| OLD | NEW |
| 1 // Copyright (c) 2008, Google Inc. | 1 // Copyright (c) 2008, Google Inc. |
| 2 // All rights reserved. | 2 // All rights reserved. |
| 3 // | 3 // |
| 4 // Redistribution and use in source and binary forms, with or without | 4 // Redistribution and use in source and binary forms, with or without |
| 5 // modification, are permitted provided that the following conditions are | 5 // modification, are permitted provided that the following conditions are |
| 6 // met: | 6 // met: |
| 7 // | 7 // |
| 8 // * Redistributions of source code must retain the above copyright | 8 // * Redistributions of source code must retain the above copyright |
| 9 // notice, this list of conditions and the following disclaimer. | 9 // notice, this list of conditions and the following disclaimer. |
| 10 // * Redistributions in binary form must reproduce the above | 10 // * Redistributions in binary form must reproduce the above |
| (...skipping 16 matching lines...) Expand all Loading... |
| 27 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE | 27 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
| 28 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | 28 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 29 | 29 |
| 30 // --- | 30 // --- |
| 31 // Author: Sanjay Ghemawat <opensource@google.com> | 31 // Author: Sanjay Ghemawat <opensource@google.com> |
| 32 | 32 |
| 33 #include "config.h" | 33 #include "config.h" |
| 34 #include "common.h" | 34 #include "common.h" |
| 35 #include "system-alloc.h" | 35 #include "system-alloc.h" |
| 36 | 36 |
| 37 #if defined(HAVE_UNISTD_H) && defined(HAVE_GETPAGESIZE) | |
| 38 #include <unistd.h> // for getpagesize | |
| 39 #endif | |
| 40 | |
| 41 namespace tcmalloc { | 37 namespace tcmalloc { |
| 42 | 38 |
| 43 // Note: the following only works for "n"s that fit in 32-bits, but | 39 // Note: the following only works for "n"s that fit in 32-bits, but |
| 44 // that is fine since we only use it for small sizes. | 40 // that is fine since we only use it for small sizes. |
| 45 static inline int LgFloor(size_t n) { | 41 static inline int LgFloor(size_t n) { |
| 46 int log = 0; | 42 int log = 0; |
| 47 for (int i = 4; i >= 0; --i) { | 43 for (int i = 4; i >= 0; --i) { |
| 48 int shift = (1 << i); | 44 int shift = (1 << i); |
| 49 size_t x = n >> shift; | 45 size_t x = n >> shift; |
| 50 if (x != 0) { | 46 if (x != 0) { |
| 51 n = x; | 47 n = x; |
| 52 log += shift; | 48 log += shift; |
| 53 } | 49 } |
| 54 } | 50 } |
| 55 ASSERT(n == 1); | 51 ASSERT(n == 1); |
| 56 return log; | 52 return log; |
| 57 } | 53 } |
| 58 | 54 |
| 59 int AlignmentForSize(size_t size) { | 55 int AlignmentForSize(size_t size) { |
| 60 int alignment = kAlignment; | 56 int alignment = kAlignment; |
| 61 if (size >= 2048) { | 57 if (size > kMaxSize) { |
| 62 // Cap alignment at 256 for large sizes. | 58 // Cap alignment at kPageSize for large sizes. |
| 63 alignment = 256; | 59 alignment = kPageSize; |
| 64 } else if (size >= 128) { | 60 } else if (size >= 128) { |
| 65 // Space wasted due to alignment is at most 1/8, i.e., 12.5%. | 61 // Space wasted due to alignment is at most 1/8, i.e., 12.5%. |
| 66 alignment = (1 << LgFloor(size)) / 8; | 62 alignment = (1 << LgFloor(size)) / 8; |
| 67 } else if (size >= 16) { | 63 } else if (size >= 16) { |
| 68 // We need an alignment of at least 16 bytes to satisfy | 64 // We need an alignment of at least 16 bytes to satisfy |
