objects that have no pointers
353 static int oop_sorting_rank(oop o) {
354 bool has_oop_ptr, has_native_ptr;
355 HeapShared::get_pointer_info(o, has_oop_ptr, has_native_ptr);
356
357 if (has_oop_ptr) {
358 if (!has_native_ptr) {
359 return 0;
360 } else {
361 return 1;
362 }
363 } else {
364 if (has_native_ptr) {
365 return 2;
366 } else {
367 return 3;
368 }
369 }
370 }
371
372 int AOTMappedHeapWriter::compare_objs_by_oop_fields(HeapObjOrder* a, HeapObjOrder* b) {
373 int rank_a = a->_rank;
374 int rank_b = b->_rank;
375
376 if (rank_a != rank_b) {
377 return rank_a - rank_b;
378 } else {
379 // If they are the same rank, sort them by their position in the _source_objs array
380 return a->_index - b->_index;
381 }
382 }
383
384 void AOTMappedHeapWriter::sort_source_objs() {
385 log_info(aot)("sorting heap objects");
386 int len = _source_objs->length();
387 _source_objs_order = new GrowableArrayCHeap<HeapObjOrder, mtClassShared>(len);
388
389 for (int i = 0; i < len; i++) {
390 oop o = _source_objs->at(i);
391 int rank = oop_sorting_rank(o);
392 HeapObjOrder os = {i, rank};
393 _source_objs_order->append(os);
394 }
395 log_info(aot)("computed ranks");
396 _source_objs_order->sort(compare_objs_by_oop_fields);
397 log_info(aot)("sorting heap objects done");
398 }
399
400 void AOTMappedHeapWriter::copy_source_objs_to_buffer(GrowableArrayCHeap<oop, mtClassShared>* roots) {
401 // There could be multiple root segments, which we want to be aligned by region.
402 // Putting them ahead of objects makes sure we waste no space.
403 copy_roots_to_buffer(roots);
404
405 sort_source_objs();
406 for (int i = 0; i < _source_objs_order->length(); i++) {
407 int src_obj_index = _source_objs_order->at(i)._index;
408 oop src_obj = _source_objs->at(src_obj_index);
409 HeapShared::CachedOopInfo* info = HeapShared::get_cached_oop_info(src_obj);
410 assert(info != nullptr, "must be");
411 size_t buffer_offset = copy_one_source_obj_to_buffer(src_obj);
412 info->set_buffer_offset(buffer_offset);
413 assert(buffer_offset <= 0x7fffffff, "sanity");
414
415 OopHandle handle(Universe::vm_global(), src_obj);
416 _buffer_offset_to_source_obj_table->put_when_absent(buffer_offset, handle);
417 _buffer_offset_to_source_obj_table->maybe_grow();
418
419 if (java_lang_Module::is_instance(src_obj)) {
420 Modules::check_archived_module_oop(src_obj);
421 }
422 }
423
424 log_info(aot)("Size of heap region = %zu bytes, %d objects, %d roots, %d native ptrs",
425 _buffer_used, _source_objs->length() + 1, roots->length(), _num_native_ptrs);
426 log_info(cds)(" strings = %8zu (%zu bytes)", _num_strings, _string_bytes);
427 log_info(cds)(" packages = %8zu", _num_packages);
428 log_info(cds)(" protection domains = %8zu", _num_protection_domains);
429 }
430
431 size_t AOTMappedHeapWriter::filler_array_byte_size(int length) {
432 size_t byte_size = objArrayOopDesc::object_size(length) * HeapWordSize;
433 return byte_size;
434 }
435
436 int AOTMappedHeapWriter::filler_array_length(size_t fill_bytes) {
437 assert(is_object_aligned(fill_bytes), "must be");
438 size_t elemSize = (UseCompressedOops ? sizeof(narrowOop) : sizeof(oop));
439
440 int initial_length = to_array_length(fill_bytes / elemSize);
441 for (int length = initial_length; length >= 0; length --) {
442 size_t array_byte_size = filler_array_byte_size(length);
443 if (array_byte_size == fill_bytes) {
444 return length;
445 }
446 }
