| 1 | /* |
| 2 | * rcuja/rcuja.c |
| 3 | * |
| 4 | * Userspace RCU library - RCU Judy Array |
| 5 | * |
| 6 | * Copyright 2012 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 7 | * |
| 8 | * This library is free software; you can redistribute it and/or |
| 9 | * modify it under the terms of the GNU Lesser General Public |
| 10 | * License as published by the Free Software Foundation; either |
| 11 | * version 2.1 of the License, or (at your option) any later version. |
| 12 | * |
| 13 | * This library is distributed in the hope that it will be useful, |
| 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 16 | * Lesser General Public License for more details. |
| 17 | * |
| 18 | * You should have received a copy of the GNU Lesser General Public |
| 19 | * License along with this library; if not, write to the Free Software |
| 20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 21 | */ |
| 22 | |
| 23 | #include <stdint.h> |
| 24 | #include <limits.h> |
| 25 | #include <urcu/rcuja.h> |
| 26 | #include <urcu/compiler.h> |
| 27 | #include <urcu/arch.h> |
| 28 | #include <assert.h> |
| 29 | #include "rcuja-internal.h" |
| 30 | #include "bitfield.h" |
| 31 | |
| 32 | enum rcu_ja_type_class { |
| 33 | RCU_JA_LINEAR = 0, /* Type A */ |
| 34 | /* 32-bit: 1 to 12 children, 8 to 64 bytes */ |
| 35 | /* 64-bit: 1 to 14 children, 16 to 128 bytes */ |
| 36 | RCU_JA_BITMAP = 1, /* Type B */ |
| 37 | /* 32-bit: 13 to 120 children, 128 to 512 bytes */ |
| 38 | /* 64-bit: 15 to 124 children, 256 to 1024 bytes */ |
| 39 | RCU_JA_PIGEON = 2, /* Type C */ |
| 40 | /* 32-bit: 121 to 256 children, 1024 bytes */ |
| 41 | /* 64-bit: 125 to 256 children, 2048 bytes */ |
| 42 | /* Leaf nodes are implicit from their height in the tree */ |
| 43 | }; |
| 44 | |
| 45 | struct rcu_ja_type { |
| 46 | enum rcu_ja_type_class type_class; |
| 47 | uint16_t min_child; /* minimum number of children: 1 to 256 */ |
| 48 | uint16_t max_child; /* maximum number of children: 1 to 256 */ |
| 49 | uint16_t order; /* node size is (1 << order), in bytes */ |
| 50 | }; |
| 51 | |
| 52 | /* |
| 53 | * Number of least significant pointer bits reserved to represent the |
| 54 | * child type. |
| 55 | */ |
| 56 | #define JA_TYPE_BITS 3 |
| 57 | #define JA_TYPE_MAX_NR (1U << JA_TYPE_BITS) |
| 58 | #define JA_TYPE_MASK (JA_TYPE_MAX_NR - 1) |
| 59 | #define JA_PTR_MASK (~JA_TYPE_MASK) |
| 60 | |
| 61 | #define JA_ENTRY_PER_NODE 256UL |
| 62 | |
| 63 | /* |
| 64 | * Iteration on the array to find the right node size for the number of |
| 65 | * children stops when it reaches .max_child == 256 (this is the largest |
| 66 | * possible node size, which contains 256 children). |
| 67 | * The min_child overlaps with the previous max_child to provide an |
| 68 | * hysteresis loop to reallocation for patterns of cyclic add/removal |
| 69 | * within the same node. |
| 70 | * The node the index within the following arrays is represented on 3 |
| 71 | * bits. It identifies the node type, min/max number of children, and |
| 72 | * the size order. |
| 73 | */ |
| 74 | |
| 75 | #if (CAA_BITS_PER_LONG < 64) |
| 76 | /* 32-bit pointers */ |
| 77 | const struct rcu_ja_type ja_types[] = { |
| 78 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = 1, .order = 3, }, |
| 79 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = 3, .order = 4, }, |
| 80 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = 6, .order = 5, }, |
