| 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 25 children, 8 to 128 bytes */ |
| 35 | /* 64-bit: 1 to 28 children, 16 to 256 bytes */ |
| 36 | RCU_JA_POOL = 1, /* Type B */ |
| 37 | /* 32-bit: 26 to 100 children, 256 to 512 bytes */ |
| 38 | /* 64-bit: 29 to 112 children, 512 to 1024 bytes */ |
| 39 | RCU_JA_PIGEON = 2, /* Type C */ |
| 40 | /* 32-bit: 101 to 256 children, 1024 bytes */ |
| 41 | /* 64-bit: 113 to 256 children, 2048 bytes */ |
| 42 | /* Leaf nodes are implicit from their height in the tree */ |
| 43 | RCU_JA_NR_TYPES, |
| 44 | }; |
| 45 | |
| 46 | struct rcu_ja_type { |
| 47 | enum rcu_ja_type_class type_class; |
| 48 | uint16_t min_child; /* minimum number of children: 1 to 256 */ |
| 49 | uint16_t max_child; /* maximum number of children: 1 to 256 */ |
| 50 | uint16_t order; /* node size is (1 << order), in bytes */ |
| 51 | uint16_t nr_pool_order; /* number of pools */ |
| 52 | uint16_t pool_size_order; /* pool size */ |
| 53 | }; |
| 54 | |
| 55 | /* |
| 56 | * Number of least significant pointer bits reserved to represent the |
| 57 | * child type. |
| 58 | */ |
| 59 | #define JA_TYPE_BITS 3 |
| 60 | #define JA_TYPE_MAX_NR (1U << JA_TYPE_BITS) |
| 61 | #define JA_TYPE_MASK (JA_TYPE_MAX_NR - 1) |
| 62 | #define JA_PTR_MASK (~JA_TYPE_MASK) |
| 63 | |
| 64 | #define JA_ENTRY_PER_NODE 256UL |
| 65 | |
| 66 | /* |
| 67 | * Iteration on the array to find the right node size for the number of |
| 68 | * children stops when it reaches .max_child == 256 (this is the largest |
| 69 | * possible node size, which contains 256 children). |
| 70 | * The min_child overlaps with the previous max_child to provide an |
| 71 | * hysteresis loop to reallocation for patterns of cyclic add/removal |
| 72 | * within the same node. |
| 73 | * The node the index within the following arrays is represented on 3 |
| 74 | * bits. It identifies the node type, min/max number of children, and |
| 75 | * the size order. |
| 76 | * The max_child values for the RCU_JA_POOL below result from |
| 77 | * statistical approximation: over million populations, the max_child |
| 78 | * covers between 97% and 99% of the populations generated. Therefore, a |
| 79 | * fallback should exist to cover the rare extreme population unbalance |
| 80 | * cases, but it will not have a major impact on speed nor space |
| 81 | * consumption, since those are rare cases. |
| 82 | */ |
| 83 | |
| 84 | #if (CAA_BITS_PER_LONG < 64) |
| 85 | /* 32-bit pointers */ |
| 86 | enum { |
| 87 | ja_type_0_max_child = 1, |
| 88 | ja_type_1_max_child = 3, |
| 89 | ja_type_2_max_child = 6, |
| 90 | ja_type_3_max_child = 12, |
| 91 | ja_type_4_max_child = 25, |
| 92 | ja_type_5_max_child = 48, |
| 93 | ja_type_6_max_child = 92, |
| 94 | ja_type_7_max_child = 256, |
| 95 | }; |
| 96 | |
| 97 | enum { |
| 98 | ja_type_5_nr_pool_order = 1, |
| 99 | ja_type_6_nr_pool_order = 2, |
| 100 | }; |
| 101 | |
| 102 | const struct rcu_ja_type ja_types[] = { |
| 103 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .order = 3, }, |
| 104 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .order = 4, }, |
| 105 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .order = 5, }, |