447
448 ShouldNotReachHere();
449 return -1;
450 }
451
452 HeapWord* AOTMappedHeapWriter::init_filler_array_at_buffer_top(int array_length, size_t fill_bytes) {
453 assert(UseCompressedClassPointers, "Archived heap only supported for compressed klasses");
454 Klass* oak = Universe::objectArrayKlass(); // already relocated to point to archived klass
455 HeapWord* mem = offset_to_buffered_address<HeapWord*>(_buffer_used);
456 memset(mem, 0, fill_bytes);
457 narrowKlass nk = ArchiveBuilder::current()->get_requested_narrow_klass(oak);
458 if (UseCompactObjectHeaders) {
459 oopDesc::release_set_mark(mem, markWord::prototype().set_narrow_klass(nk));
460 } else {
461 oopDesc::set_mark(mem, markWord::prototype());
462 cast_to_oop(mem)->set_narrow_klass(nk);
463 }
464 arrayOopDesc::set_length(mem, array_length);
465 return mem;
466 }
467
468 void AOTMappedHeapWriter::maybe_fill_gc_region_gap(size_t required_byte_size) {
469 // We fill only with arrays (so we don't need to use a single HeapWord filler if the
470 // leftover space is smaller than a zero-sized array object). Therefore, we need to
471 // make sure there's enough space of min_filler_byte_size in the current region after
472 // required_byte_size has been allocated. If not, fill the remainder of the current
473 // region.
474 size_t min_filler_byte_size = filler_array_byte_size(0);
475 size_t new_used = _buffer_used + required_byte_size + min_filler_byte_size;
476
477 const size_t cur_min_region_bottom = align_down(_buffer_used, MIN_GC_REGION_ALIGNMENT);
478 const size_t next_min_region_bottom = align_down(new_used, MIN_GC_REGION_ALIGNMENT);
479
480 if (cur_min_region_bottom != next_min_region_bottom) {
481 // Make sure that no objects span across MIN_GC_REGION_ALIGNMENT. This way
482 // we can map the region in any region-based collector.
483 assert(next_min_region_bottom > cur_min_region_bottom, "must be");
484 assert(next_min_region_bottom - cur_min_region_bottom == MIN_GC_REGION_ALIGNMENT,
485 "no buffered object can be larger than %d bytes", MIN_GC_REGION_ALIGNMENT);
486
487 const size_t filler_end = next_min_region_bottom;
488 const size_t fill_bytes = filler_end - _buffer_used;
489 assert(fill_bytes > 0, "must be");
490 ensure_buffer_space(filler_end);
491
492 int array_length = filler_array_length(fill_bytes);
493 log_info(aot, heap)("Inserting filler obj array of %d elements (%zu bytes total) @ buffer offset %zu",
494 array_length, fill_bytes, _buffer_used);
495 HeapWord* filler = init_filler_array_at_buffer_top(array_length, fill_bytes);
496 _buffer_used = filler_end;
497 _fillers->put(buffered_address_to_offset((address)filler), fill_bytes);
498 }
499 }
500
501 size_t AOTMappedHeapWriter::get_filler_size_at(address buffered_addr) {
502 size_t* p = _fillers->get(buffered_address_to_offset(buffered_addr));
503 if (p != nullptr) {
504 assert(*p > 0, "filler must be larger than zero bytes");
505 return *p;
506 } else {
507 return 0; // buffered_addr is not a filler
508 }
509 }
510
511 template <typename T>
512 void update_buffered_object_field(address buffered_obj, int field_offset, T value) {
513 T* field_addr = cast_to_oop(buffered_obj)->field_addr<T>(field_offset);
514 *field_addr = value;
515 }
516
517 void AOTMappedHeapWriter::update_stats(oop src_obj) {
518 if (java_lang_String::is_instance(src_obj)) {
519 _num_strings ++;
520 _string_bytes += src_obj->size() * HeapWordSize;