| 69 // requirements for some SSE types. | 65 // requirements for some SSE types. |
| 70 alignment = 16; | 66 alignment = 16; |
| 71 } | 67 } |
| 68 // Maximum alignment allowed is page size alignment. |
| 69 if (alignment > kPageSize) { |
| 70 alignment = kPageSize; |
| 71 } |
| 72 CHECK_CONDITION(size < 16 || alignment >= 16); | 72 CHECK_CONDITION(size < 16 || alignment >= 16); |
| 73 CHECK_CONDITION((alignment & (alignment - 1)) == 0); | 73 CHECK_CONDITION((alignment & (alignment - 1)) == 0); |
| 74 return alignment; | 74 return alignment; |
| 75 } | 75 } |
| 76 | 76 |
| 77 int SizeMap::NumMoveSize(size_t size) { | 77 int SizeMap::NumMoveSize(size_t size) { |
| 78 if (size == 0) return 0; | 78 if (size == 0) return 0; |
| 79 // Use approx 64k transfers between thread and central caches. | 79 // Use approx 64k transfers between thread and central caches. |
| 80 int num = static_cast<int>(64.0 * 1024.0 / size); | 80 int num = static_cast<int>(64.0 * 1024.0 / size); |
| 81 if (num < 2) num = 2; | 81 if (num < 2) num = 2; |
| (...skipping 10 matching lines...) Expand all Loading... |
| 92 // This value strikes a balance between the constraints above. | 92 // This value strikes a balance between the constraints above. |
| 93 if (num > 32) num = 32; | 93 if (num > 32) num = 32; |
| 94 | 94 |
| 95 return num; | 95 return num; |
| 96 } | 96 } |
| 97 | 97 |
| 98 // Initialize the mapping arrays | 98 // Initialize the mapping arrays |
| 99 void SizeMap::Init() { | 99 void SizeMap::Init() { |
| 100 // Do some sanity checking on add_amount[]/shift_amount[]/class_array[] | 100 // Do some sanity checking on add_amount[]/shift_amount[]/class_array[] |
| 101 if (ClassIndex(0) < 0) { | 101 if (ClassIndex(0) < 0) { |
| 102 CRASH("Invalid class index %d for size 0\n", ClassIndex(0)); | 102 Log(kCrash, __FILE__, __LINE__, |
| 103 "Invalid class index for size 0", ClassIndex(0)); |
| 103 } | 104 } |
| 104 if (ClassIndex(kMaxSize) >= sizeof(class_array_)) { | 105 if (ClassIndex(kMaxSize) >= sizeof(class_array_)) { |
| 105 CRASH("Invalid class index %d for kMaxSize\n", ClassIndex(kMaxSize)); | 106 Log(kCrash, __FILE__, __LINE__, |
| 107 "Invalid class index for kMaxSize", ClassIndex(kMaxSize)); |
| 106 } | 108 } |
| 107 | 109 |
| 108 // Compute the size classes we want to use | 110 // Compute the size classes we want to use |
| 109 int sc = 1; // Next size class to assign | 111 int sc = 1; // Next size class to assign |
| 110 int alignment = kAlignment; | 112 int alignment = kAlignment; |
| 111 CHECK_CONDITION(kAlignment <= 16); | 113 CHECK_CONDITION(kAlignment <= 16); |
| 112 int last_lg = -1; | 114 for (size_t size = kAlignment; size <= kMaxSize; size += alignment) { |
| 113 for (size_t size = kMinClassSize; size <= kMaxSize; size += alignment) { | 115 alignment = AlignmentForSize(size); |
| 114 int lg = LgFloor(size); | |
| 115 if (lg > last_lg) { | |
| 116 // Increase alignment every so often to reduce number of size classes. | |
| 117 alignment = AlignmentForSize(size); | |
| 118 last_lg = lg; | |
| 119 } | |
| 120 CHECK_CONDITION((size % alignment) == 0); | 116 CHECK_CONDITION((size % alignment) == 0); |
| 121 | 117 |
| 122 // Allocate enough pages so leftover is less than 1/8 of total. | 118 int blocks_to_move = NumMoveSize(size) / 4; |
| 123 // This bounds wasted space to at most 12.5%. | 119 size_t psize = 0; |
