1/*
2 * Copyright (c) 2006, 2019, 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#ifndef SHARE_GC_PARALLEL_MUTABLENUMASPACE_HPP
26#define SHARE_GC_PARALLEL_MUTABLENUMASPACE_HPP
27
28#include "gc/parallel/mutableSpace.hpp"
29#include "gc/shared/gcUtil.hpp"
30#include "utilities/macros.hpp"
31
32/*
33 * The NUMA-aware allocator (MutableNUMASpace) is basically a modification
34 * of MutableSpace which preserves interfaces but implements different
35 * functionality. The space is split into chunks for each locality group
36 * (resizing for adaptive size policy is also supported). For each thread
37 * allocations are performed in the chunk corresponding to the home locality
38 * group of the thread. Whenever any chunk fills-in the young generation
39 * collection occurs.
40 * The chunks can be also be adaptively resized. The idea behind the adaptive
41 * sizing is to reduce the loss of the space in the eden due to fragmentation.
42 * The main cause of fragmentation is uneven allocation rates of threads.
43 * The allocation rate difference between locality groups may be caused either by
44 * application specifics or by uneven LWP distribution by the OS. Besides,
45 * application can have less threads then the number of locality groups.
46 * In order to resize the chunk we measure the allocation rate of the
47 * application between collections. After that we reshape the chunks to reflect
48 * the allocation rate pattern. The AdaptiveWeightedAverage exponentially
49 * decaying average is used to smooth the measurements. The NUMASpaceResizeRate
50 * parameter is used to control the adaptation speed by restricting the number of
51 * bytes that can be moved during the adaptation phase.
52 * Chunks may contain pages from a wrong locality group. The page-scanner has
53 * been introduced to address the problem. Remote pages typically appear due to
54 * the memory shortage in the target locality group. Besides Solaris would
55 * allocate a large page from the remote locality group even if there are small
56 * local pages available. The page-scanner scans the pages right after the
57 * collection and frees remote pages in hope that subsequent reallocation would
58 * be more successful. This approach proved to be useful on systems with high
59 * load where multiple processes are competing for the memory.
60 */
61
62class MutableNUMASpace : public MutableSpace {
63 friend class VMStructs;
64
65 class LGRPSpace : public CHeapObj<mtGC> {
66 int _lgrp_id;
67 MutableSpace* _space;
68 MemRegion _invalid_region;
69 AdaptiveWeightedAverage *_alloc_rate;
70 bool _allocation_failed;
71
72 struct SpaceStats {
73 size_t _local_space, _remote_space, _unbiased_space, _uncommited_space;
74 size_t _large_pages, _small_pages;
75
76 SpaceStats() {
77 _local_space = 0;
78 _remote_space = 0;
79 _unbiased_space = 0;
80 _uncommited_space = 0;
81 _large_pages = 0;
82 _small_pages = 0;
83 }
84 };
85
86 SpaceStats _space_stats;
87
88 char* _last_page_scanned;
89 char* last_page_scanned() { return _last_page_scanned; }
90 void set_last_page_scanned(char* p) { _last_page_scanned = p; }
91 public:
92 LGRPSpace(int l, size_t alignment) : _lgrp_id(l), _allocation_failed(false), _last_page_scanned(NULL) {
93 _space = new MutableSpace(alignment);
94 _alloc_rate = new AdaptiveWeightedAverage(NUMAChunkResizeWeight);
95 }
96 ~LGRPSpace() {
97 delete _space;
98 delete _alloc_rate;
99 }
100
101 void add_invalid_region(MemRegion r) {
102 if (!_invalid_region.is_empty()) {
103 _invalid_region.set_start(MIN2(_invalid_region.start(), r.start()));
104 _invalid_region.set_end(MAX2(_invalid_region.end(), r.end()));
105 } else {
106 _invalid_region = r;
107 }
108 }
109
110 static bool equals(void* lgrp_id_value, LGRPSpace* p) {
111 return *(int*)lgrp_id_value == p->lgrp_id();
112 }
113
114 // Report a failed allocation.
115 void set_allocation_failed() { _allocation_failed = true; }
116
117 void sample() {
118 // If there was a failed allocation make allocation rate equal
119 // to the size of the whole chunk. This ensures the progress of
120 // the adaptation process.