521 _string_bytes += java_lang_String::value(src_obj)->size() * HeapWordSize;
522 } else {
523 Klass* k = src_obj->klass();
524 Symbol* name = k->name();
525 if (name->equals("java/lang/NamedPackage") || name->equals("java/lang/Package")) {
526 _num_packages ++;
527 } else if (name->equals("java/security/ProtectionDomain")) {
528 _num_protection_domains ++;
529 }
530 }
531 }
532
533 size_t AOTMappedHeapWriter::copy_one_source_obj_to_buffer(oop src_obj) {
534 update_stats(src_obj);
535
536 assert(!is_too_large_to_archive(src_obj), "already checked");
537 size_t byte_size = src_obj->size() * HeapWordSize;
538 assert(byte_size > 0, "no zero-size objects");
539
540 // For region-based collectors such as G1, the archive heap may be mapped into
541 // multiple regions. We need to make sure that we don't have an object that can possible
542 // span across two regions.
543 maybe_fill_gc_region_gap(byte_size);
544
545 size_t new_used = _buffer_used + byte_size;
546 assert(new_used > _buffer_used, "no wrap around");
547
548 size_t cur_min_region_bottom = align_down(_buffer_used, MIN_GC_REGION_ALIGNMENT);
549 size_t next_min_region_bottom = align_down(new_used, MIN_GC_REGION_ALIGNMENT);
550 assert(cur_min_region_bottom == next_min_region_bottom, "no object should cross minimal GC region boundaries");
551
552 ensure_buffer_space(new_used);
553
554 address from = cast_from_oop<address>(src_obj);
555 address to = offset_to_buffered_address<address>(_buffer_used);
556 assert(is_object_aligned(_buffer_used), "sanity");
557 assert(is_object_aligned(byte_size), "sanity");
558 memcpy(to, from, byte_size);
559
560 // These native pointers will be restored explicitly at run time.
561 if (java_lang_Module::is_instance(src_obj)) {
562 update_buffered_object_field<ModuleEntry*>(to, java_lang_Module::module_entry_offset(), nullptr);
563 } else if (java_lang_ClassLoader::is_instance(src_obj)) {
564 #ifdef ASSERT
565 // We only archive these loaders
566 if (src_obj != SystemDictionary::java_platform_loader() &&
567 src_obj != SystemDictionary::java_system_loader()) {
568 assert(src_obj->klass()->name()->equals("jdk/internal/loader/ClassLoaders$BootClassLoader"), "must be");
569 }
570 #endif
571 update_buffered_object_field<ClassLoaderData*>(to, java_lang_ClassLoader::loader_data_offset(), nullptr);
572 }
573
574 size_t buffered_obj_offset = _buffer_used;
575 _buffer_used = new_used;
576
577 return buffered_obj_offset;
578 }
579
580 // Set the range [_requested_bottom, _requested_top), the requested address range of all
581 // the archived heap objects in the production run.
582 //
583 // (1) UseCompressedOops == true && !is_writing_deterministic_heap()
584 //
585 // The archived objects are stored using the COOPS encoding of the assembly phase.
586 // We pick a range within the heap used by the assembly phase.
587 //
588 // In the production run, if different COOPS encodings are used:
589 // - The heap contents needs to be relocated.
590 //
591 // (2) UseCompressedOops == true && is_writing_deterministic_heap()
592 //
593 // We always use zero-based, zero-shift encoding. _requested_top is aligned to 0x10000000.
594 //
595 // (3) UseCompressedOops == false:
596 //
597 // In the production run, the heap range is usually picked (randomly) by the OS, so we
598 // will almost always need to perform relocation, regardless of how we pick the requested
599 // address range.
