1 /* 2 * Copyright (c) 2006, 2009, 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 /* 26 * The NUMA-aware allocator (MutableNUMASpace) is basically a modification 27 * of MutableSpace which preserves interfaces but implements different 28 * functionality. The space is split into chunks for each locality group 29 * (resizing for adaptive size policy is also supported). For each thread 30 * allocations are performed in the chunk corresponding to the home locality 31 * group of the thread. Whenever any chunk fills-in the young generation 32 * collection occurs. 33 * The chunks can be also be adaptively resized. The idea behind the adaptive 34 * sizing is to reduce the loss of the space in the eden due to fragmentation. 35 * The main cause of fragmentation is uneven allocation rates of threads. 36 * The allocation rate difference between locality groups may be caused either by 37 * application specifics or by uneven LWP distribution by the OS. Besides, 38 * application can have less threads then the number of locality groups. 39 * In order to resize the chunk we measure the allocation rate of the 40 * application between collections. After that we reshape the chunks to reflect 41 * the allocation rate pattern. The AdaptiveWeightedAverage exponentially 42 * decaying average is used to smooth the measurements. The NUMASpaceResizeRate 43 * parameter is used to control the adaptation speed by restricting the number of 44 * bytes that can be moved during the adaptation phase. 45 * Chunks may contain pages from a wrong locality group. The page-scanner has 46 * been introduced to address the problem. Remote pages typically appear due to 47 * the memory shortage in the target locality group. Besides Solaris would 48 * allocate a large page from the remote locality group even if there are small 49 * local pages available. The page-scanner scans the pages right after the 50 * collection and frees remote pages in hope that subsequent reallocation would 51 * be more successful. This approach proved to be useful on systems with high 52 * load where multiple processes are competing for the memory. 53 */ 54 55 class MutableNUMASpace : public MutableSpace { 56 friend class VMStructs; 57 58 class LGRPSpace : public CHeapObj { 59 int _lgrp_id; 60 MutableSpace* _space; 61 MemRegion _invalid_region; 62 AdaptiveWeightedAverage *_alloc_rate; 63 bool _allocation_failed; 64 65 struct SpaceStats { 66 size_t _local_space, _remote_space, _unbiased_space, _uncommited_space; 67 size_t _large_pages, _small_pages; 68 69 SpaceStats() { 70 _local_space = 0; 71 _remote_space = 0; 72 _unbiased_space = 0; 73 _uncommited_space = 0; 74 _large_pages = 0; 75 _small_pages = 0; 76 } 77 }; 78 79 SpaceStats _space_stats; 80 81 char* _last_page_scanned; 82 char* last_page_scanned() { return _last_page_scanned; } 83 void set_last_page_scanned(char* p) { _last_page_scanned = p; } 84 public: 85 LGRPSpace(int l, size_t alignment) : _lgrp_id(l), _last_page_scanned(NULL), _allocation_failed(false) { 86 _space = new MutableSpace(alignment); 87 _alloc_rate = new AdaptiveWeightedAverage(NUMAChunkResizeWeight); 88 } 89 ~LGRPSpace() { 90 delete _space; 91 delete _alloc_rate; 92 } 93 94 void add_invalid_region(MemRegion r) { 95 if (!_invalid_region.is_empty()) { 96 _invalid_region.set_start(MIN2(_invalid_region.start(), r.start())); 97 _invalid_region.set_end(MAX2(_invalid_region.end(), r.end())); 98 } else { 99 _invalid_region = r; 100 } 101 } 102 103 static bool equals(void* lgrp_id_value, LGRPSpace* p) { 104 return *(int*)lgrp_id_value == p->lgrp_id(); 105 } 106 107 // Report a failed allocation. 108 void set_allocation_failed() { _allocation_failed = true; } 109 110 void sample() { 111 // If there was a failed allocation make allocation rate equal 112 // to the size of the whole chunk. This ensures the progress of 113 // the adaptation process. 