| 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 | #define _LGPL_SOURCE |
| 24 | #include <stdint.h> |
| 25 | #include <errno.h> |
| 26 | #include <limits.h> |
| 27 | #include <urcu/rcuja.h> |
| 28 | #include <urcu/compiler.h> |
| 29 | #include <urcu/arch.h> |
| 30 | #include <assert.h> |
| 31 | #include <urcu-pointer.h> |
| 32 | #include <urcu/uatomic.h> |
| 33 | #include <stdint.h> |
| 34 | |
| 35 | #include "rcuja-internal.h" |
| 36 | #include "bitfield.h" |
| 37 | |
| 38 | enum cds_ja_type_class { |
| 39 | RCU_JA_LINEAR = 0, /* Type A */ |
| 40 | /* 32-bit: 1 to 25 children, 8 to 128 bytes */ |
| 41 | /* 64-bit: 1 to 28 children, 16 to 256 bytes */ |
| 42 | RCU_JA_POOL = 1, /* Type B */ |
| 43 | /* 32-bit: 26 to 100 children, 256 to 512 bytes */ |
| 44 | /* 64-bit: 29 to 112 children, 512 to 1024 bytes */ |
| 45 | RCU_JA_PIGEON = 2, /* Type C */ |
| 46 | /* 32-bit: 101 to 256 children, 1024 bytes */ |
| 47 | /* 64-bit: 113 to 256 children, 2048 bytes */ |
| 48 | /* Leaf nodes are implicit from their height in the tree */ |
| 49 | RCU_JA_NR_TYPES, |
| 50 | |
| 51 | RCU_JA_NULL, /* not an encoded type, but keeps code regular */ |
| 52 | }; |
| 53 | |
| 54 | struct cds_ja_type { |
| 55 | enum cds_ja_type_class type_class; |
| 56 | uint16_t min_child; /* minimum number of children: 1 to 256 */ |
| 57 | uint16_t max_child; /* maximum number of children: 1 to 256 */ |
| 58 | uint16_t max_linear_child; /* per-pool max nr. children: 1 to 256 */ |
| 59 | uint16_t order; /* node size is (1 << order), in bytes */ |
| 60 | uint16_t nr_pool_order; /* number of pools */ |
| 61 | uint16_t pool_size_order; /* pool size */ |
| 62 | }; |
| 63 | |
| 64 | /* |
| 65 | * Iteration on the array to find the right node size for the number of |
| 66 | * children stops when it reaches .max_child == 256 (this is the largest |
| 67 | * possible node size, which contains 256 children). |
| 68 | * The min_child overlaps with the previous max_child to provide an |
| 69 | * hysteresis loop to reallocation for patterns of cyclic add/removal |
| 70 | * within the same node. |
| 71 | * The node the index within the following arrays is represented on 3 |
| 72 | * bits. It identifies the node type, min/max number of children, and |
| 73 | * the size order. |
| 74 | * The max_child values for the RCU_JA_POOL below result from |
| 75 | * statistical approximation: over million populations, the max_child |
| 76 | * covers between 97% and 99% of the populations generated. Therefore, a |
| 77 | * fallback should exist to cover the rare extreme population unbalance |
| 78 | * cases, but it will not have a major impact on speed nor space |
| 79 | * consumption, since those are rare cases. |
| 80 | */ |
| 81 | |
| 82 | #if (CAA_BITS_PER_LONG < 64) |
| 83 | /* 32-bit pointers */ |
| 84 | enum { |
| 85 | ja_type_0_max_child = 1, |
| 86 | ja_type_1_max_child = 3, |
| 87 | ja_type_2_max_child = 6, |
| 88 | ja_type_3_max_child = 12, |
| 89 | ja_type_4_max_child = 25, |
| 90 | ja_type_5_max_child = 48, |
| 91 | ja_type_6_max_child = 92, |
| 92 | ja_type_7_max_child = 256, |
| 93 | ja_type_8_max_child = 0, /* NULL */ |
| 94 | }; |
| 95 | |
| 96 | enum { |
| 97 | ja_type_0_max_linear_child = 1, |
| 98 | ja_type_1_max_linear_child = 3, |
| 99 | ja_type_2_max_linear_child = 6, |
| 100 | ja_type_3_max_linear_child = 12, |
| 101 | ja_type_4_max_linear_child = 25, |
| 102 | ja_type_5_max_linear_child = 24, |
| 103 | ja_type_6_max_linear_child = 23, |
| 104 | }; |
| 105 | |
| 106 | enum { |
| 107 | ja_type_5_nr_pool_order = 1, |
| 108 | ja_type_6_nr_pool_order = 2, |
| 109 | }; |
| 110 | |
| 111 | const struct cds_ja_type ja_types[] = { |
| 112 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .max_linear_child = ja_type_0_max_linear_child, .order = 3, }, |
| 113 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .max_linear_child = ja_type_1_max_linear_child, .order = 4, }, |
| 114 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .max_linear_child = ja_type_2_max_linear_child, .order = 5, }, |
| 115 | { .type_class = RCU_JA_LINEAR, .min_child = 4, .max_child = ja_type_3_max_child, .max_linear_child = ja_type_3_max_linear_child, .order = 6, }, |
| 116 | { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .max_linear_child = ja_type_4_max_linear_child, .order = 7, }, |
| 117 | |
| 118 | /* Pools may fill sooner than max_child */ |
| 119 | { .type_class = RCU_JA_POOL, .min_child = 20, .max_child = ja_type_5_max_child, .max_linear_child = ja_type_5_max_linear_child, .order = 8, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 7, }, |
| 120 | { .type_class = RCU_JA_POOL, .min_child = 45, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 9, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 7, }, |
| 121 | |
| 122 | /* |
| 123 | * TODO: Upon node removal below min_child, if child pool is |
| 124 | * filled beyond capacity, we need to roll back to pigeon. |
| 125 | */ |
| 126 | { .type_class = RCU_JA_PIGEON, .min_child = 89, .max_child = ja_type_7_max_child, .order = 10, }, |
| 127 | |
| 128 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, |
| 129 | }; |
| 130 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
| 131 | /* 64-bit pointers */ |
| 132 | enum { |
| 133 | ja_type_0_max_child = 1, |
| 134 | ja_type_1_max_child = 3, |
| 135 | ja_type_2_max_child = 7, |
| 136 | ja_type_3_max_child = 14, |
| 137 | ja_type_4_max_child = 28, |
| 138 | ja_type_5_max_child = 54, |
| 139 | ja_type_6_max_child = 104, |
| 140 | ja_type_7_max_child = 256, |
| 141 | ja_type_8_max_child = 256, |
| 142 | }; |
| 143 | |
| 144 | enum { |
| 145 | ja_type_0_max_linear_child = 1, |
| 146 | ja_type_1_max_linear_child = 3, |
| 147 | ja_type_2_max_linear_child = 7, |
| 148 | ja_type_3_max_linear_child = 14, |
| 149 | ja_type_4_max_linear_child = 28, |
| 150 | ja_type_5_max_linear_child = 27, |
| 151 | ja_type_6_max_linear_child = 26, |
| 152 | }; |
| 153 | |
| 154 | enum { |
| 155 | ja_type_5_nr_pool_order = 1, |
| 156 | ja_type_6_nr_pool_order = 2, |
| 157 | }; |
| 158 | |
| 159 | const struct cds_ja_type ja_types[] = { |
| 160 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_0_max_child, .max_linear_child = ja_type_0_max_linear_child, .order = 4, }, |
| 161 | { .type_class = RCU_JA_LINEAR, .min_child = 1, .max_child = ja_type_1_max_child, .max_linear_child = ja_type_1_max_linear_child, .order = 5, }, |
| 162 | { .type_class = RCU_JA_LINEAR, .min_child = 3, .max_child = ja_type_2_max_child, .max_linear_child = ja_type_2_max_linear_child, .order = 6, }, |
| 163 | { .type_class = RCU_JA_LINEAR, .min_child = 5, .max_child = ja_type_3_max_child, .max_linear_child = ja_type_3_max_linear_child, .order = 7, }, |
| 164 | { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .max_linear_child = ja_type_4_max_linear_child, .order = 8, }, |
| 165 | |
| 166 | /* Pools may fill sooner than max_child. */ |
| 167 | { .type_class = RCU_JA_POOL, .min_child = 22, .max_child = ja_type_5_max_child, .max_linear_child = ja_type_5_max_linear_child, .order = 9, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 8, }, |
| 168 | { .type_class = RCU_JA_POOL, .min_child = 51, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 10, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 8, }, |
| 169 | |
| 170 | /* |
| 171 | * TODO: Upon node removal below min_child, if child pool is |