600 //
601 // So we just hard code it to NOCOOPS_REQUESTED_BASE.
602 //
603 void AOTMappedHeapWriter::set_requested_address_range(AOTMappedHeapInfo* info) {
604 assert(!info->is_used(), "only set once");
605
606 size_t heap_region_byte_size = _buffer_used;
607 assert(heap_region_byte_size > 0, "must archived at least one object!");
608
609 if (UseCompressedOops) {
610 if (is_writing_deterministic_heap()) {
611 // Pick a heap range so that requested addresses can be encoded with zero-base/no shift.
612 // We align the requested bottom to at least 1 MB: if the production run uses G1 with a small
613 // heap (e.g., -Xmx256m), it's likely that we can map the archived objects at the
614 // requested location to avoid relocation.
615 //
616 // For other collectors or larger heaps, relocation is unavoidable, but is usually
617 // quite cheap. If you really want to avoid relocation, use the AOT workflow instead.
618 address heap_end = (address)0x100000000;
619 size_t alignment = MAX2(MIN_GC_REGION_ALIGNMENT, 1024 * 1024);
620 if (align_up(heap_region_byte_size, alignment) >= (size_t)heap_end) {
621 log_error(aot, heap)("cached heap space is too large: %zu bytes", heap_region_byte_size);
622 AOTMetaspace::unrecoverable_writing_error();
623 }
624 _requested_bottom = align_down(heap_end - heap_region_byte_size, alignment);
625 } else if (UseG1GC) {
626 // For G1, pick the range at the top of the current heap. If the exact same heap sizes
627 // are used in the production run, it's likely that we can map the archived objects
628 // at the requested location to avoid relocation.
629 address heap_end = (address)G1CollectedHeap::heap()->reserved().end();
630 log_info(aot, heap)("Heap end = %p", heap_end);
631 _requested_bottom = align_down(heap_end - heap_region_byte_size, G1HeapRegion::GrainBytes);
632 _requested_bottom = align_down(_requested_bottom, MIN_GC_REGION_ALIGNMENT);
633 assert(is_aligned(_requested_bottom, G1HeapRegion::GrainBytes), "sanity");
634 } else {
635 _requested_bottom = align_up(CompressedOops::begin(), MIN_GC_REGION_ALIGNMENT);
636 }
637 } else {
638 // We always write the objects as if the heap started at this address. This
639 // makes the contents of the archive heap deterministic.
640 //
641 // Note that at runtime, the heap address is selected by the OS, so the archive
642 // heap will not be mapped at 0x10000000, and the contents need to be patched.
643 _requested_bottom = align_up((address)NOCOOPS_REQUESTED_BASE, MIN_GC_REGION_ALIGNMENT);
644 }
645
646 assert(is_aligned(_requested_bottom, MIN_GC_REGION_ALIGNMENT), "sanity");
647
648 _requested_top = _requested_bottom + _buffer_used;
649
650 info->set_buffer_region(MemRegion(offset_to_buffered_address<HeapWord*>(0),
651 offset_to_buffered_address<HeapWord*>(_buffer_used)));
652 info->set_root_segments(_heap_root_segments);
653 }
654
655 // Oop relocation
656
657 template <typename T> T* AOTMappedHeapWriter::requested_addr_to_buffered_addr(T* p) {
658 assert(is_in_requested_range(cast_to_oop(p)), "must be");
659
660 address addr = address(p);
661 assert(addr >= _requested_bottom, "must be");
662 size_t offset = addr - _requested_bottom;
663 return offset_to_buffered_address<T*>(offset);
664 }
665