| 106 | { .type_class = RCU_JA_LINEAR, .min_child = 4, .max_child = ja_type_3_max_child, .order = 6, }, |
| 107 | { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .order = 7, }, |
| 108 | |
| 109 | /* Pools may fill sooner than max_child */ |
| 110 | { .type_class = RCU_JA_POOL, .min_child = 20, .max_child = ja_type_5_max_child, .order = 8, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 7, }, |
| 111 | { .type_class = RCU_JA_POOL, .min_child = 45, .max_child = ja_type_6_max_child, .order = 9, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 7, }, |
| 112 | |
| 113 | /* |
| 114 | * TODO: Upon node removal below min_child, if child pool is |
| 115 | * filled beyond capacity, we need to roll back to pigeon. |
| 116 | */ |
| 117 | { .type_class = RCU_JA_PIGEON, .min_child = 89, .max_child = ja_type_7_max_child, .order = 10, }, |
| 118 | }; |
| 119 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
| 120 | /* 64-bit pointers */ |
| 121 | enum { |
| 122 | ja_type_0_max_child = 1, |
| 123 | ja_type_1_max_child = 3, |
| 124 | ja_type_2_max_child = 7, |
| 125 | ja_type_3_max_child = 14, |
| 126 | ja_type_4_max_child = 28, |
| 127 | ja_type_5_max_child = 54, |
| 128 | ja_type_6_max_child = 104, |
| 129 | ja_type_7_max_child = 256, |
| 130 | }; |
| 131 | |
| 132 | enum { |
| 133 | ja_type_5_nr_pool_order = 1, |
| 134 | ja_type_6_nr_pool_order = 2, |
| 135 | }; |
| 136 | |
| 137 | const struct rcu_ja_type ja_types[] = { |
| 138 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .order = 4, }, |
| 139 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .order = 5, }, |
| 140 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .order = 6, }, |
| 141 | { .type_class = RCU_JA_LINEAR, .min_child = 5, .max_child = ja_type_3_max_child, .order = 7, }, |
| 142 | { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .order = 8, }, |
| 143 | |
| 144 | /* Pools may fill sooner than max_child. */ |
| 145 | { .type_class = RCU_JA_POOL, .min_child = 22, .max_child = ja_type_5_max_child, .order = 9, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 8, }, |
| 146 | { .type_class = RCU_JA_POOL, .min_child = 51, .max_child = ja_type_6_max_child, .order = 10, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 8, }, |
| 147 | |
| 148 | /* |
| 149 | * TODO: Upon node removal below min_child, if child pool is |
| 150 | * filled beyond capacity, we need to roll back to pigeon. |
| 151 | */ |
| 152 | { .type_class = RCU_JA_PIGEON, .min_child = 101, .max_child = ja_type_7_max_child, .order = 11, }, |
| 153 | }; |
| 154 | #endif /* !(BITS_PER_LONG < 64) */ |
| 155 | |
| 156 | static inline __attribute__((unused)) |
| 157 | void static_array_size_check(void) |
| 158 | { |
| 159 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) > JA_TYPE_MAX_NR); |
| 160 | } |
| 161 | |
| 162 | /* Never declared. Opaque type used to store flagged node pointers. */ |
| 163 | struct rcu_ja_node_flag; |
| 164 | |
| 165 | /* |
| 166 | * The rcu_ja_node contains the compressed node data needed for |
| 167 | * read-side. For linear and pool node configurations, it starts with a |