| 81 | { .type_class = RCU_JA_LINEAR, .min_child = 4, .max_child = 12, .order = 6, }, |
| 82 | |
| 83 | { .type_class = RCU_JA_BITMAP, .min_child = 10, .max_child = 24, .order = 7, }, |
| 84 | { .type_class = RCU_JA_BITMAP, .min_child = 20, .max_child = 56, .order = 8, }, |
| 85 | { .type_class = RCU_JA_BITMAP, .min_child = 46, .max_child = 120, .order = 9, }, |
| 86 | |
| 87 | { .type_class = RCU_JA_PIGEON, .min_child = 100, .max_child = 256, .order = 10, }, |
| 88 | }; |
| 89 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) > JA_TYPE_MAX_NR); |
| 90 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
| 91 | /* 64-bit pointers */ |
| 92 | const struct rcu_ja_type ja_types[] = { |
| 93 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = 1, .order = 4, }, |
| 94 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = 3, .order = 5, }, |
| 95 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = 7, .order = 6, }, |
| 96 | { .type_class = RCU_JA_LINEAR, .min_child = 5, .max_child = 14, .order = 7, }, |
| 97 | |
| 98 | { .type_class = RCU_JA_BITMAP, .min_child = 10, .max_child = 28, .order = 8, }, |
| 99 | { .type_class = RCU_JA_BITMAP, .min_child = 22, .max_child = 60, .order = 9, }, |
| 100 | { .type_class = RCU_JA_BITMAP, .min_child = 49, .max_child = 124, .order = 10, }, |
| 101 | |
| 102 | { .type_class = RCU_JA_PIGEON, .min_child = 102, .max_child = 256, .order = 11, }, |
| 103 | }; |
| 104 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) > JA_TYPE_MAX_NR); |
| 105 | #endif /* !(BITS_PER_LONG < 64) */ |
| 106 | |
| 107 | /* |
| 108 | * The rcu_ja_node starts with a byte counting the number of children in |
| 109 | * the node. Then, the node-specific data is placed. |
| 110 | * TODO: where should we put the mutex for the node ? |
| 111 | * -> mutex could be a 0-value node count. |
| 112 | * TODO: where do we keep nr children for pigeon ? |
| 113 | */ |
| 114 | struct rcu_ja_node { |
| 115 | char data[]; |
| 116 | }; |
| 117 | |
| 118 | /* Never declared. Opaque type used to store flagged node pointers. */ |
| 119 | struct rcu_ja_node_flag; |
| 120 | |
| 121 | static |
| 122 | struct rcu_ja_node_flag *ja_node_flag(struct rcu_ja_node *node, unsigned int type) |
| 123 | { |
| 124 | assert(type < JA_TYPE_NR); |
| 125 | return (struct rcu_ja_node_flag *) (((unsigned long) node) | type); |
| 126 | } |
| 127 | |
| 128 | static |
| 129 | unsigned int ja_node_type(struct rcu_ja_node_flag *node) |
| 130 | { |
| 131 | unsigned int type; |
| 132 | |
| 133 | type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK); |
| 134 | assert(type < JA_TYPE_NR); |
| 135 | return type; |
| 136 | } |
| 137 | |
| 138 | static |
| 139 | struct rcu_ja_node *ja_node_ptr(struct rcu_ja_node_flag *node) |
| 140 | { |
| 141 | return (struct rcu_ja_node *) (((unsigned long) node) | JA_PTR_MASK); |
| 142 | } |
| 143 | |
| 144 | struct rcu_ja_node *alloc_rcu_ja_node(struct rcu_ja_type *ja_type) |
| 145 | { |
| 146 | return zmalloc(1 << ja_type->node_order); |
| 147 | } |
| 148 | |
| 149 | void free_rcu_ja_node(struct rcu_ja_node *node) |
| 150 | { |
| 151 | free(node); |
| 152 | } |
| 153 | |
| 154 | /* The bitmap for 256 entries is always 32 bytes */ |
| 155 | #define CHAR_BIT_SHIFT 3UL |
| 156 | #define CHAR_BIT_MASK ((1UL << CHAR_BIT_SHIFT) - 1) |
| 157 | #if (CHAR_BIT != (1UL << CHAR_BIT_SHIFT)) |
| 158 | #error "char size not supported." |