666 template <typename T> oop AOTMappedHeapWriter::load_source_oop_from_buffer(T* buffered_addr) {
667 oop o = load_oop_from_buffer(buffered_addr);
668 assert(!in_buffer(cast_from_oop<address>(o)), "must point to source oop");
669 return o;
670 }
671
672 template <typename T> void AOTMappedHeapWriter::store_requested_oop_in_buffer(T* buffered_addr,
673 oop request_oop) {
674 assert(request_oop == nullptr || is_in_requested_range(request_oop), "must be");
675 store_oop_in_buffer(buffered_addr, request_oop);
676 }
677
678 inline void AOTMappedHeapWriter::store_oop_in_buffer(oop* buffered_addr, oop requested_obj) {
679 *buffered_addr = requested_obj;
680 }
681
682 inline void AOTMappedHeapWriter::store_oop_in_buffer(narrowOop* buffered_addr, oop requested_obj) {
683 narrowOop val = CompressedOops::encode(requested_obj);
684 *buffered_addr = val;
685 }
686
687 oop AOTMappedHeapWriter::load_oop_from_buffer(oop* buffered_addr) {
688 return *buffered_addr;
689 }
690
691 oop AOTMappedHeapWriter::load_oop_from_buffer(narrowOop* buffered_addr) {
692 return CompressedOops::decode(*buffered_addr);
693 }
694
695 template <typename T> void AOTMappedHeapWriter::relocate_field_in_buffer(T* field_addr_in_buffer, oop source_referent, CHeapBitMap* oopmap) {
696 oop request_referent = source_obj_to_requested_obj(source_referent);
697 if (UseCompressedOops && is_writing_deterministic_heap()) {
698 // We use zero-based, 0-shift encoding, so the narrowOop is just the lower
699 // 32 bits of request_referent
700 intptr_t addr = cast_from_oop<intptr_t>(request_referent);
701 *((narrowOop*)field_addr_in_buffer) = CompressedOops::narrow_oop_cast(addr);
702 } else {
703 store_requested_oop_in_buffer<T>(field_addr_in_buffer, request_referent);
704 }
705 if (request_referent != nullptr) {
706 mark_oop_pointer<T>(field_addr_in_buffer, oopmap);
707 }
708 }
709
710 template <typename T> void AOTMappedHeapWriter::mark_oop_pointer(T* buffered_addr, CHeapBitMap* oopmap) {
711 T* request_p = (T*)(buffered_addr_to_requested_addr((address)buffered_addr));
712 address requested_region_bottom;
713
714 assert(request_p >= (T*)_requested_bottom, "sanity");
715 assert(request_p < (T*)_requested_top, "sanity");
716 requested_region_bottom = _requested_bottom;
717
718 // Mark the pointer in the oopmap
719 T* region_bottom = (T*)requested_region_bottom;
720 assert(request_p >= region_bottom, "must be");
721 BitMap::idx_t idx = request_p - region_bottom;
722 assert(idx < oopmap->size(), "overflow");
723 oopmap->set_bit(idx);
724 }
725
726 void AOTMappedHeapWriter::update_header_for_requested_obj(oop requested_obj, oop src_obj, Klass* src_klass) {
727 assert(UseCompressedClassPointers, "Archived heap only supported for compressed klasses");
728 narrowKlass nk = ArchiveBuilder::current()->get_requested_narrow_klass(src_klass);
729 address buffered_addr = requested_addr_to_buffered_addr(cast_from_oop<address>(requested_obj));
730
731 oop fake_oop = cast_to_oop(buffered_addr);
732 if (UseCompactObjectHeaders) {
733 fake_oop->set_mark(markWord::prototype().set_narrow_klass(nk));
734 } else {
735 fake_oop->set_narrow_klass(nk);
736 }
737
738 if (src_obj == nullptr) {
739 return;
740 }
741 // We need to retain the identity_hash, because it may have been used by some hashtables
742 // in the shared heap.