| 124 size_t psize = kPageSize; | 120 do { |
| 125 while ((psize % size) > (psize >> 3)) { | |
| 126 psize += kPageSize; | 121 psize += kPageSize; |
| 127 } | 122 // Allocate enough pages so leftover is less than 1/8 of total. |
| 123 // This bounds wasted space to at most 12.5%. |
| 124 while ((psize % size) > (psize >> 3)) { |
| 125 psize += kPageSize; |
| 126 } |
| 127 // Continue to add pages until there are at least as many objects in |
| 128 // the span as are needed when moving objects from the central |
| 129 // freelists and spans to the thread caches. |
| 130 } while ((psize / size) < (blocks_to_move)); |
| 128 const size_t my_pages = psize >> kPageShift; | 131 const size_t my_pages = psize >> kPageShift; |
| 129 | 132 |
| 130 if (sc > 1 && my_pages == class_to_pages_[sc-1]) { | 133 if (sc > 1 && my_pages == class_to_pages_[sc-1]) { |
| 131 // See if we can merge this into the previous class without | 134 // See if we can merge this into the previous class without |
| 132 // increasing the fragmentation of the previous class. | 135 // increasing the fragmentation of the previous class. |
| 133 const size_t my_objects = (my_pages << kPageShift) / size; | 136 const size_t my_objects = (my_pages << kPageShift) / size; |
| 134 const size_t prev_objects = (class_to_pages_[sc-1] << kPageShift) | 137 const size_t prev_objects = (class_to_pages_[sc-1] << kPageShift) |
| 135 / class_to_size_[sc-1]; | 138 / class_to_size_[sc-1]; |
| 136 if (my_objects == prev_objects) { | 139 if (my_objects == prev_objects) { |
| 137 // Adjust last class to include this size | 140 // Adjust last class to include this size |
| 138 class_to_size_[sc-1] = size; | 141 class_to_size_[sc-1] = size; |
| 139 continue; | 142 continue; |
| 140 } | 143 } |
| 141 } | 144 } |
| 142 | 145 |
| 143 // Add new class | 146 // Add new class |
| 144 class_to_pages_[sc] = my_pages; | 147 class_to_pages_[sc] = my_pages; |
| 145 class_to_size_[sc] = size; | 148 class_to_size_[sc] = size; |
| 146 sc++; | 149 sc++; |
| 147 } | 150 } |
| 148 if (sc != kNumClasses) { | 151 if (sc != kNumClasses) { |
| 149 CRASH("wrong number of size classes: found %d instead of %d\n", | 152 Log(kCrash, __FILE__, __LINE__, |
| 150 sc, int(kNumClasses)); | 153 "wrong number of size classes: (found vs. expected )", sc, kNumClasses); |
| 151 } | 154 } |
| 152 | 155 |
| 153 // Initialize the mapping arrays | 156 // Initialize the mapping arrays |
| 154 int next_size = 0; | 157 int next_size = 0; |
| 155 for (int c = 1; c < kNumClasses; c++) { | 158 for (int c = 1; c < kNumClasses; c++) { |
| 156 const int max_size_in_class = class_to_size_[c]; | 159 const int max_size_in_class = class_to_size_[c]; |
| 157 for (int s = next_size; s <= max_size_in_class; s += kAlignment) { | 160 for (int s = next_size; s <= max_size_in_class; s += kAlignment) { |
| 158 class_array_[ClassIndex(s)] = c; | 161 class_array_[ClassIndex(s)] = c; |
| 159 } | 162 } |
| 160 next_size = max_size_in_class + kAlignment; | 163 next_size = max_size_in_class + kAlignment; |
| 161 } | 164 } |
| 162 | 165 |
| 163 // Double-check sizes just to be safe | 166 // Double-check sizes just to be safe |
| 164 for (size_t size = 0; size <= kMaxSize; size++) { | 167 for (size_t size = 0; size <= kMaxSize; size++) { |
| 165 const int sc = SizeClass(size); | 168 const int sc = SizeClass(size); |