114 size_t alloc_rate_sample; 115 if (_allocation_failed) { 116 alloc_rate_sample = space()->capacity_in_bytes(); 117 _allocation_failed = false; 118 } else { 119 alloc_rate_sample = space()->used_in_bytes(); 120 } 121 alloc_rate()->sample(alloc_rate_sample); 122 } 123 124 MemRegion invalid_region() const { return _invalid_region; } 125 void set_invalid_region(MemRegion r) { _invalid_region = r; } 126 int lgrp_id() const { return _lgrp_id; } 127 MutableSpace* space() const { return _space; } 128 AdaptiveWeightedAverage* alloc_rate() const { return _alloc_rate; } 129 void clear_alloc_rate() { _alloc_rate->clear(); } 130 SpaceStats* space_stats() { return &_space_stats; } 131 void clear_space_stats() { _space_stats = SpaceStats(); } 132 133 void accumulate_statistics(size_t page_size); 134 void scan_pages(size_t page_size, size_t page_count); 135 }; 136 137 GrowableArray<LGRPSpace*>* _lgrp_spaces; 138 size_t _page_size; 139 unsigned _adaptation_cycles, _samples_count; 140 141 void set_page_size(size_t psz) { _page_size = psz; } 142 size_t page_size() const { return _page_size; } 143 144 unsigned adaptation_cycles() { return _adaptation_cycles; } 145 void set_adaptation_cycles(int v) { _adaptation_cycles = v; } 146 147 unsigned samples_count() { return _samples_count; } 148 void increment_samples_count() { ++_samples_count; } 149 150 size_t _base_space_size; 151 void set_base_space_size(size_t v) { _base_space_size = v; } 152 size_t base_space_size() const { return _base_space_size; } 153 154 // Check if the NUMA topology has changed. Add and remove spaces if needed. 155 // The update can be forced by setting the force parameter equal to true. 156 bool update_layout(bool force); 157 // Bias region towards the lgrp. 158 void bias_region(MemRegion mr, int lgrp_id); 159 // Free pages in a given region. 160 void free_region(MemRegion mr); 161 // Get current chunk size. 162 size_t current_chunk_size(int i); 163 // Get default chunk size (equally divide the space). 164 size_t default_chunk_size(); 165 // Adapt the chunk size to follow the allocation rate. 166 size_t adaptive_chunk_size(int i, size_t limit); 167 // Scan and free invalid pages. 168 void scan_pages(size_t page_count); 169 // Return the bottom_region and the top_region. Align them to page_size() boundary. 170 // |------------------new_region---------------------------------| 171 // |----bottom_region--|---intersection---|------top_region------| 172 void select_tails(MemRegion new_region, MemRegion intersection, 173 MemRegion* bottom_region, MemRegion *top_region); 174 // Try to merge the invalid region with the bottom or top region by decreasing 175 // the intersection area. Return the invalid_region aligned to the page_size() 176 // boundary if it's inside the intersection. Return non-empty invalid_region 177 // if it lies inside the intersection (also page-aligned). 178 // |------------------new_region---------------------------------| 179 // |----------------|-------invalid---|--------------------------| 180 // |----bottom_region--|---intersection---|------top_region------| 181 void merge_regions(MemRegion new_region, MemRegion* intersection, 182 MemRegion *invalid_region); 183 184 public: 185 GrowableArray<LGRPSpace*>* lgrp_spaces() const { return _lgrp_spaces; } 186 MutableNUMASpace(size_t alignment); 187 virtual ~MutableNUMASpace(); 188 // Space initialization. 189 virtual void initialize(MemRegion mr, bool clear_space, bool mangle_space, bool setup_pages = SetupPages); 190 // Update space layout if necessary. Do all adaptive resizing job. 191 virtual void update(); 192 // Update allocation rate averages. 193 virtual void accumulate_statistics(); 194 195 virtual void clear(bool mangle_space); 196 virtual void mangle_unused_area() PRODUCT_RETURN; 197 virtual void mangle_unused_area_complete() PRODUCT_RETURN; 198 virtual void mangle_region(MemRegion mr) PRODUCT_RETURN; 199 virtual void check_mangled_unused_area(HeapWord* limit) PRODUCT_RETURN; 200 virtual void check_mangled_unused_area_complete() PRODUCT_RETURN; 201 virtual void set_top_for_allocations(HeapWord* v) PRODUCT_RETURN; 202 virtual void set_top_for_allocations() PRODUCT_RETURN; 203 204 virtual void ensure_parsability(); 205 virtual size_t used_in_words() const; 206 virtual size_t free_in_words() const; 207 208 using MutableSpace::capacity_in_words; 209 virtual size_t capacity_in_words(Thread* thr) const; 210 virtual size_t tlab_capacity(Thread* thr) const; 211 virtual size_t unsafe_max_tlab_alloc(Thread* thr) const; 212 213 // Allocation (return NULL if full) 214 virtual HeapWord* allocate(size_t word_size); 215 virtual HeapWord* cas_allocate(size_t word_size); 216 217 // Debugging 218 virtual void print_on(outputStream* st) const; 219 virtual void print_short_on(outputStream* st) const; 220 virtual void verify(bool allow_dirty); 221 222 virtual void set_top(HeapWord* value); 223 };