743 if (!src_obj->fast_no_hash_check()) {
744 intptr_t src_hash = src_obj->identity_hash();
745 if (UseCompactObjectHeaders) {
746 fake_oop->set_mark(markWord::prototype().set_narrow_klass(nk).copy_set_hash(src_hash));
747 } else {
748 fake_oop->set_mark(markWord::prototype().copy_set_hash(src_hash));
749 }
750 assert(fake_oop->mark().is_unlocked(), "sanity");
751
752 DEBUG_ONLY(intptr_t archived_hash = fake_oop->identity_hash());
753 assert(src_hash == archived_hash, "Different hash codes: original " INTPTR_FORMAT ", archived " INTPTR_FORMAT, src_hash, archived_hash);
754 }
755 // Strip age bits.
756 fake_oop->set_mark(fake_oop->mark().set_age(0));
757 }
758
759 class AOTMappedHeapWriter::EmbeddedOopRelocator: public BasicOopIterateClosure {
760 oop _src_obj;
761 address _buffered_obj;
762 CHeapBitMap* _oopmap;
763 bool _is_java_lang_ref;
764 public:
765 EmbeddedOopRelocator(oop src_obj, address buffered_obj, CHeapBitMap* oopmap) :
766 _src_obj(src_obj), _buffered_obj(buffered_obj), _oopmap(oopmap)
767 {
768 _is_java_lang_ref = AOTReferenceObjSupport::check_if_ref_obj(src_obj);
769 }
770
771 void do_oop(narrowOop *p) { EmbeddedOopRelocator::do_oop_work(p); }
772 void do_oop( oop *p) { EmbeddedOopRelocator::do_oop_work(p); }
773
774 private:
775 template <class T> void do_oop_work(T *p) {
776 int field_offset = pointer_delta_as_int((char*)p, cast_from_oop<char*>(_src_obj));
777 T* field_addr = (T*)(_buffered_obj + field_offset);
778 oop referent = load_source_oop_from_buffer<T>(field_addr);
779 referent = HeapShared::maybe_remap_referent(_is_java_lang_ref, field_offset, referent);
780 AOTMappedHeapWriter::relocate_field_in_buffer<T>(field_addr, referent, _oopmap);
781 }
782 };
783
784 static void log_bitmap_usage(const char* which, BitMap* bitmap, size_t total_bits) {
785 // The whole heap is covered by total_bits, but there are only non-zero bits within [start ... end).
786 size_t start = bitmap->find_first_set_bit(0);
787 size_t end = bitmap->size();
788 log_info(aot)("%s = %7zu ... %7zu (%3zu%% ... %3zu%% = %3zu%%)", which,
789 start, end,
790 start * 100 / total_bits,
791 end * 100 / total_bits,
792 (end - start) * 100 / total_bits);
793 }
794
795 // Update all oop fields embedded in the buffered objects
796 void AOTMappedHeapWriter::relocate_embedded_oops(GrowableArrayCHeap<oop, mtClassShared>* roots,
797 AOTMappedHeapInfo* heap_info) {
798 size_t oopmap_unit = (UseCompressedOops ? sizeof(narrowOop) : sizeof(oop));
799 size_t heap_region_byte_size = _buffer_used;
800 heap_info->oopmap()->resize(heap_region_byte_size / oopmap_unit);
801
802 for (int i = 0; i < _source_objs_order->length(); i++) {
803 int src_obj_index = _source_objs_order->at(i)._index;
804 oop src_obj = _source_objs->at(src_obj_index);
805 HeapShared::CachedOopInfo* info = HeapShared::get_cached_oop_info(src_obj);
806 assert(info != nullptr, "must be");
807 oop requested_obj = requested_obj_from_buffer_offset(info->buffer_offset());
808 update_header_for_requested_obj(requested_obj, src_obj, src_obj->klass());
809 address buffered_obj = offset_to_buffered_address<address>(info->buffer_offset());
810 EmbeddedOopRelocator relocator(src_obj, buffered_obj, heap_info->oopmap());
811 src_obj->oop_iterate(&relocator);
812 mark_native_pointers(src_obj);
813 };
814
815 // Relocate HeapShared::roots(), which is created in copy_roots_to_buffer() and
816 // doesn't have a corresponding src_obj, so we can't use EmbeddedOopRelocator on it.