| 172 | * filled beyond capacity, we need to roll back to pigeon. |
| 173 | */ |
| 174 | { .type_class = RCU_JA_PIGEON, .min_child = 101, .max_child = ja_type_7_max_child, .order = 11, }, |
| 175 | |
| 176 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, |
| 177 | }; |
| 178 | #endif /* !(BITS_PER_LONG < 64) */ |
| 179 | |
| 180 | static inline __attribute__((unused)) |
| 181 | void static_array_size_check(void) |
| 182 | { |
| 183 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) < JA_TYPE_MAX_NR); |
| 184 | } |
| 185 | |
| 186 | /* |
| 187 | * The cds_ja_node contains the compressed node data needed for |
| 188 | * read-side. For linear and pool node configurations, it starts with a |
| 189 | * byte counting the number of children in the node. Then, the |
| 190 | * node-specific data is placed. |
| 191 | * The node mutex, if any is needed, protecting concurrent updated of |
| 192 | * each node is placed in a separate hash table indexed by node address. |
| 193 | * For the pigeon configuration, the number of children is also kept in |
| 194 | * a separate hash table, indexed by node address, because it is only |
| 195 | * required for updates. |
| 196 | */ |
| 197 | |
| 198 | #define DECLARE_LINEAR_NODE(index) \ |
| 199 | struct { \ |
| 200 | uint8_t nr_child; \ |
| 201 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ |
| 202 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
| 203 | } |
| 204 | |
| 205 | #define DECLARE_POOL_NODE(index) \ |
| 206 | struct { \ |
| 207 | struct { \ |
| 208 | uint8_t nr_child; \ |
| 209 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ |
| 210 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
| 211 | } linear[1U << ja_type_## index ##_nr_pool_order]; \ |
| 212 | } |
| 213 | |
| 214 | struct cds_ja_inode { |
| 215 | union { |
| 216 | /* Linear configuration */ |
| 217 | DECLARE_LINEAR_NODE(0) conf_0; |
| 218 | DECLARE_LINEAR_NODE(1) conf_1; |
| 219 | DECLARE_LINEAR_NODE(2) conf_2; |
| 220 | DECLARE_LINEAR_NODE(3) conf_3; |
| 221 | DECLARE_LINEAR_NODE(4) conf_4; |
| 222 | |
| 223 | /* Pool configuration */ |
| 224 | DECLARE_POOL_NODE(5) conf_5; |
| 225 | DECLARE_POOL_NODE(6) conf_6; |
| 226 | |
| 227 | /* Pigeon configuration */ |
| 228 | struct { |
| 229 | struct cds_ja_inode_flag *child[ja_type_7_max_child]; |
| 230 | } conf_7; |
| 231 | /* data aliasing nodes for computed accesses */ |
| 232 | uint8_t data[sizeof(struct cds_ja_inode_flag *) * ja_type_7_max_child]; |
| 233 | } u; |
| 234 | }; |
| 235 | |
| 236 | enum ja_recompact { |
| 237 | JA_RECOMPACT, |
| 238 | JA_RECOMPACT_ADD, |
| 239 | JA_RECOMPACT_DEL, |
| 240 | }; |
| 241 | |
| 242 | struct cds_ja_inode *alloc_cds_ja_node(const struct cds_ja_type *ja_type) |
| 243 | { |
| 244 | return calloc(1U << ja_type->order, sizeof(char)); |
| 245 | } |
| 246 | |
| 247 | void free_cds_ja_node(struct cds_ja_inode *node) |
| 248 | { |
| 249 | free(node); |
| 250 | } |
| 251 | |
| 252 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) |
| 253 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) |
| 254 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) |
| 255 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) |
| 256 | |
| 257 | static |
| 258 | uint8_t *align_ptr_size(uint8_t *ptr) |
| 259 | { |
| 260 | return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *)); |
| 261 | } |
| 262 | |
| 263 | static |
| 264 | uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type, |
| 265 | struct cds_ja_inode *node) |
| 266 | { |
| 267 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 268 | return rcu_dereference(node->u.data[0]); |
| 269 | } |
| 270 | |
| 271 | /* |
| 272 | * The order in which values and pointers are does does not matter: if |
| 273 | * a value is missing, we return NULL. If a value is there, but its |
| 274 | * associated pointers is still NULL, we return NULL too. |
| 275 | */ |
| 276 | static |
| 277 | struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, |
| 278 | struct cds_ja_inode *node, |
| 279 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
| 280 | struct cds_ja_inode_flag **child_node_flag_v, |
| 281 | struct cds_ja_inode_flag ***node_flag_ptr, |
| 282 | uint8_t n) |
| 283 | { |
| 284 | uint8_t nr_child; |
| 285 | uint8_t *values; |
| 286 | struct cds_ja_inode_flag **pointers; |
| 287 | struct cds_ja_inode_flag *ptr; |
| 288 | unsigned int i; |
| 289 | |
| 290 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 291 | |
| 292 | nr_child = ja_linear_node_get_nr_child(type, node); |
| 293 | cmm_smp_rmb(); /* read nr_child before values and pointers */ |
| 294 | assert(nr_child <= type->max_linear_child); |
| 295 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); |
| 296 | |
| 297 | values = &node->u.data[1]; |
| 298 | for (i = 0; i < nr_child; i++) { |
| 299 | if (CMM_LOAD_SHARED(values[i]) == n) |
| 300 | break; |
| 301 | } |
| 302 | if (i >= nr_child) { |
| 303 | if (caa_unlikely(node_flag_ptr)) |
| 304 | *node_flag_ptr = NULL; |
| 305 | return NULL; |
| 306 | } |
| 307 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
| 308 | ptr = rcu_dereference(pointers[i]); |
| 309 | if (caa_unlikely(child_node_flag_ptr) && ptr) |
| 310 | *child_node_flag_ptr = &pointers[i]; |
| 311 | if (caa_unlikely(child_node_flag_v) && ptr) |
| 312 | *child_node_flag_v = ptr; |
| 313 | if (caa_unlikely(node_flag_ptr)) |
| 314 | *node_flag_ptr = &pointers[i]; |
| 315 | return ptr; |
| 316 | } |
| 317 | |
| 318 | static |
| 319 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, |
| 320 | struct cds_ja_inode *node, |
| 321 | uint8_t i, |
| 322 | uint8_t *v, |
| 323 | struct cds_ja_inode_flag **iter) |
| 324 | { |
| 325 | uint8_t *values; |
| 326 | struct cds_ja_inode_flag **pointers; |
| 327 | |
| 328 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 329 | assert(i < ja_linear_node_get_nr_child(type, node)); |
| 330 | |
| 331 | values = &node->u.data[1]; |
| 332 | *v = values[i]; |
| 333 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
| 334 | *iter = pointers[i]; |
| 335 | } |
| 336 | |
| 337 | static |
| 338 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, |
| 339 | struct cds_ja_inode *node, |
| 340 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
| 341 | struct cds_ja_inode_flag **child_node_flag_v, |
| 342 | struct cds_ja_inode_flag ***node_flag_ptr, |
| 343 | uint8_t n) |
| 344 | { |
| 345 | struct cds_ja_inode *linear; |
| 346 | |
| 347 | assert(type->type_class == RCU_JA_POOL); |
| 348 | /* |
| 349 | * TODO: currently, we select the pool by highest bits. We |
| 350 | * should support various encodings. |
| 351 | */ |
| 352 | linear = (struct cds_ja_inode *) |
| 353 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
| 354 | return ja_linear_node_get_nth(type, linear, child_node_flag_ptr, |
| 355 | child_node_flag_v, node_flag_ptr, n); |
| 356 | } |
| 357 | |
| 358 | static |
| 359 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, |
| 360 | struct cds_ja_inode *node, |
| 361 | uint8_t i) |
| 362 | { |
| 363 | assert(type->type_class == RCU_JA_POOL); |
| 364 | return (struct cds_ja_inode *) |
| 365 | &node->u.data[(unsigned int) i << type->pool_size_order]; |
| 366 | } |
| 367 | |
| 368 | static |
| 369 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, |
| 370 | struct cds_ja_inode *node, |