| 166 if (sc <= 0 || sc >= kNumClasses) { | 169 if (sc <= 0 || sc >= kNumClasses) { |
| 167 CRASH("Bad size class %d for %" PRIuS "\n", sc, size); | 170 Log(kCrash, __FILE__, __LINE__, |
| 171 "Bad size class (class, size)", sc, size); |
| 168 } | 172 } |
| 169 if (sc > 1 && size <= class_to_size_[sc-1]) { | 173 if (sc > 1 && size <= class_to_size_[sc-1]) { |
| 170 CRASH("Allocating unnecessarily large class %d for %" PRIuS | 174 Log(kCrash, __FILE__, __LINE__, |
| 171 "\n", sc, size); | 175 "Allocating unnecessarily large class (class, size)", sc, size); |
| 172 } | 176 } |
| 173 const size_t s = class_to_size_[sc]; | 177 const size_t s = class_to_size_[sc]; |
| 174 if (size > s) { | 178 if (size > s || s == 0) { |
| 175 CRASH("Bad size %" PRIuS " for %" PRIuS " (sc = %d)\n", s, size, sc); | 179 Log(kCrash, __FILE__, __LINE__, |
| 176 } | 180 "Bad (class, size, requested)", sc, s, size); |
| 177 if (s == 0) { | |
| 178 CRASH("Bad size %" PRIuS " for %" PRIuS " (sc = %d)\n", s, size, sc); | |
| 179 } | 181 } |
| 180 } | 182 } |
| 181 | 183 |
| 182 // Initialize the num_objects_to_move array. | 184 // Initialize the num_objects_to_move array. |
| 183 for (size_t cl = 1; cl < kNumClasses; ++cl) { | 185 for (size_t cl = 1; cl < kNumClasses; ++cl) { |
| 184 num_objects_to_move_[cl] = NumMoveSize(ByteSizeForClass(cl)); | 186 num_objects_to_move_[cl] = NumMoveSize(ByteSizeForClass(cl)); |
| 185 } | 187 } |
| 186 } | 188 } |
| 187 | 189 |
| 188 void SizeMap::Dump(TCMalloc_Printer* out) { | |
| 189 // Dump class sizes and maximum external wastage per size class | |
| 190 for (size_t cl = 1; cl < kNumClasses; ++cl) { | |
| 191 const int alloc_size = class_to_pages_[cl] << kPageShift; | |
| 192 const int alloc_objs = alloc_size / class_to_size_[cl]; | |
| 193 const int min_used = (class_to_size_[cl-1] + 1) * alloc_objs; | |
| 194 const int max_waste = alloc_size - min_used; | |
| 195 out->printf("SC %3d [ %8d .. %8d ] from %8d ; %2.0f%% maxwaste\n", | |
| 196 int(cl), | |
| 197 int(class_to_size_[cl-1] + 1), | |
| 198 int(class_to_size_[cl]), | |
| 199 int(class_to_pages_[cl] << kPageShift), | |
| 200 max_waste * 100.0 / alloc_size | |
| 201 ); | |
| 202 } | |
| 203 } | |
| 204 | |
| 205 // Metadata allocator -- keeps stats about how many bytes allocated. | 190 // Metadata allocator -- keeps stats about how many bytes allocated. |
| 206 static uint64_t metadata_system_bytes_ = 0; | 191 static uint64_t metadata_system_bytes_ = 0; |
| 207 void* MetaDataAlloc(size_t bytes) { | 192 void* MetaDataAlloc(size_t bytes) { |
| 208 static size_t pagesize; | 193 void* result = TCMalloc_SystemAlloc(bytes, NULL); |
| 209 #ifdef HAVE_GETPAGESIZE | |
| 210 if (pagesize == 0) | |
| 211 pagesize = getpagesize(); | |
| 212 #endif | |
| 213 | |
| 214 void* result = TCMalloc_SystemAlloc(bytes, NULL, pagesize); | |
| 215 if (result != NULL) { | 194 if (result != NULL) { |
| 216 metadata_system_bytes_ += bytes; | 195 metadata_system_bytes_ += bytes; |
| 217 } | 196 } |
| 218 return result; | 197 return result; |
| 219 } | 198 } |
| 220 | 199 |
| 221 uint64_t metadata_system_bytes() { return metadata_system_bytes_; } | 200 uint64_t metadata_system_bytes() { return metadata_system_bytes_; } |
| 222 | 201 |
| 223 void increment_metadata_system_bytes(size_t bytes) { | |
| 224 metadata_system_bytes_ += bytes; | |
| 225 } | |
| 226 | |
| 227 } // namespace tcmalloc | 202 } // namespace tcmalloc |
| OLD | NEW |