817 for (size_t seg_idx = 0; seg_idx < _heap_root_segments.count(); seg_idx++) {
818 size_t seg_offset = _heap_root_segments.segment_offset(seg_idx);
819
820 objArrayOop requested_obj = (objArrayOop)requested_obj_from_buffer_offset(seg_offset);
821 update_header_for_requested_obj(requested_obj, nullptr, Universe::objectArrayKlass());
822 address buffered_obj = offset_to_buffered_address<address>(seg_offset);
823 int length = _heap_root_segments.size_in_elems(seg_idx);
824
825 size_t elem_size = UseCompressedOops ? sizeof(narrowOop) : sizeof(oop);
826
827 for (int i = 0; i < length; i++) {
828 // There is no source object; these are native oops - load, translate and
829 // write back
830 size_t elem_offset = objArrayOopDesc::base_offset_in_bytes() + elem_size * i;
831 HeapWord* elem_addr = (HeapWord*)(buffered_obj + elem_offset);
832 oop obj = NativeAccess<>::oop_load(elem_addr);
833 obj = HeapShared::maybe_remap_referent(false /* is_reference_field */, elem_offset, obj);
834 if (UseCompressedOops) {
835 relocate_field_in_buffer<narrowOop>((narrowOop*)elem_addr, obj, heap_info->oopmap());
836 } else {
837 relocate_field_in_buffer<oop>((oop*)elem_addr, obj, heap_info->oopmap());
838 }
839 }
840 }
841
842 compute_ptrmap(heap_info);
843
844 size_t total_bytes = (size_t)_buffer->length();
845 log_bitmap_usage("oopmap", heap_info->oopmap(), total_bytes / (UseCompressedOops ? sizeof(narrowOop) : sizeof(oop)));
846 log_bitmap_usage("ptrmap", heap_info->ptrmap(), total_bytes / sizeof(address));
847 }
848
849 void AOTMappedHeapWriter::mark_native_pointer(oop src_obj, int field_offset) {
850 Metadata* ptr = src_obj->metadata_field_acquire(field_offset);
851 if (ptr != nullptr) {
852 NativePointerInfo info;
853 info._src_obj = src_obj;
854 info._field_offset = field_offset;
855 _native_pointers->append(info);
856 HeapShared::set_has_native_pointers(src_obj);
857 _num_native_ptrs ++;
858 }
859 }
860
861 void AOTMappedHeapWriter::mark_native_pointers(oop orig_obj) {
862 HeapShared::do_metadata_offsets(orig_obj, [&](int offset) {
863 mark_native_pointer(orig_obj, offset);
864 });
865 }
866
867 void AOTMappedHeapWriter::compute_ptrmap(AOTMappedHeapInfo* heap_info) {
868 int num_non_null_ptrs = 0;
869 Metadata** bottom = (Metadata**) _requested_bottom;
870 Metadata** top = (Metadata**) _requested_top; // exclusive
871 heap_info->ptrmap()->resize(top - bottom);
872
873 BitMap::idx_t max_idx = 32; // paranoid - don't make it too small
874 for (int i = 0; i < _native_pointers->length(); i++) {
875 NativePointerInfo info = _native_pointers->at(i);
876 oop src_obj = info._src_obj;
877 int field_offset = info._field_offset;
878 HeapShared::CachedOopInfo* p = HeapShared::get_cached_oop_info(src_obj);
879 // requested_field_addr = the address of this field in the requested space
880 oop requested_obj = requested_obj_from_buffer_offset(p->buffer_offset());
881 Metadata** requested_field_addr = (Metadata**)(cast_from_oop<address>(requested_obj) + field_offset);
882 assert(bottom <= requested_field_addr && requested_field_addr < top, "range check");
883
884 // Mark this field in the bitmap
885 BitMap::idx_t idx = requested_field_addr - bottom;
886 heap_info->ptrmap()->set_bit(idx);
887 num_non_null_ptrs ++;
888 max_idx = MAX2(max_idx, idx);
889
890 // Set the native pointer to the requested address of the metadata (at runtime, the metadata will have
891 // this address if the RO/RW regions are mapped at the default location).