| 371 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
| 372 | struct cds_ja_inode_flag **child_node_flag_v, |
| 373 | struct cds_ja_inode_flag ***node_flag_ptr, |
| 374 | uint8_t n) |
| 375 | { |
| 376 | struct cds_ja_inode_flag **child_node_flag; |
| 377 | struct cds_ja_inode_flag *child_node_flag_read; |
| 378 | |
| 379 | assert(type->type_class == RCU_JA_PIGEON); |
| 380 | child_node_flag = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
| 381 | child_node_flag_read = rcu_dereference(*child_node_flag); |
| 382 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", |
| 383 | child_node_flag); |
| 384 | if (caa_unlikely(child_node_flag_ptr) && child_node_flag_read) |
| 385 | *child_node_flag_ptr = child_node_flag; |
| 386 | if (caa_unlikely(child_node_flag_v) && child_node_flag_read) |
| 387 | *child_node_flag_v = child_node_flag_read; |
| 388 | if (caa_unlikely(node_flag_ptr)) |
| 389 | *node_flag_ptr = child_node_flag; |
| 390 | return child_node_flag_read; |
| 391 | } |
| 392 | |
| 393 | static |
| 394 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, |
| 395 | struct cds_ja_inode *node, |
| 396 | uint8_t i) |
| 397 | { |
| 398 | return ja_pigeon_node_get_nth(type, node, NULL, NULL, NULL, i); |
| 399 | } |
| 400 | |
| 401 | /* |
| 402 | * ja_node_get_nth: get nth item from a node. |
| 403 | * node_flag is already rcu_dereference'd. |
| 404 | */ |
| 405 | static |
| 406 | struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, |
| 407 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
| 408 | struct cds_ja_inode_flag **child_node_flag, |
| 409 | struct cds_ja_inode_flag ***node_flag_ptr, |
| 410 | uint8_t n) |
| 411 | { |
| 412 | unsigned int type_index; |
| 413 | struct cds_ja_inode *node; |
| 414 | const struct cds_ja_type *type; |
| 415 | |
| 416 | node = ja_node_ptr(node_flag); |
| 417 | assert(node != NULL); |
| 418 | type_index = ja_node_type(node_flag); |
| 419 | type = &ja_types[type_index]; |
| 420 | |
| 421 | switch (type->type_class) { |
| 422 | case RCU_JA_LINEAR: |
| 423 | return ja_linear_node_get_nth(type, node, |
| 424 | child_node_flag_ptr, child_node_flag, |
| 425 | node_flag_ptr, n); |
| 426 | case RCU_JA_POOL: |
| 427 | return ja_pool_node_get_nth(type, node, |
| 428 | child_node_flag_ptr, child_node_flag, |
| 429 | node_flag_ptr, n); |
| 430 | case RCU_JA_PIGEON: |
| 431 | return ja_pigeon_node_get_nth(type, node, |
| 432 | child_node_flag_ptr, child_node_flag, |
| 433 | node_flag_ptr, n); |
| 434 | default: |
| 435 | assert(0); |
| 436 | return (void *) -1UL; |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | static |
| 441 | int ja_linear_node_set_nth(const struct cds_ja_type *type, |
| 442 | struct cds_ja_inode *node, |
| 443 | struct cds_ja_shadow_node *shadow_node, |
| 444 | uint8_t n, |
| 445 | struct cds_ja_inode_flag *child_node_flag) |
| 446 | { |
| 447 | uint8_t nr_child; |
| 448 | uint8_t *values, *nr_child_ptr; |
| 449 | struct cds_ja_inode_flag **pointers; |
| 450 | unsigned int i, unused = 0; |
| 451 | |
| 452 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 453 | |
| 454 | nr_child_ptr = &node->u.data[0]; |
| 455 | dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr); |
| 456 | nr_child = *nr_child_ptr; |
| 457 | assert(nr_child <= type->max_linear_child); |
| 458 | |
| 459 | values = &node->u.data[1]; |
| 460 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
| 461 | /* Check if node value is already populated */ |
| 462 | for (i = 0; i < nr_child; i++) { |
| 463 | if (values[i] == n) { |
| 464 | if (pointers[i]) |
| 465 | return -EEXIST; |
| 466 | else |
| 467 | break; |
| 468 | } else { |
| 469 | if (!pointers[i]) |
| 470 | unused++; |
| 471 | } |
| 472 | } |
| 473 | if (i == nr_child && nr_child >= type->max_linear_child) { |
| 474 | if (unused) |
| 475 | return -ERANGE; /* recompact node */ |
| 476 | else |
| 477 | return -ENOSPC; /* No space left in this node type */ |
| 478 | } |
| 479 | |
| 480 | assert(pointers[i] == NULL); |
| 481 | rcu_assign_pointer(pointers[i], child_node_flag); |
| 482 | /* If we expanded the nr_child, increment it */ |
| 483 | if (i == nr_child) { |
| 484 | CMM_STORE_SHARED(values[nr_child], n); |
| 485 | /* write pointer and value before nr_child */ |
| 486 | cmm_smp_wmb(); |
| 487 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); |
| 488 | } |
| 489 | shadow_node->nr_child++; |
| 490 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", |
| 491 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
| 492 | (unsigned int) shadow_node->nr_child, |
| 493 | node, shadow_node); |
| 494 | |
| 495 | return 0; |
| 496 | } |
| 497 | |
| 498 | static |
| 499 | int ja_pool_node_set_nth(const struct cds_ja_type *type, |
| 500 | struct cds_ja_inode *node, |
| 501 | struct cds_ja_shadow_node *shadow_node, |
| 502 | uint8_t n, |
| 503 | struct cds_ja_inode_flag *child_node_flag) |
| 504 | { |
| 505 | struct cds_ja_inode *linear; |
| 506 | |
| 507 | assert(type->type_class == RCU_JA_POOL); |
| 508 | linear = (struct cds_ja_inode *) |
| 509 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
| 510 | return ja_linear_node_set_nth(type, linear, shadow_node, |
| 511 | n, child_node_flag); |
| 512 | } |
| 513 | |
| 514 | static |
| 515 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, |
| 516 | struct cds_ja_inode *node, |
| 517 | struct cds_ja_shadow_node *shadow_node, |
| 518 | uint8_t n, |
| 519 | struct cds_ja_inode_flag *child_node_flag) |
| 520 | { |
| 521 | struct cds_ja_inode_flag **ptr; |
| 522 | |
| 523 | assert(type->type_class == RCU_JA_PIGEON); |
| 524 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
| 525 | if (*ptr) |
| 526 | return -EEXIST; |
| 527 | rcu_assign_pointer(*ptr, child_node_flag); |
| 528 | shadow_node->nr_child++; |
| 529 | return 0; |
| 530 | } |
| 531 | |
| 532 | /* |
| 533 | * _ja_node_set_nth: set nth item within a node. Return an error |
| 534 | * (negative error value) if it is already there. |
| 535 | */ |
| 536 | static |
| 537 | int _ja_node_set_nth(const struct cds_ja_type *type, |
| 538 | struct cds_ja_inode *node, |
| 539 | struct cds_ja_shadow_node *shadow_node, |
| 540 | uint8_t n, |
| 541 | struct cds_ja_inode_flag *child_node_flag) |
| 542 | { |
| 543 | switch (type->type_class) { |
| 544 | case RCU_JA_LINEAR: |
| 545 | return ja_linear_node_set_nth(type, node, shadow_node, n, |
| 546 | child_node_flag); |
| 547 | case RCU_JA_POOL: |
| 548 | return ja_pool_node_set_nth(type, node, shadow_node, n, |
| 549 | child_node_flag); |
| 550 | case RCU_JA_PIGEON: |
| 551 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, |
| 552 | child_node_flag); |
| 553 | case RCU_JA_NULL: |
| 554 | return -ENOSPC; |
| 555 | default: |
| 556 | assert(0); |
| 557 | return -EINVAL; |
| 558 | } |
| 559 | |
| 560 | return 0; |
| 561 | } |
| 562 | |
| 563 | static |
| 564 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, |
| 565 | struct cds_ja_inode *node, |
| 566 | struct cds_ja_shadow_node *shadow_node, |
| 567 | struct cds_ja_inode_flag **node_flag_ptr) |
| 568 | { |
| 569 | uint8_t nr_child; |
| 570 | uint8_t *nr_child_ptr; |
| 571 | |
| 572 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 573 | |
| 574 | nr_child_ptr = &node->u.data[0]; |
| 575 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); |
| 576 | nr_child = *nr_child_ptr; |
