1 /*
2 * Copyright (c) 2019, 2025, 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 #include "gc/z/zGeneration.inline.hpp"
25 #include "gc/z/zObjArrayAllocator.hpp"
26 #include "gc/z/zThreadLocalData.hpp"
27 #include "gc/z/zUtils.inline.hpp"
28 #include "oops/arrayKlass.hpp"
29 #include "runtime/interfaceSupport.inline.hpp"
30 #include "utilities/debug.hpp"
31 #include "utilities/globalDefinitions.hpp"
32
33 ZObjArrayAllocator::ZObjArrayAllocator(Klass* klass, size_t word_size, int length, bool do_zero, Thread* thread)
34 : ObjArrayAllocator(klass, word_size, length, do_zero, thread) {}
35
36 void ZObjArrayAllocator::yield_for_safepoint() const {
37 ThreadBlockInVM tbivm(JavaThread::cast(_thread));
38 }
39
40 oop ZObjArrayAllocator::initialize(HeapWord* mem) const {
41 // ZGC specializes the initialization by performing segmented clearing
42 // to allow shorter time-to-safepoints.
43
44 if (!_do_zero) {
45 // No need for ZGC specialization
46 return ObjArrayAllocator::initialize(mem);
47 }
48
49 // A max segment size of 64K was chosen because microbenchmarking
50 // suggested that it offered a good trade-off between allocation
51 // time and time-to-safepoint
52 const size_t segment_max = ZUtils::bytes_to_words(64 * K);
53
54 if (_word_size <= segment_max || ArrayKlass::cast(_klass)->is_flatArray_klass()) {
55 // To small to use segmented clearing
56 return ObjArrayAllocator::initialize(mem);
57 }
58
59 // Segmented clearing
60
61 // The array is going to be exposed before it has been completely
62 // cleared, therefore we can't expose the header at the end of this
63 // function. Instead explicitly initialize it according to our needs.
64
65 // Signal to the ZIterator that this is an invisible root, by setting
66 // the mark word to "marked". Reset to prototype() after the clearing.
67 if (UseCompactObjectHeaders) {
68 oopDesc::release_set_mark(mem, _klass->prototype_header().set_marked());
69 } else {
70 if (EnableValhalla) {
71 arrayOopDesc::set_mark(mem, _klass->prototype_header().set_marked());
72 } else {
73 arrayOopDesc::set_mark(mem, markWord::prototype().set_marked());
74 }
75 arrayOopDesc::release_set_klass(mem, _klass);
76 }
77 assert(_length >= 0, "length should be non-negative");
78 arrayOopDesc::set_length(mem, _length);
79
80 // Keep the array alive across safepoints through an invisible
81 // root. Invisible roots are not visited by the heap iterator
82 // and the marking logic will not attempt to follow its elements.
83 // Relocation and remembered set code know how to dodge iterating
84 // over such objects.
85 ZThreadLocalData::set_invisible_root(_thread, (zaddress_unsafe*)&mem);
86
87 const BasicType element_type = ArrayKlass::cast(_klass)->element_type();
88 const size_t base_offset_in_bytes = (size_t)arrayOopDesc::base_offset_in_bytes(element_type);
89 const size_t process_start_offset_in_bytes = align_up(base_offset_in_bytes, (size_t)BytesPerWord);
90
91 if (process_start_offset_in_bytes != base_offset_in_bytes) {
92 // initialize_memory can only fill word aligned memory,
93 // fill the first 4 bytes here.
94 assert(process_start_offset_in_bytes - base_offset_in_bytes == 4, "Must be 4-byte aligned");
95 assert(!is_reference_type(element_type), "Only TypeArrays can be 4-byte aligned");
96 *reinterpret_cast<int*>(reinterpret_cast<char*>(mem) + base_offset_in_bytes) = 0;
97 }
98
99 // Note: initialize_memory may clear padding bytes at the end
100 const size_t process_start_offset = ZUtils::bytes_to_words(process_start_offset_in_bytes);
101 const size_t process_size = _word_size - process_start_offset;
102
103 uint32_t old_seqnum_before = ZGeneration::old()->seqnum();
104 uint32_t young_seqnum_before = ZGeneration::young()->seqnum();
105 uintptr_t color_before = ZPointerStoreGoodMask;
106 auto gc_safepoint_happened = [&]() {
107 return old_seqnum_before != ZGeneration::old()->seqnum() ||
108 young_seqnum_before != ZGeneration::young()->seqnum() ||
109 color_before != ZPointerStoreGoodMask;
110 };
111
112 bool seen_gc_safepoint = false;
113
114 auto initialize_memory = [&]() {
115 for (size_t processed = 0; processed < process_size; processed += segment_max) {
116 // Clear segment
117 uintptr_t* const start = (uintptr_t*)(mem + process_start_offset + processed);
118 const size_t remaining = process_size - processed;
119 const size_t segment = MIN2(remaining, segment_max);
120 // Usually, the young marking code has the responsibility to color
121 // raw nulls, before they end up in the old generation. However, the
122 // invisible roots are hidden from the marking code, and therefore
123 // we must color the nulls already here in the initialization. The
124 // color we choose must be store bad for any subsequent stores, regardless
125 // of how many GC flips later it will arrive. That's why we OR in 11
126 // (ZPointerRememberedMask) in the remembered bits, similar to how
127 // forgotten old oops also have 11, for the very same reason.
128 // However, we opportunistically try to color without the 11 remembered
129 // bits, hoping to not get interrupted in the middle of a GC safepoint.
130 // Most of the time, we manage to do that, and can the avoid having GC
131 // barriers trigger slow paths for this.
132 const uintptr_t colored_null = seen_gc_safepoint ? (ZPointerStoreGoodMask | ZPointerRememberedMask)
133 : ZPointerStoreGoodMask;
134 const uintptr_t fill_value = is_reference_type(element_type) ? colored_null : 0;
135 ZUtils::fill(start, segment, fill_value);
136
137 // Safepoint
138 yield_for_safepoint();
139
140 // Deal with safepoints
141 if (is_reference_type(element_type) && !seen_gc_safepoint && gc_safepoint_happened()) {
142 // The first time we observe a GC safepoint in the yield point,
143 // we have to restart processing with 11 remembered bits.
144 seen_gc_safepoint = true;
145 return false;
146 }
147 }
148 return true;
149 };
150
151 mem_zap_start_padding(mem);
152
153 if (!initialize_memory()) {
154 // Re-color with 11 remset bits if we got intercepted by a GC safepoint
155 const bool result = initialize_memory();
156 assert(result, "Array initialization should always succeed the second time");
157 }
158
159 mem_zap_end_padding(mem);
160
161 ZThreadLocalData::clear_invisible_root(_thread);
162
163 // Signal to the ZIterator that this is no longer an invisible root
164 if (UseCompactObjectHeaders || EnableValhalla) {
165 oopDesc::release_set_mark(mem, _klass->prototype_header());
166 } else {
167 oopDesc::release_set_mark(mem, markWord::prototype());
168 }
169
170 return cast_to_oop(mem);
171 }