121 size_t alloc_rate_sample;
122 if (_allocation_failed) {
123 alloc_rate_sample = space()->capacity_in_bytes();
124 _allocation_failed = false;
125 } else {
126 alloc_rate_sample = space()->used_in_bytes();
127 }
128 alloc_rate()->sample(alloc_rate_sample);
129 }
130
131 MemRegion invalid_region() const { return _invalid_region; }
132 void set_invalid_region(MemRegion r) { _invalid_region = r; }
133 int lgrp_id() const { return _lgrp_id; }
134 MutableSpace* space() const { return _space; }
135 AdaptiveWeightedAverage* alloc_rate() const { return _alloc_rate; }
136 void clear_alloc_rate() { _alloc_rate->clear(); }
137 SpaceStats* space_stats() { return &_space_stats; }
138 void clear_space_stats() { _space_stats = SpaceStats(); }
139
140 void accumulate_statistics(size_t page_size);
141 void scan_pages(size_t page_size, size_t page_count);
142 };
143
144 GrowableArray<LGRPSpace*>* _lgrp_spaces;
145 size_t _page_size;
146 unsigned _adaptation_cycles, _samples_count;
147
148 bool _must_use_large_pages;
149
150 void set_page_size(size_t psz) { _page_size = psz; }
151 size_t page_size() const { return _page_size; }
152
153 unsigned adaptation_cycles() { return _adaptation_cycles; }
154 void set_adaptation_cycles(int v) { _adaptation_cycles = v; }
155
156 unsigned samples_count() { return _samples_count; }
157 void increment_samples_count() { ++_samples_count; }
158
159 size_t _base_space_size;
160 void set_base_space_size(size_t v) { _base_space_size = v; }
161 size_t base_space_size() const { return _base_space_size; }
162
163 // Check if the NUMA topology has changed. Add and remove spaces if needed.
164 // The update can be forced by setting the force parameter equal to true.
165 bool update_layout(bool force);
166 // Bias region towards the lgrp.
167 void bias_region(MemRegion mr, int lgrp_id);
168 // Free pages in a given region.
169 void free_region(MemRegion mr);
170 // Get current chunk size.
171 size_t current_chunk_size(int i);
172 // Get default chunk size (equally divide the space).
173 size_t default_chunk_size();
174 // Adapt the chunk size to follow the allocation rate.
175 size_t adaptive_chunk_size(int i, size_t limit);
176 // Scan and free invalid pages.
177 void scan_pages(size_t page_count);
178 // Return the bottom_region and the top_region. Align them to page_size() boundary.
179 // |------------------new_region---------------------------------|
180 // |----bottom_region--|---intersection---|------top_region------|
181 void select_tails(MemRegion new_region, MemRegion intersection,
182 MemRegion* bottom_region, MemRegion *top_region);
183 // Try to merge the invalid region with the bottom or top region by decreasing
184 // the intersection area. Return the invalid_region aligned to the page_size()
185 // boundary if it's inside the intersection. Return non-empty invalid_region
186 // if it lies inside the intersection (also page-aligned).
187 // |------------------new_region---------------------------------|
188 // |----------------|-------invalid---|--------------------------|
189 // |----bottom_region--|---intersection---|------top_region------|
190 void merge_regions(MemRegion new_region, MemRegion* intersection,
191 MemRegion *invalid_region);
192
193 public:
194 GrowableArray<LGRPSpace*>* lgrp_spaces() const { return _lgrp_spaces; }
195 MutableNUMASpace(size_t alignment);
196 virtual ~MutableNUMASpace();
197 // Space initialization.
198 virtual void initialize(MemRegion mr, bool clear_space, bool mangle_space, bool setup_pages = SetupPages);
199 // Update space layout if necessary. Do all adaptive resizing job.
200 virtual void update();
201 // Update allocation rate averages.
202 virtual void accumulate_statistics();
203
204 virtual void clear(bool mangle_space);
205 virtual void mangle_unused_area() PRODUCT_RETURN;
206 virtual void mangle_unused_area_complete() PRODUCT_RETURN;
207 virtual void mangle_region(MemRegion mr) PRODUCT_RETURN;
208 virtual void check_mangled_unused_area(HeapWord* limit) PRODUCT_RETURN;
209 virtual void check_mangled_unused_area_complete() PRODUCT_RETURN;
210 virtual void set_top_for_allocations(HeapWord* v) PRODUCT_RETURN;
211 virtual void set_top_for_allocations() PRODUCT_RETURN;
212
213 virtual void ensure_parsability();
214 virtual size_t used_in_words() const;
215 virtual size_t free_in_words() const;
216
217 using MutableSpace::capacity_in_words;
218 virtual size_t capacity_in_words(Thread* thr) const;
219 virtual size_t tlab_capacity(Thread* thr) const;
220 virtual size_t tlab_used(Thread* thr) const;
221 virtual size_t unsafe_max_tlab_alloc(Thread* thr) const;
222
223 // Allocation (return NULL if full)
224 virtual HeapWord* allocate(size_t word_size);
225 virtual HeapWord* cas_allocate(size_t word_size);
226
227 // Debugging
228 virtual void print_on(outputStream* st) const;
229 virtual void print_short_on(outputStream* st) const;
230 virtual void verify();
231
232 virtual void set_top(HeapWord* value);
233};
234
235#endif // SHARE_GC_PARALLEL_MUTABLENUMASPACE_HPP
236