| 577 | assert(nr_child <= type->max_linear_child); |
| 578 | |
| 579 | if (shadow_node->fallback_removal_count) { |
| 580 | shadow_node->fallback_removal_count--; |
| 581 | } else { |
| 582 | if (shadow_node->nr_child <= type->min_child) { |
| 583 | /* We need to try recompacting the node */ |
| 584 | return -EFBIG; |
| 585 | } |
| 586 | } |
| 587 | assert(*node_flag_ptr != NULL); |
| 588 | rcu_assign_pointer(*node_flag_ptr, NULL); |
| 589 | /* |
| 590 | * Value and nr_child are never changed (would cause ABA issue). |
| 591 | * Instead, we leave the pointer to NULL and recompact the node |
| 592 | * once in a while. It is allowed to set a NULL pointer to a new |
| 593 | * value without recompaction though. |
| 594 | * Only update the shadow node accounting. |
| 595 | */ |
| 596 | shadow_node->nr_child--; |
| 597 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", |
| 598 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
| 599 | (unsigned int) shadow_node->nr_child, |
| 600 | node, shadow_node); |
| 601 | |
| 602 | return 0; |
| 603 | } |
| 604 | |
| 605 | static |
| 606 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, |
| 607 | struct cds_ja_inode *node, |
| 608 | struct cds_ja_shadow_node *shadow_node, |
| 609 | struct cds_ja_inode_flag **node_flag_ptr, |
| 610 | uint8_t n) |
| 611 | { |
| 612 | struct cds_ja_inode *linear; |
| 613 | |
| 614 | assert(type->type_class == RCU_JA_POOL); |
| 615 | linear = (struct cds_ja_inode *) |
| 616 | &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; |
| 617 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); |
| 618 | } |
| 619 | |
| 620 | static |
| 621 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, |
| 622 | struct cds_ja_inode *node, |
| 623 | struct cds_ja_shadow_node *shadow_node, |
| 624 | struct cds_ja_inode_flag **node_flag_ptr) |
| 625 | { |
| 626 | assert(type->type_class == RCU_JA_PIGEON); |
| 627 | dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); |
| 628 | rcu_assign_pointer(*node_flag_ptr, NULL); |
| 629 | shadow_node->nr_child--; |
| 630 | return 0; |
| 631 | } |
| 632 | |
| 633 | /* |
| 634 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error |
| 635 | * (negative error value) if it is not found (-ENOENT). |
| 636 | */ |
| 637 | static |
| 638 | int _ja_node_clear_ptr(const struct cds_ja_type *type, |
| 639 | struct cds_ja_inode *node, |
| 640 | struct cds_ja_shadow_node *shadow_node, |
| 641 | struct cds_ja_inode_flag **node_flag_ptr, |
| 642 | uint8_t n) |
| 643 | { |
| 644 | switch (type->type_class) { |
| 645 | case RCU_JA_LINEAR: |
| 646 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
| 647 | case RCU_JA_POOL: |
| 648 | return ja_pool_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
| 649 | case RCU_JA_PIGEON: |
| 650 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
| 651 | case RCU_JA_NULL: |
| 652 | return -ENOENT; |
| 653 | default: |
| 654 | assert(0); |
| 655 | return -EINVAL; |
| 656 | } |
| 657 | |
| 658 | return 0; |
| 659 | } |
| 660 | |
| 661 | /* |
| 662 | * ja_node_recompact_add: recompact a node, adding a new child. |
| 663 | * TODO: for pool type, take selection bit(s) into account. |
| 664 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
| 665 | * error value otherwise. |
| 666 | */ |
| 667 | static |
| 668 | int ja_node_recompact(enum ja_recompact mode, |
| 669 | struct cds_ja *ja, |
| 670 | unsigned int old_type_index, |
| 671 | const struct cds_ja_type *old_type, |
| 672 | struct cds_ja_inode *old_node, |
| 673 | struct cds_ja_shadow_node *shadow_node, |
| 674 | struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, |
| 675 | struct cds_ja_inode_flag *child_node_flag, |
| 676 | struct cds_ja_inode_flag **nullify_node_flag_ptr) |
| 677 | { |
| 678 | unsigned int new_type_index; |
| 679 | struct cds_ja_inode *new_node; |
| 680 | struct cds_ja_shadow_node *new_shadow_node = NULL; |
| 681 | const struct cds_ja_type *new_type; |
| 682 | struct cds_ja_inode_flag *new_node_flag, *old_node_flag; |
| 683 | int ret; |
| 684 | int fallback = 0; |
| 685 | |
| 686 | old_node_flag = *old_node_flag_ptr; |
| 687 | |
| 688 | switch (mode) { |
| 689 | case JA_RECOMPACT: |
| 690 | new_type_index = old_type_index; |
| 691 | break; |
| 692 | case JA_RECOMPACT_ADD: |
| 693 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { |
| 694 | new_type_index = 0; |
| 695 | } else { |
| 696 | new_type_index = old_type_index + 1; |
| 697 | } |
| 698 | break; |
| 699 | case JA_RECOMPACT_DEL: |
| 700 | if (old_type_index == 0) { |
| 701 | new_type_index = NODE_INDEX_NULL; |
| 702 | } else { |
| 703 | new_type_index = old_type_index - 1; |
| 704 | } |
| 705 | break; |
| 706 | default: |
| 707 | assert(0); |
| 708 | } |
| 709 | |
| 710 | retry: /* for fallback */ |
| 711 | dbg_printf("Recompact from type %d to type %d\n", |
| 712 | old_type_index, new_type_index); |
| 713 | new_type = &ja_types[new_type_index]; |
| 714 | if (new_type_index != NODE_INDEX_NULL) { |
| 715 | new_node = alloc_cds_ja_node(new_type); |
| 716 | if (!new_node) |
| 717 | return -ENOMEM; |
| 718 | new_node_flag = ja_node_flag(new_node, new_type_index); |
| 719 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); |
| 720 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja); |
| 721 | if (!new_shadow_node) { |
| 722 | free(new_node); |
| 723 | return -ENOMEM; |
| 724 | } |
| 725 | if (fallback) |
| 726 | new_shadow_node->fallback_removal_count = |
| 727 | JA_FALLBACK_REMOVAL_COUNT; |
| 728 | } else { |
| 729 | new_node = NULL; |
| 730 | new_node_flag = NULL; |
| 731 | } |
| 732 | |
| 733 | assert(mode != JA_RECOMPACT_ADD || old_type->type_class != RCU_JA_PIGEON); |
| 734 | |
| 735 | if (new_type_index == NODE_INDEX_NULL) |
| 736 | goto skip_copy; |
| 737 | |
| 738 | switch (old_type->type_class) { |
| 739 | case RCU_JA_LINEAR: |
| 740 | { |
| 741 | uint8_t nr_child = |
| 742 | ja_linear_node_get_nr_child(old_type, old_node); |
| 743 | unsigned int i; |
| 744 | |
| 745 | for (i = 0; i < nr_child; i++) { |
| 746 | struct cds_ja_inode_flag *iter; |
| 747 | uint8_t v; |
| 748 | |
| 749 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); |
| 750 | if (!iter) |
| 751 | continue; |
| 752 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 753 | continue; |
| 754 | ret = _ja_node_set_nth(new_type, new_node, |
| 755 | new_shadow_node, |
| 756 | v, iter); |
| 757 | if (new_type->type_class == RCU_JA_POOL && ret) { |
| 758 | goto fallback_toosmall; |
| 759 | } |
| 760 | assert(!ret); |
| 761 | } |
| 762 | break; |
| 763 | } |
| 764 | case RCU_JA_POOL: |
| 765 | { |
| 766 | unsigned int pool_nr; |
| 767 | |
| 768 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { |
| 769 | struct cds_ja_inode *pool = |
| 770 | ja_pool_node_get_ith_pool(old_type, |
| 771 | old_node, pool_nr); |
| 772 | uint8_t nr_child = |
| 773 | ja_linear_node_get_nr_child(old_type, pool); |
| 774 | unsigned int j; |
| 775 | |
| 776 | for (j = 0; j < nr_child; j++) { |
| 777 | struct cds_ja_inode_flag *iter; |
| 778 | uint8_t v; |
| 779 | |
| 780 | ja_linear_node_get_ith_pos(old_type, pool, |
| 781 | j, &v, &iter); |
| 782 | if (!iter) |
| 783 | continue; |
| 784 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 785 | continue; |
| 786 | ret = _ja_node_set_nth(new_type, new_node, |
| 787 | new_shadow_node, |