892
893 Metadata** buffered_field_addr = requested_addr_to_buffered_addr(requested_field_addr);
894 Metadata* native_ptr = *buffered_field_addr;
895 guarantee(native_ptr != nullptr, "sanity");
896
897 if (RegeneratedClasses::has_been_regenerated(native_ptr)) {
898 native_ptr = RegeneratedClasses::get_regenerated_object(native_ptr);
899 }
900
901 guarantee(ArchiveBuilder::current()->has_been_archived((address)native_ptr),
902 "Metadata %p should have been archived", native_ptr);
903
904 address buffered_native_ptr = ArchiveBuilder::current()->get_buffered_addr((address)native_ptr);
905 address requested_native_ptr = ArchiveBuilder::current()->to_requested(buffered_native_ptr);
906 *buffered_field_addr = (Metadata*)requested_native_ptr;
907 }
908
909 heap_info->ptrmap()->resize(max_idx + 1);
910 log_info(aot, heap)("calculate_ptrmap: marked %d non-null native pointers for heap region (%zu bits)",
911 num_non_null_ptrs, size_t(heap_info->ptrmap()->size()));
912 }
913
914 AOTMapLogger::OopDataIterator* AOTMappedHeapWriter::oop_iterator(AOTMappedHeapInfo* heap_info) {
915 class MappedWriterOopIterator : public AOTMappedHeapOopIterator {
916 public:
917 MappedWriterOopIterator(address buffer_start,
918 address buffer_end,
919 address requested_base,
920 address requested_start,
921 int requested_shift,
922 size_t num_root_segments) :
923 AOTMappedHeapOopIterator(buffer_start,
924 buffer_end,
925 requested_base,
926 requested_start,
927 requested_shift,
928 num_root_segments) {}
929
930 AOTMapLogger::OopData capture(address buffered_addr) override {
931 oopDesc* raw_oop = (oopDesc*)buffered_addr;
932 size_t size = size_of_buffered_oop(buffered_addr);
933 address requested_addr = buffered_addr_to_requested_addr(buffered_addr);
934 intptr_t target_location = (intptr_t)requested_addr;
935 uint64_t pd = (uint64_t)(pointer_delta(buffered_addr, _buffer_start, 1));
936 uint32_t narrow_location = checked_cast<uint32_t>(_buffer_start_narrow_oop + (pd >> _requested_shift));
937 Klass* klass = real_klass_of_buffered_oop(buffered_addr);
938
939 return { buffered_addr,
940 requested_addr,
941 target_location,
942 narrow_location,
943 raw_oop,
944 klass,
945 size,
946 false };
947 }
948 };
949
950 MemRegion r = heap_info->buffer_region();
951 address buffer_start = address(r.start());
952 address buffer_end = address(r.end());
953
954 address requested_base = UseCompressedOops ? AOTMappedHeapWriter::narrow_oop_base() : (address)AOTMappedHeapWriter::NOCOOPS_REQUESTED_BASE;
955 address requested_start = UseCompressedOops ? AOTMappedHeapWriter::buffered_addr_to_requested_addr(buffer_start) : requested_base;
956 int requested_shift = AOTMappedHeapWriter::narrow_oop_shift();
957
958 return new MappedWriterOopIterator(buffer_start,
959 buffer_end,
960 requested_base,
961 requested_start,
962 requested_shift,
963 heap_info->root_segments().count());
964 }
965
966 #endif // INCLUDE_CDS_JAVA_HEAP