| 168 | * byte counting the number of children in the node. Then, the |
| 169 | * node-specific data is placed. |
| 170 | * The node mutex, if any is needed, protecting concurrent updated of |
| 171 | * each node is placed in a separate hash table indexed by node address. |
| 172 | * For the pigeon configuration, the number of children is also kept in |
| 173 | * a separate hash table, indexed by node address, because it is only |
| 174 | * required for updates. |
| 175 | */ |
| 176 | |
| 177 | #define DECLARE_LINEAR_NODE(index) \ |
| 178 | struct { \ |
| 179 | uint8_t nr_child; \ |
| 180 | uint8_t child_value[ja_type_## index ##_max_child]; \ |
| 181 | struct rcu_ja_node_flag *child_ptr[ja_type_## index ##_max_child]; \ |
| 182 | } |
| 183 | |
| 184 | #define DECLARE_POOL_NODE(index) \ |
| 185 | struct { \ |
| 186 | struct { \ |
| 187 | uint8_t nr_child; \ |
| 188 | uint8_t child_value[ja_type_## index ##_max_child >> ja_type_## index ##_nr_pool_order]; \ |
| 189 | struct rcu_ja_node_flag *child_ptr[ja_type_## index ##_max_child >> ja_type_## index ##_nr_pool_order]; \ |
| 190 | } linear[1U << ja_type_## index ##_nr_pool_order]; \ |
| 191 | } |
| 192 | |
| 193 | struct rcu_ja_node { |
| 194 | union { |
| 195 | /* Linear configuration */ |
| 196 | DECLARE_LINEAR_NODE(0) conf_0; |
| 197 | DECLARE_LINEAR_NODE(1) conf_1; |
| 198 | DECLARE_LINEAR_NODE(2) conf_2; |
| 199 | DECLARE_LINEAR_NODE(3) conf_3; |
| 200 | DECLARE_LINEAR_NODE(4) conf_4; |
| 201 | |
| 202 | /* Pool configuration */ |
| 203 | DECLARE_POOL_NODE(5) conf_5; |
| 204 | DECLARE_POOL_NODE(6) conf_6; |
| 205 | |
| 206 | /* Pigeon configuration */ |
| 207 | struct { |
| 208 | struct rcu_ja_node_flag *child[ja_type_7_max_child]; |
| 209 | } conf_7; |
| 210 | /* data aliasing nodes for computed accesses */ |
| 211 | uint8_t data[sizeof(struct rcu_ja_node_flag *) * ja_type_7_max_child]; |
| 212 | } u; |
| 213 | }; |
| 214 | |
| 215 | static |
| 216 | struct rcu_ja_node_flag *ja_node_flag(struct rcu_ja_node *node, unsigned int type) |
| 217 | { |
| 218 | assert(type < RCU_JA_NR_TYPES); |
| 219 | return (struct rcu_ja_node_flag *) (((unsigned long) node) | type); |
| 220 | } |
| 221 | |
| 222 | static |
| 223 | unsigned int ja_node_type(struct rcu_ja_node_flag *node) |
| 224 | { |
| 225 | unsigned int type; |
| 226 | |
| 227 | type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK); |
| 228 | assert(type < RCU_JA_NR_TYPES); |
| 229 | return type; |
| 230 | } |
| 231 | |
| 232 | static |
| 233 | struct rcu_ja_node *ja_node_ptr(struct rcu_ja_node_flag *node) |
| 234 | { |
| 235 | return (struct rcu_ja_node *) (((unsigned long) node) | JA_PTR_MASK); |
| 236 | } |
| 237 | |
| 238 | struct rcu_ja_node *alloc_rcu_ja_node(struct rcu_ja_type *ja_type) |
| 239 | { |
| 240 | return calloc(1U << ja_type->order, sizeof(char)); |
| 241 | } |
| 242 | |
| 243 | void free_rcu_ja_node(struct rcu_ja_node *node) |
| 244 | { |
| 245 | free(node); |
| 246 | } |
| 247 | |
| 248 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) |
| 249 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) |
| 250 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) |