| 788 | v, iter); |
| 789 | if (new_type->type_class == RCU_JA_POOL |
| 790 | && ret) { |
| 791 | goto fallback_toosmall; |
| 792 | } |
| 793 | assert(!ret); |
| 794 | } |
| 795 | } |
| 796 | break; |
| 797 | } |
| 798 | case RCU_JA_NULL: |
| 799 | assert(mode == JA_RECOMPACT_ADD); |
| 800 | break; |
| 801 | case RCU_JA_PIGEON: |
| 802 | { |
| 803 | uint8_t nr_child; |
| 804 | unsigned int i; |
| 805 | |
| 806 | assert(mode == JA_RECOMPACT_DEL); |
| 807 | nr_child = shadow_node->nr_child; |
| 808 | for (i = 0; i < nr_child; i++) { |
| 809 | struct cds_ja_inode_flag *iter; |
| 810 | |
| 811 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); |
| 812 | if (!iter) |
| 813 | continue; |
| 814 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 815 | continue; |
| 816 | ret = _ja_node_set_nth(new_type, new_node, |
| 817 | new_shadow_node, |
| 818 | i, iter); |
| 819 | if (new_type->type_class == RCU_JA_POOL && ret) { |
| 820 | goto fallback_toosmall; |
| 821 | } |
| 822 | assert(!ret); |
| 823 | } |
| 824 | break; |
| 825 | } |
| 826 | default: |
| 827 | assert(0); |
| 828 | ret = -EINVAL; |
| 829 | goto end; |
| 830 | } |
| 831 | skip_copy: |
| 832 | |
| 833 | if (mode == JA_RECOMPACT_ADD) { |
| 834 | /* add node */ |
| 835 | ret = _ja_node_set_nth(new_type, new_node, |
| 836 | new_shadow_node, |
| 837 | n, child_node_flag); |
| 838 | assert(!ret); |
| 839 | } |
| 840 | /* Return pointer to new recompacted node through old_node_flag_ptr */ |
| 841 | *old_node_flag_ptr = new_node_flag; |
| 842 | if (old_node) { |
| 843 | int flags; |
| 844 | |
| 845 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; |
| 846 | /* |
| 847 | * It is OK to free the lock associated with a node |
| 848 | * going to NULL, since we are holding the parent lock. |
| 849 | * This synchronizes removal with re-add of that node. |
| 850 | */ |
| 851 | if (new_type_index == NODE_INDEX_NULL) |
| 852 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; |
| 853 | ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, |
| 854 | flags); |
| 855 | assert(!ret); |
| 856 | } |
| 857 | |
| 858 | ret = 0; |
| 859 | end: |
| 860 | return ret; |
| 861 | |
| 862 | fallback_toosmall: |
| 863 | /* fallback if next pool is too small */ |
| 864 | assert(new_shadow_node); |
| 865 | ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, |
| 866 | RCUJA_SHADOW_CLEAR_FREE_NODE); |
| 867 | assert(!ret); |
| 868 | |
| 869 | /* Choose fallback type: pigeon */ |
| 870 | new_type_index = (1UL << JA_TYPE_BITS) - 1; |
| 871 | dbg_printf("Fallback to type %d\n", new_type_index); |
| 872 | uatomic_inc(&ja->nr_fallback); |
| 873 | fallback = 1; |
| 874 | goto retry; |
| 875 | } |
| 876 | |
| 877 | /* |
| 878 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
| 879 | * error value otherwise. |
| 880 | */ |
| 881 | static |
| 882 | int ja_node_set_nth(struct cds_ja *ja, |
| 883 | struct cds_ja_inode_flag **node_flag, uint8_t n, |
| 884 | struct cds_ja_inode_flag *child_node_flag, |
| 885 | struct cds_ja_shadow_node *shadow_node) |
| 886 | { |
| 887 | int ret; |
| 888 | unsigned int type_index; |
| 889 | const struct cds_ja_type *type; |
| 890 | struct cds_ja_inode *node; |
| 891 | |
| 892 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", |
| 893 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); |
| 894 | |
| 895 | node = ja_node_ptr(*node_flag); |
| 896 | type_index = ja_node_type(*node_flag); |
| 897 | type = &ja_types[type_index]; |
| 898 | ret = _ja_node_set_nth(type, node, shadow_node, |
| 899 | n, child_node_flag); |
| 900 | switch (ret) { |
| 901 | case -ENOSPC: |
| 902 | /* Not enough space in node, need to recompact. */ |
| 903 | ret = ja_node_recompact(JA_RECOMPACT_ADD, ja, type_index, type, node, |
| 904 | shadow_node, node_flag, n, child_node_flag, NULL); |
| 905 | break; |
| 906 | case -ERANGE: |
| 907 | /* Node needs to be recompacted. */ |
| 908 | ret = ja_node_recompact(JA_RECOMPACT, ja, type_index, type, node, |
| 909 | shadow_node, node_flag, n, child_node_flag, NULL); |
| 910 | break; |
| 911 | } |
| 912 | return ret; |
| 913 | } |
| 914 | |
| 915 | /* |
| 916 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
| 917 | * error value otherwise. |
| 918 | */ |
| 919 | static |
| 920 | int ja_node_clear_ptr(struct cds_ja *ja, |
| 921 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ |
| 922 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ |
| 923 | struct cds_ja_shadow_node *shadow_node, /* of parent */ |
| 924 | uint8_t n) |
| 925 | { |
| 926 | int ret; |
| 927 | unsigned int type_index; |
| 928 | const struct cds_ja_type *type; |
| 929 | struct cds_ja_inode *node; |
| 930 | |
| 931 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", |
| 932 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); |
| 933 | |
| 934 | node = ja_node_ptr(*parent_node_flag_ptr); |
| 935 | type_index = ja_node_type(*parent_node_flag_ptr); |
| 936 | type = &ja_types[type_index]; |
| 937 | ret = _ja_node_clear_ptr(type, node, shadow_node, node_flag_ptr, n); |
| 938 | if (ret == -EFBIG) { |
| 939 | /* Should to try recompaction. */ |
| 940 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, |
| 941 | shadow_node, parent_node_flag_ptr, n, NULL, |
| 942 | node_flag_ptr); |
| 943 | } |
| 944 | return ret; |
| 945 | } |
| 946 | |
| 947 | struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key) |
| 948 | { |
| 949 | unsigned int tree_depth, i; |
| 950 | struct cds_ja_inode_flag *node_flag; |
| 951 | struct cds_hlist_head head = { NULL }; |
| 952 | |
| 953 | if (caa_unlikely(key > ja->key_max)) |
| 954 | return head; |
| 955 | tree_depth = ja->tree_depth; |
| 956 | node_flag = rcu_dereference(ja->root); |
| 957 | |
| 958 | /* level 0: root node */ |
| 959 | if (!ja_node_ptr(node_flag)) |
| 960 | return head; |
| 961 | |
| 962 | for (i = 1; i < tree_depth; i++) { |
| 963 | uint8_t iter_key; |
| 964 | |
| 965 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
| 966 | node_flag = ja_node_get_nth(node_flag, NULL, NULL, NULL, |
| 967 | iter_key); |
| 968 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", |
| 969 | (unsigned int) iter_key, node_flag); |
| 970 | if (!ja_node_ptr(node_flag)) |
| 971 | return head; |
| 972 | } |
| 973 | |
| 974 | /* Last level lookup succeded. We got an actual match. */ |
| 975 | head.next = (struct cds_hlist_node *) node_flag; |
| 976 | return head; |
| 977 | } |
| 978 | |
| 979 | /* |
| 980 | * We reached an unpopulated node. Create it and the children we need, |
| 981 | * and then attach the entire branch to the current node. This may |
| 982 | * trigger recompaction of the current node. Locks needed: node lock |
| 983 | * (for add), and, possibly, parent node lock (to update pointer due to |
| 984 | * node recompaction). |
| 985 | * |
| 986 | * First take node lock, check if recompaction is needed, then take |
| 987 | * parent lock (if needed). Then we can proceed to create the new |
| 988 | * branch. Publish the new branch, and release locks. |
| 989 | * TODO: we currently always take the parent lock even when not needed. |
| 990 | */ |
| 991 | static |
| 992 | int ja_attach_node(struct cds_ja *ja, |
| 993 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
| 994 | struct cds_ja_inode_flag *attach_node_flag, |
| 995 | struct cds_ja_inode_flag **node_flag_ptr, |