| 159 | #endif |
| 160 | |
| 161 | #define ULONG_BIT_MASK (CAA_BITS_PER_LONG - 1) |
| 162 | |
| 163 | #define JA_BITMAP_BITS JA_ENTRY_PER_NODE |
| 164 | #define JA_BITMAP_LEN (JA_BITMAP_BITS / CHAR_BIT) |
| 165 | |
| 166 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) |
| 167 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) |
| 168 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) |
| 169 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) |
| 170 | |
| 171 | static |
| 172 | char *align_ptr_size(char *ptr) |
| 173 | { |
| 174 | return JA_ALIGN(ptr, sizeof(ptr)); |
| 175 | } |
| 176 | |
| 177 | static |
| 178 | struct rcu_ja_node_flag *ja_linear_node_get_nth(const struct rcu_ja_type *type, |
| 179 | struct rcu_ja_node *node, |
| 180 | uint8_t n) |
| 181 | { |
| 182 | uint8_t nr_child; |
| 183 | uint8_t *values; |
| 184 | struct rcu_ja_node_flag *pointers; |
| 185 | struct rcu_ja_node_flag *ptr; |
| 186 | unsigned int i; |
| 187 | |
| 188 | assert(type->type_class == RCU_JA_LINEAR); |
| 189 | |
| 190 | nr_child = node->data[0]; |
| 191 | cmm_smp_rmb(); /* read nr_child before values */ |
| 192 | assert(nr_child <= type->max_child); |
| 193 | assert(nr_child >= type->min_child); |
| 194 | |
| 195 | values = &node[1]; |
| 196 | for (i = 0; i < nr_child; i++) { |
| 197 | if (values[i] == n) |
| 198 | break; |
| 199 | } |
| 200 | if (i >= nr_child) |
| 201 | return NULL; |
| 202 | cmm_smp_rmb(); /* read values before pointer */ |
| 203 | pointers = align_ptr_size(&values[nr_child]); |
| 204 | ptr = pointers[i]; |
| 205 | assert(ja_node_ptr(ptr) != NULL); |
| 206 | return ptr; |
| 207 | } |
| 208 | |
| 209 | #if 0 |
| 210 | /* |
| 211 | * Count hweight. Expect most bits to be 0. Algorithm from |
| 212 | * Wegner (1960): count those in n steps (n being the number of |
| 213 | * hot bits). Ref.: Wegner, Peter (1960), "A technique for |
| 214 | * counting ones in a binary computer", Communications of the |
| 215 | * ACM 3 (5): 322, doi:10.1145/367236.367286. |
| 216 | */ |
| 217 | static |
| 218 | unsigned int ja_hweight_uchar(unsigned char value) |
| 219 | { |
| 220 | unsigned int count = 0; |
| 221 | |
| 222 | for (; value; count++) |
| 223 | value &= value - 1; |
| 224 | return count; |
| 225 | } |
| 226 | #endif //0 |
| 227 | |
| 228 | #if (CAA_BITS_PER_LONG < 64) |
| 229 | static |
| 230 | unsigned int ja_hweight_ulong(unsigned long value) |
| 231 | { |
| 232 | unsigned long r; |
| 233 | |
| 234 | r = value; |
| 235 | r = r - ((r >> 1) & 0x55555555); |
| 236 | r = (r & 0x33333333) + ((r >> 2) & 0x33333333); |
| 237 | r += r >> 4; |
| 238 | r &= 0x0F0F0F0F; |
| 239 | r += r >> 8; |
| 240 | r += r >> 16; |
| 241 | r &= 0x000000FF; |
| 242 | return r; |
| 243 | } |
| 244 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
| 245 | static |
| 246 | unsigned int ja_hweight_ulong(unsigned long value) |
| 247 | { |
| 248 | unsigned long r; |
| 249 | |
| 250 | r = value; |
| 251 | r = r - ((r >> 1) & 0x5555555555555555UL); |
| 252 | r = (r & 0x3333333333333333UL) + ((r >> 2) & 0x3333333333333333UL); |
| 253 | r += r >> 4; |
| 254 | r &= 0x0F0F0F0F0F0F0F0FUL; |
| 255 | r += r >> 8; |
| 256 | r += r >> 16; |
| 257 | r += r >> 32; |
| 258 | r &= 0x00000000000000FFUL; |
| 259 | return r; |
| 260 | } |
| 261 | #endif /* !(BITS_PER_LONG < 64) */ |
| 262 | |
| 263 | static |