| 251 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) |
| 252 | |
| 253 | static |
| 254 | uint8_t *align_ptr_size(uint8_t *ptr) |
| 255 | { |
| 256 | return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *)); |
| 257 | } |
| 258 | |
| 259 | static |
| 260 | struct rcu_ja_node_flag *ja_linear_node_get_nth(const struct rcu_ja_type *type, |
| 261 | struct rcu_ja_node *node, |
| 262 | uint8_t n) |
| 263 | { |
| 264 | uint8_t nr_child; |
| 265 | uint8_t *values; |
| 266 | struct rcu_ja_node_flag **pointers; |
| 267 | struct rcu_ja_node_flag *ptr; |
| 268 | unsigned int i; |
| 269 | |
| 270 | assert(!type || type->type_class == RCU_JA_LINEAR); |
| 271 | |
| 272 | nr_child = node->u.data[0]; |
| 273 | cmm_smp_rmb(); /* read nr_child before values */ |
| 274 | assert(!type || nr_child <= type->max_child); |
| 275 | assert(!type || nr_child >= type->min_child); |
| 276 | |
| 277 | values = &node->u.data[1]; |
| 278 | for (i = 0; i < nr_child; i++) { |
| 279 | if (values[i] == n) |
| 280 | break; |
| 281 | } |
| 282 | if (i >= nr_child) |
| 283 | return NULL; |
| 284 | cmm_smp_rmb(); /* read values before pointer */ |
| 285 | pointers = (struct rcu_ja_node_flag **) align_ptr_size(&values[nr_child]); |
| 286 | ptr = pointers[i]; |
| 287 | assert(ja_node_ptr(ptr) != NULL); |
| 288 | return ptr; |
| 289 | } |
| 290 | |
| 291 | static |
| 292 | struct rcu_ja_node_flag *ja_pool_node_get_nth(const struct rcu_ja_type *type, |
| 293 | struct rcu_ja_node *node, |
| 294 | uint8_t n) |
| 295 | { |
| 296 | struct rcu_ja_node *linear; |
| 297 | |
| 298 | assert(type->type_class == RCU_JA_POOL); |
| 299 | linear = (struct rcu_ja_node *) |
| 300 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
| 301 | return ja_linear_node_get_nth(NULL, linear, n); |
| 302 | } |
| 303 | |
| 304 | static |
| 305 | struct rcu_ja_node_flag *ja_pigeon_node_get_nth(const struct rcu_ja_type *type, |
| 306 | struct rcu_ja_node *node, |
| 307 | uint8_t n) |
| 308 | { |
| 309 | assert(type->type_class == RCU_JA_PIGEON); |
| 310 | return ((struct rcu_ja_node_flag **) node->u.data)[n]; |
| 311 | } |
| 312 | |
| 313 | /* ja_node_get_nth: get nth item from a node */ |
| 314 | static |
| 315 | struct rcu_ja_node_flag *ja_node_get_nth(struct rcu_ja_node_flag *node_flag, |
| 316 | uint8_t n) |
| 317 | { |
| 318 | unsigned int type_index; |
| 319 | struct rcu_ja_node *node; |
| 320 | const struct rcu_ja_type *type; |
| 321 | |
| 322 | node_flag = rcu_dereference(node_flag); |
| 323 | node = ja_node_ptr(node_flag); |
| 324 | assert(node != NULL); |
| 325 | type_index = ja_node_type(node_flag); |
| 326 | type = &ja_types[type_index]; |
| 327 | |
| 328 | switch (type->type_class) { |
| 329 | case RCU_JA_LINEAR: |
| 330 | return ja_linear_node_get_nth(type, node, n); |
| 331 | case RCU_JA_POOL: |
| 332 | return ja_pool_node_get_nth(type, node, n); |
| 333 | case RCU_JA_PIGEON: |
| 334 | return ja_pigeon_node_get_nth(type, node, n); |
| 335 | default: |
| 336 | assert(0); |
| 337 | return (void *) -1UL; |
| 338 | } |
| 339 | } |
| 340 | |
| 341 | /* |
| 342 | * ja_node_set_nth: set nth item within a node. asserts that it is not |
| 343 | * there yet. |
| 344 | */ |
| 345 | |