| 996 | struct cds_ja_inode_flag *node_flag, |
| 997 | struct cds_ja_inode_flag *parent_node_flag, |
| 998 | uint64_t key, |
| 999 | unsigned int level, |
| 1000 | struct cds_ja_node *child_node) |
| 1001 | { |
| 1002 | struct cds_ja_shadow_node *shadow_node = NULL, |
| 1003 | *parent_shadow_node = NULL; |
| 1004 | struct cds_ja_inode *node = ja_node_ptr(node_flag); |
| 1005 | struct cds_ja_inode *parent_node = ja_node_ptr(parent_node_flag); |
| 1006 | struct cds_hlist_head head; |
| 1007 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; |
| 1008 | int ret, i; |
| 1009 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; |
| 1010 | int nr_created_nodes = 0; |
| 1011 | |
| 1012 | dbg_printf("Attach node at level %u (node %p, node_flag %p)\n", |
| 1013 | level, node, node_flag); |
| 1014 | |
| 1015 | assert(node); |
| 1016 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, node_flag); |
| 1017 | if (!shadow_node) { |
| 1018 | ret = -EAGAIN; |
| 1019 | goto end; |
| 1020 | } |
| 1021 | if (parent_node) { |
| 1022 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
| 1023 | parent_node_flag); |
| 1024 | if (!parent_shadow_node) { |
| 1025 | ret = -EAGAIN; |
| 1026 | goto unlock_shadow; |
| 1027 | } |
| 1028 | } |
| 1029 | |
| 1030 | if (node_flag_ptr && ja_node_ptr(*node_flag_ptr)) { |
| 1031 | /* |
| 1032 | * Target node has been updated between RCU lookup and |
| 1033 | * lock acquisition. We need to re-try lookup and |
| 1034 | * attach. |
| 1035 | */ |
| 1036 | ret = -EAGAIN; |
| 1037 | goto unlock_parent; |
| 1038 | } |
| 1039 | |
| 1040 | if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != |
| 1041 | ja_node_ptr(attach_node_flag)) { |
| 1042 | /* |
| 1043 | * Target node has been updated between RCU lookup and |
| 1044 | * lock acquisition. We need to re-try lookup and |
| 1045 | * attach. |
| 1046 | */ |
| 1047 | ret = -EAGAIN; |
| 1048 | goto unlock_parent; |
| 1049 | } |
| 1050 | |
| 1051 | /* Create new branch, starting from bottom */ |
| 1052 | CDS_INIT_HLIST_HEAD(&head); |
| 1053 | cds_hlist_add_head_rcu(&child_node->list, &head); |
| 1054 | iter_node_flag = (struct cds_ja_inode_flag *) head.next; |
| 1055 | |
| 1056 | for (i = ja->tree_depth; i > (int) level; i--) { |
| 1057 | uint8_t iter_key; |
| 1058 | |
| 1059 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i))); |
| 1060 | dbg_printf("branch creation level %d, key %u\n", |
| 1061 | i - 1, (unsigned int) iter_key); |
| 1062 | iter_dest_node_flag = NULL; |
| 1063 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, |
| 1064 | iter_key, |
| 1065 | iter_node_flag, |
| 1066 | NULL); |
| 1067 | if (ret) |
| 1068 | goto check_error; |
| 1069 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; |
| 1070 | iter_node_flag = iter_dest_node_flag; |
| 1071 | } |
| 1072 | |
| 1073 | if (level > 1) { |
| 1074 | uint8_t iter_key; |
| 1075 | |
| 1076 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); |
| 1077 | /* We need to use set_nth on the previous level. */ |
| 1078 | iter_dest_node_flag = node_flag; |
| 1079 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, |
| 1080 | iter_key, |
| 1081 | iter_node_flag, |
| 1082 | shadow_node); |
| 1083 | if (ret) |
| 1084 | goto check_error; |
| 1085 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; |
| 1086 | iter_node_flag = iter_dest_node_flag; |
| 1087 | } |
| 1088 | |
| 1089 | /* Publish new branch */ |
| 1090 | dbg_printf("Publish branch %p, replacing %p\n", |
| 1091 | iter_node_flag, *attach_node_flag_ptr); |
| 1092 | rcu_assign_pointer(*attach_node_flag_ptr, iter_node_flag); |
| 1093 | |
| 1094 | /* Success */ |
| 1095 | ret = 0; |
| 1096 | |
| 1097 | check_error: |
| 1098 | if (ret) { |
| 1099 | for (i = 0; i < nr_created_nodes; i++) { |
| 1100 | int tmpret; |
| 1101 | int flags; |
| 1102 | |
| 1103 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; |
| 1104 | if (i) |
| 1105 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; |
| 1106 | tmpret = rcuja_shadow_clear(ja->ht, |
| 1107 | created_nodes[i], |
| 1108 | NULL, |
| 1109 | flags); |
| 1110 | assert(!tmpret); |
| 1111 | } |
| 1112 | } |
| 1113 | unlock_parent: |
| 1114 | if (parent_shadow_node) |
| 1115 | rcuja_shadow_unlock(parent_shadow_node); |
| 1116 | unlock_shadow: |
| 1117 | if (shadow_node) |
| 1118 | rcuja_shadow_unlock(shadow_node); |
| 1119 | end: |
| 1120 | return ret; |
| 1121 | } |
| 1122 | |
| 1123 | /* |
| 1124 | * Lock the parent containing the hlist head pointer, and add node to list of |
| 1125 | * duplicates. Failure can happen if concurrent update changes the |
| 1126 | * parent before we get the lock. We return -EAGAIN in that case. |
| 1127 | * Return 0 on success, negative error value on failure. |
| 1128 | */ |
| 1129 | static |
| 1130 | int ja_chain_node(struct cds_ja *ja, |
| 1131 | struct cds_ja_inode_flag *parent_node_flag, |
| 1132 | struct cds_ja_inode_flag **node_flag_ptr, |
| 1133 | struct cds_ja_inode_flag *node_flag, |
| 1134 | struct cds_hlist_head *head, |
| 1135 | struct cds_ja_node *node) |
| 1136 | { |
| 1137 | struct cds_ja_shadow_node *shadow_node; |
| 1138 | int ret = 0; |
| 1139 | |
| 1140 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
| 1141 | if (!shadow_node) { |
| 1142 | return -EAGAIN; |
| 1143 | } |
| 1144 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
| 1145 | ret = -EAGAIN; |
| 1146 | goto end; |
| 1147 | } |
| 1148 | cds_hlist_add_head_rcu(&node->list, head); |
| 1149 | end: |
| 1150 | rcuja_shadow_unlock(shadow_node); |
| 1151 | return ret; |
| 1152 | } |
| 1153 | |
| 1154 | int cds_ja_add(struct cds_ja *ja, uint64_t key, |
| 1155 | struct cds_ja_node *new_node) |
| 1156 | { |
| 1157 | unsigned int tree_depth, i; |
| 1158 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
| 1159 | **node_flag_ptr; |
| 1160 | struct cds_ja_inode_flag *node_flag, |
| 1161 | *parent_node_flag, |
| 1162 | *parent2_node_flag, |
| 1163 | *attach_node_flag; |
| 1164 | int ret; |
| 1165 | |
| 1166 | if (caa_unlikely(key > ja->key_max)) { |
| 1167 | return -EINVAL; |
| 1168 | } |
| 1169 | tree_depth = ja->tree_depth; |
| 1170 | |
| 1171 | retry: |
| 1172 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", |
| 1173 | key, new_node); |
| 1174 | parent2_node_flag = NULL; |
| 1175 | parent_node_flag = |
| 1176 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ |
| 1177 | attach_node_flag_ptr = &ja->root; |
| 1178 | attach_node_flag = rcu_dereference(ja->root); |
| 1179 | node_flag_ptr = &ja->root; |
| 1180 | node_flag = rcu_dereference(ja->root); |
| 1181 | |
| 1182 | /* Iterate on all internal levels */ |
| 1183 | for (i = 1; i < tree_depth; i++) { |
| 1184 | uint8_t iter_key; |
| 1185 | |
| 1186 | dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
| 1187 | attach_node_flag_ptr, node_flag_ptr, node_flag); |
| 1188 | if (!ja_node_ptr(node_flag)) { |
| 1189 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
| 1190 | attach_node_flag, |
| 1191 | node_flag_ptr, |
| 1192 | parent_node_flag, |
| 1193 | parent2_node_flag, |
| 1194 | key, i, new_node); |
| 1195 | if (ret == -EAGAIN || ret == -EEXIST) |
| 1196 | goto retry; |
| 1197 | else |
| 1198 | goto end; |
| 1199 | } |
| 1200 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
| 1201 | parent2_node_flag = parent_node_flag; |
| 1202 | parent_node_flag = node_flag; |
| 1203 | node_flag = ja_node_get_nth(node_flag, |