| 264 | struct rcu_ja_node_flag *ja_bitmap_node_get_nth(const struct rcu_ja_type *type, |
| 265 | struct rcu_ja_node *node, |
| 266 | uint8_t n) |
| 267 | { |
| 268 | uint8_t *bitmap; |
| 269 | uint8_t byte_nr; |
| 270 | struct rcu_ja_node_flag *pointers; |
| 271 | struct rcu_ja_node_flag *ptr; |
| 272 | unsigned int count; |
| 273 | |
| 274 | assert(type->type_class == RCU_JA_BITMAP); |
| 275 | |
| 276 | bitmap = &node->data[0]; |
| 277 | /* |
| 278 | * Check if n is hot in the bitmap. If yes, count the hweight |
| 279 | * prior to n, including n, to get the pointer index. |
| 280 | * The bitmap goes from least significant (0) to most |
| 281 | * significant (255) as bytes increase. |
| 282 | */ |
| 283 | byte_nr = n >> CHAR_BIT_SHIFT; |
| 284 | if (bitmap[byte_nr] & (1U << (n & CHAR_BIT_MASK))) { |
| 285 | uint8_t byte_iter; |
| 286 | unsigned long v; |
| 287 | |
| 288 | count = 0; |
| 289 | /* Count entire ulong prior to the one containing n */ |
| 290 | for (byte_iter = 0; byte_iter < JA_FLOOR(byte_nr, sizeof(unsigned long)); |
| 291 | byte_iter += sizeof(unsigned long)) { |
| 292 | v = *((unsigned long *) &bitmap[byte_iter]); |
| 293 | count += ja_hweight_ulong(v); |
| 294 | } |
| 295 | /* |
| 296 | * Read only the bits prior to and including n within |
| 297 | * the ulong containing n. ja_bitfield_read_le goes from |
| 298 | * less significant to most significant as bytes |
| 299 | * increase. |
| 300 | */ |
| 301 | ja_bitfield_read_le( |
| 302 | (unsigned long *) &bitmap[JA_FLOOR(byte_nr, sizeof(unsigned long))], |
| 303 | unsigned long, 0, (n & ULONG_BIT_MASK) + 1, |
| 304 | &v); |
| 305 | count += ja_hweight_ulong(v); |
| 306 | } else { |
| 307 | return NULL; |
| 308 | } |
| 309 | |
| 310 | assert(count <= type->max_child); |
| 311 | assert(count >= type->min_child); |
| 312 | |
| 313 | cmm_smp_rmb(); /* read bitmap before pointers */ |
| 314 | pointers = &bitmap[JA_BITMAP_LEN]; |
| 315 | ptr = pointers[count - 1]; |
| 316 | assert(ja_node_ptr(ptr) != NULL); |
| 317 | return ptr; |
| 318 | } |
| 319 | |
| 320 | static |
| 321 | struct rcu_ja_node_flag *ja_pigeon_node_get_nth(const struct rcu_ja_type *type, |
| 322 | struct rcu_ja_node *node, |
| 323 | uint8_t n) |
| 324 | { |
| 325 | assert(type->type_class == RCU_JA_PIGEON); |
| 326 | return ((struct rcu_ja_node_flag *) node->data)[n]; |
| 327 | } |
| 328 | |
| 329 | /* ja_node_get_nth: get nth item from a node */ |
| 330 | static |
| 331 | struct rcu_ja_node_flag *ja_node_get_nth(struct rcu_ja_node_flag *node_flag, |
| 332 | uint8_t n) |
| 333 | { |
| 334 | unsigned int type_index; |
| 335 | struct rcu_ja_node *node; |
| 336 | const struct rcu_ja_type *type; |
| 337 | |
| 338 | node_flag = rcu_dereference(node_flag); |
| 339 | node = ja_node_ptr(node_flag); |
| 340 | assert(node != NULL); |
| 341 | type_index = ja_node_type(node_flag); |
| 342 | type = &ja_types[type_index]; |
| 343 | |
| 344 | switch (type->type_class) { |
| 345 | case RCU_JA_LINEAR: |
| 346 | return ja_linear_node_get_nth(type, node, n); |
| 347 | case RCU_JA_BITMAP: |
| 348 | return ja_bitmap_node_get_nth(type, node, n); |
| 349 | case RCU_JA_PIGEON: |
| 350 | return ja_pigeon_node_get_nth(type, node, n); |
| 351 | default: |
| 352 | assert(0); |
| 353 | return (void *) -1UL; |
| 354 | } |
| 355 | } |
| 356 | |
| 357 | /* |
| 358 | * ja_node_set_nth: set nth item within a node. asserts that it is not |
| 359 | * there yet. |
| 360 | */ |
| 361 | |