| 1204 | &attach_node_flag_ptr, |
| 1205 | &attach_node_flag, |
| 1206 | &node_flag_ptr, |
| 1207 | iter_key); |
| 1208 | dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n", |
| 1209 | (unsigned int) iter_key, node_flag, |
| 1210 | attach_node_flag_ptr, |
| 1211 | node_flag_ptr); |
| 1212 | } |
| 1213 | |
| 1214 | /* |
| 1215 | * We reached bottom of tree, simply add node to last internal |
| 1216 | * level, or chain it if key is already present. |
| 1217 | */ |
| 1218 | if (!ja_node_ptr(node_flag)) { |
| 1219 | dbg_printf("cds_ja_add attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
| 1220 | attach_node_flag_ptr, node_flag_ptr, node_flag); |
| 1221 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
| 1222 | attach_node_flag, |
| 1223 | node_flag_ptr, parent_node_flag, |
| 1224 | parent2_node_flag, key, i, new_node); |
| 1225 | } else { |
| 1226 | ret = ja_chain_node(ja, |
| 1227 | parent_node_flag, |
| 1228 | node_flag_ptr, |
| 1229 | node_flag, |
| 1230 | (struct cds_hlist_head *) attach_node_flag_ptr, |
| 1231 | new_node); |
| 1232 | } |
| 1233 | if (ret == -EAGAIN || ret == -EEXIST) |
| 1234 | goto retry; |
| 1235 | end: |
| 1236 | return ret; |
| 1237 | } |
| 1238 | |
| 1239 | /* |
| 1240 | * Note: there is no need to lookup the pointer address associated with |
| 1241 | * each node's nth item after taking the lock: it's already been done by |
| 1242 | * cds_ja_del while holding the rcu read-side lock, and our node rules |
| 1243 | * ensure that when a match value -> pointer is found in a node, it is |
| 1244 | * _NEVER_ changed for that node without recompaction, and recompaction |
| 1245 | * reallocates the node. |
| 1246 | * However, when a child is removed from "linear" nodes, its pointer |
| 1247 | * is set to NULL. We therefore check, while holding the locks, if this |
| 1248 | * pointer is NULL, and return -ENOENT to the caller if it is the case. |
| 1249 | */ |
| 1250 | static |
| 1251 | int ja_detach_node(struct cds_ja *ja, |
| 1252 | struct cds_ja_inode_flag **snapshot, |
| 1253 | struct cds_ja_inode_flag ***snapshot_ptr, |
| 1254 | uint8_t *snapshot_n, |
| 1255 | int nr_snapshot, |
| 1256 | uint64_t key, |
| 1257 | struct cds_ja_node *node) |
| 1258 | { |
| 1259 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; |
| 1260 | struct cds_ja_inode_flag **node_flag_ptr = NULL, |
| 1261 | *parent_node_flag = NULL, |
| 1262 | **parent_node_flag_ptr = NULL; |
| 1263 | struct cds_ja_inode_flag *iter_node_flag; |
| 1264 | int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; |
| 1265 | uint8_t n = 0; |
| 1266 | |
| 1267 | assert(nr_snapshot == ja->tree_depth + 1); |
| 1268 | |
| 1269 | /* |
| 1270 | * From the last internal level node going up, get the node |
| 1271 | * lock, check if the node has only one child left. If it is the |
| 1272 | * case, we continue iterating upward. When we reach a node |
| 1273 | * which has more that one child left, we lock the parent, and |
| 1274 | * proceed to the node deletion (removing its children too). |
| 1275 | */ |
| 1276 | for (i = nr_snapshot - 2; i >= 1; i--) { |
| 1277 | struct cds_ja_shadow_node *shadow_node; |
| 1278 | |
| 1279 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
| 1280 | snapshot[i]); |
| 1281 | if (!shadow_node) { |
| 1282 | ret = -EAGAIN; |
| 1283 | goto end; |
| 1284 | } |
| 1285 | shadow_nodes[nr_shadow++] = shadow_node; |
| 1286 | |
| 1287 | /* |
| 1288 | * Check if node has been removed between RCU |
| 1289 | * lookup and lock acquisition. |
| 1290 | */ |
| 1291 | assert(snapshot_ptr[i + 1]); |
| 1292 | if (ja_node_ptr(*snapshot_ptr[i + 1]) |
| 1293 | != ja_node_ptr(snapshot[i + 1])) { |
| 1294 | ret = -ENOENT; |
| 1295 | goto end; |
| 1296 | } |
| 1297 | |
| 1298 | assert(shadow_node->nr_child > 0); |
| 1299 | if (shadow_node->nr_child == 1 && i > 1) |
| 1300 | nr_clear++; |
| 1301 | nr_branch++; |
| 1302 | if (shadow_node->nr_child > 1 || i == 1) { |
| 1303 | /* Lock parent and break */ |
| 1304 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
| 1305 | snapshot[i - 1]); |
| 1306 | if (!shadow_node) { |
| 1307 | ret = -EAGAIN; |
| 1308 | goto end; |
| 1309 | } |
| 1310 | shadow_nodes[nr_shadow++] = shadow_node; |
| 1311 | |
| 1312 | /* |
| 1313 | * Check if node has been removed between RCU |
| 1314 | * lookup and lock acquisition. |
| 1315 | */ |
| 1316 | assert(snapshot_ptr[i]); |
| 1317 | if (ja_node_ptr(*snapshot_ptr[i]) |
| 1318 | != ja_node_ptr(snapshot[i])) { |
| 1319 | ret = -ENOENT; |
| 1320 | goto end; |
| 1321 | } |
| 1322 | |
| 1323 | node_flag_ptr = snapshot_ptr[i + 1]; |
| 1324 | n = snapshot_n[i + 1]; |
| 1325 | parent_node_flag_ptr = snapshot_ptr[i]; |
| 1326 | parent_node_flag = snapshot[i]; |
| 1327 | |
| 1328 | if (i > 1) { |
| 1329 | /* |
| 1330 | * Lock parent's parent, in case we need |
| 1331 | * to recompact parent. |
| 1332 | */ |
| 1333 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
| 1334 | snapshot[i - 2]); |
| 1335 | if (!shadow_node) { |
| 1336 | ret = -EAGAIN; |
| 1337 | goto end; |
| 1338 | } |
| 1339 | shadow_nodes[nr_shadow++] = shadow_node; |
| 1340 | |
| 1341 | /* |
| 1342 | * Check if node has been removed between RCU |
| 1343 | * lookup and lock acquisition. |
| 1344 | */ |
| 1345 | assert(snapshot_ptr[i - 1]); |
| 1346 | if (ja_node_ptr(*snapshot_ptr[i - 1]) |
| 1347 | != ja_node_ptr(snapshot[i - 1])) { |
| 1348 | ret = -ENOENT; |
| 1349 | goto end; |
| 1350 | } |
| 1351 | } |
| 1352 | |
| 1353 | break; |
| 1354 | } |
| 1355 | } |
| 1356 | |
| 1357 | /* |
| 1358 | * At this point, we want to delete all nodes that are about to |
| 1359 | * be removed from shadow_nodes (except the last one, which is |
| 1360 | * either the root or the parent of the upmost node with 1 |
| 1361 | * child). OK to free lock here, because RCU read lock is held, |
| 1362 | * and free only performed in call_rcu. |
| 1363 | */ |
| 1364 | |
| 1365 | for (i = 0; i < nr_clear; i++) { |
| 1366 | ret = rcuja_shadow_clear(ja->ht, |
| 1367 | shadow_nodes[i]->node_flag, |
| 1368 | shadow_nodes[i], |
| 1369 | RCUJA_SHADOW_CLEAR_FREE_NODE |
| 1370 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); |
| 1371 | assert(!ret); |
| 1372 | } |
| 1373 | |
| 1374 | iter_node_flag = parent_node_flag; |
| 1375 | /* Remove from parent */ |
| 1376 | ret = ja_node_clear_ptr(ja, |
| 1377 | node_flag_ptr, /* Pointer to location to nullify */ |
| 1378 | &iter_node_flag, /* Old new parent ptr in its parent */ |
| 1379 | shadow_nodes[nr_branch - 1], /* of parent */ |
| 1380 | n); |
| 1381 | if (ret) |
| 1382 | goto end; |
| 1383 | |
| 1384 | dbg_printf("ja_detach_node: publish %p instead of %p\n", |
| 1385 | iter_node_flag, *parent_node_flag_ptr); |
| 1386 | /* Update address of parent ptr in its parent */ |
| 1387 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); |
| 1388 | |
| 1389 | end: |
| 1390 | for (i = 0; i < nr_shadow; i++) |
| 1391 | rcuja_shadow_unlock(shadow_nodes[i]); |
| 1392 | return ret; |
| 1393 | } |
| 1394 | |
| 1395 | static |
| 1396 | int ja_unchain_node(struct cds_ja *ja, |
| 1397 | struct cds_ja_inode_flag *parent_node_flag, |
| 1398 | struct cds_ja_inode_flag **node_flag_ptr, |
| 1399 | struct cds_ja_inode_flag *node_flag, |
| 1400 | struct cds_ja_node *node) |
| 1401 | { |
| 1402 | struct cds_ja_shadow_node *shadow_node; |
| 1403 | struct cds_hlist_node *hlist_node; |
| 1404 | struct cds_hlist_head hlist_head; |
| 1405 | int ret = 0, count = 0, found = 0; |
| 1406 | |
| 1407 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
| 1408 | if (!shadow_node) |
| 1409 | return -EAGAIN; |
| 1410 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
| 1411 | ret = -EAGAIN; |
| 1412 | goto end; |
| 1413 | } |
| 1414 | hlist_head.next = (struct cds_hlist_node *) ja_node_ptr(node_flag); |
| 1415 | /* |
| 1416 | * Retry if another thread removed all but one of duplicates |
| 1417 | * since check (this check was performed without lock). |
| 1418 | * Ensure that the node we are about to remove is still in the |
| 1419 | * list (while holding lock). |
| 1420 | */ |
| 1421 | cds_hlist_for_each_rcu(hlist_node, &hlist_head) { |
| 1422 | if (count == 0) { |
| 1423 | /* FIXME: currently a work-around */ |
| 1424 | hlist_node->prev = (struct cds_hlist_node *) node_flag_ptr; |
| 1425 | } |
| 1426 | count++; |
| 1427 | if (hlist_node == &node->list) |
| 1428 | found++; |
| 1429 | } |
| 1430 | assert(found <= 1); |
| 1431 | if (!found || count == 1) { |
| 1432 | ret = -EAGAIN; |
| 1433 | goto end; |
| 1434 | } |
| 1435 | cds_hlist_del_rcu(&node->list); |
| 1436 | /* |
| 1437 | * Validate that we indeed removed the node from linked list. |
| 1438 | */ |
| 1439 | assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node); |
| 1440 | end: |
| 1441 | rcuja_shadow_unlock(shadow_node); |
| 1442 | return ret; |
| 1443 | } |
| 1444 | |
| 1445 | /* |
| 1446 | * Called with RCU read lock held. |
| 1447 | */ |
| 1448 | int cds_ja_del(struct cds_ja *ja, uint64_t key, |
| 1449 | struct cds_ja_node *node) |
| 1450 | { |
| 1451 | unsigned int tree_depth, i; |
| 1452 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; |
| 1453 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; |
| 1454 | uint8_t snapshot_n[JA_MAX_DEPTH]; |
| 1455 | struct cds_ja_inode_flag *node_flag; |
| 1456 | struct cds_ja_inode_flag **prev_node_flag_ptr, |
| 1457 | **node_flag_ptr; |
| 1458 | int nr_snapshot; |
| 1459 | int ret; |
| 1460 | |
| 1461 | if (caa_unlikely(key > ja->key_max)) |
| 1462 | return -EINVAL; |
| 1463 | tree_depth = ja->tree_depth; |
| 1464 | |
| 1465 | retry: |
| 1466 | nr_snapshot = 0; |
| 1467 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", |
| 1468 | key, node); |
| 1469 | |
| 1470 | /* snapshot for level 0 is only for shadow node lookup */ |
| 1471 | snapshot_n[0] = 0; |
| 1472 | snapshot_n[1] = 0; |
| 1473 | snapshot_ptr[nr_snapshot] = NULL; |
| 1474 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; |
| 1475 | node_flag = rcu_dereference(ja->root); |
| 1476 | prev_node_flag_ptr = &ja->root; |
| 1477 | node_flag_ptr = &ja->root; |
| 1478 | |
| 1479 | /* Iterate on all internal levels */ |
| 1480 | for (i = 1; i < tree_depth; i++) { |
| 1481 | uint8_t iter_key; |
| 1482 | |
| 1483 | dbg_printf("cds_ja_del iter node_flag %p\n", |
| 1484 | node_flag); |
| 1485 | if (!ja_node_ptr(node_flag)) { |
| 1486 | return -ENOENT; |
| 1487 | } |
| 1488 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
| 1489 | snapshot_n[nr_snapshot + 1] = iter_key; |
| 1490 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
| 1491 | snapshot[nr_snapshot++] = node_flag; |
| 1492 | node_flag = ja_node_get_nth(node_flag, |
| 1493 | &prev_node_flag_ptr, |
| 1494 | NULL, |
| 1495 | &node_flag_ptr, |
| 1496 | iter_key); |
| 1497 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", |
| 1498 | (unsigned int) iter_key, node_flag, |
| 1499 | prev_node_flag_ptr); |
| 1500 | } |
| 1501 | /* |
| 1502 | * We reached bottom of tree, try to find the node we are trying |
| 1503 | * to remove. Fail if we cannot find it. |
| 1504 | */ |
| 1505 | if (!ja_node_ptr(node_flag)) { |
| 1506 | dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", |
| 1507 | key); |
| 1508 | return -ENOENT; |
| 1509 | } else { |
| 1510 | struct cds_hlist_head hlist_head; |
| 1511 | struct cds_hlist_node *hlist_node; |
| 1512 | struct cds_ja_node *entry, *match = NULL; |
| 1513 | int count = 0; |
| 1514 | |
| 1515 | hlist_head.next = |
| 1516 | (struct cds_hlist_node *) ja_node_ptr(node_flag); |
| 1517 | cds_hlist_for_each_entry_rcu(entry, |
| 1518 | hlist_node, |
| 1519 | &hlist_head, |
| 1520 | list) { |
| 1521 | dbg_printf("cds_ja_del: compare %p with entry %p\n", node, entry); |
| 1522 | if (entry == node) |
| 1523 | match = entry; |
| 1524 | count++; |
| 1525 | } |
| 1526 | if (!match) { |
| 1527 | dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); |
| 1528 | return -ENOENT; |
| 1529 | } |
| 1530 | assert(count > 0); |
| 1531 | if (count == 1) { |
| 1532 | /* |
| 1533 | * Removing last of duplicates. Last snapshot |
| 1534 | * does not have a shadow node (external leafs). |
| 1535 | */ |
| 1536 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
| 1537 | snapshot[nr_snapshot++] = node_flag; |
| 1538 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, |
| 1539 | snapshot_n, nr_snapshot, key, node); |
| 1540 | } else { |
| 1541 | ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], |
| 1542 | node_flag_ptr, node_flag, match); |
| 1543 | } |
| 1544 | } |
| 1545 | /* |
| 1546 | * Explanation of -ENOENT handling: caused by concurrent delete |
| 1547 | * between RCU lookup and actual removal. Need to re-do the |
| 1548 | * lookup and removal attempt. |
| 1549 | */ |
| 1550 | if (ret == -EAGAIN || ret == -ENOENT) |
| 1551 | goto retry; |
| 1552 | return ret; |
| 1553 | } |
| 1554 | |
| 1555 | struct cds_ja *_cds_ja_new(unsigned int key_bits, |
| 1556 | const struct rcu_flavor_struct *flavor) |
| 1557 | { |
| 1558 | struct cds_ja *ja; |
| 1559 | int ret; |
| 1560 | struct cds_ja_shadow_node *root_shadow_node; |
| 1561 | |
| 1562 | ja = calloc(sizeof(*ja), 1); |
| 1563 | if (!ja) |
| 1564 | goto ja_error; |
| 1565 | |
| 1566 | switch (key_bits) { |
| 1567 | case 8: |
| 1568 | case 16: |
| 1569 | case 24: |
| 1570 | case 32: |
| 1571 | case 40: |
| 1572 | case 48: |
| 1573 | case 56: |
| 1574 | ja->key_max = (1ULL << key_bits) - 1; |
| 1575 | break; |
| 1576 | case 64: |
| 1577 | ja->key_max = UINT64_MAX; |
| 1578 | break; |
| 1579 | default: |
| 1580 | goto check_error; |
| 1581 | } |
| 1582 | |
| 1583 | /* ja->root is NULL */ |
| 1584 | /* tree_depth 0 is for pointer to root node */ |
| 1585 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; |
| 1586 | assert(ja->tree_depth <= JA_MAX_DEPTH); |
| 1587 | ja->ht = rcuja_create_ht(flavor); |
| 1588 | if (!ja->ht) |
| 1589 | goto ht_error; |
| 1590 | |
| 1591 | /* |
| 1592 | * Note: we should not free this node until judy array destroy. |
| 1593 | */ |
| 1594 | root_shadow_node = rcuja_shadow_set(ja->ht, |
| 1595 | (struct cds_ja_inode_flag *) &ja->root, |
| 1596 | NULL, ja); |
| 1597 | if (!root_shadow_node) { |
| 1598 | ret = -ENOMEM; |
| 1599 | goto ht_node_error; |
| 1600 | } |
| 1601 | root_shadow_node->level = 0; |
| 1602 | |
| 1603 | return ja; |
| 1604 | |
| 1605 | ht_node_error: |
| 1606 | ret = rcuja_delete_ht(ja->ht); |
| 1607 | assert(!ret); |
| 1608 | ht_error: |
| 1609 | check_error: |
| 1610 | free(ja); |
| 1611 | ja_error: |
| 1612 | return NULL; |
| 1613 | } |
| 1614 | |
| 1615 | /* |
| 1616 | * Called from RCU read-side CS. |
| 1617 | */ |
| 1618 | __attribute__((visibility("protected"))) |
| 1619 | void rcuja_free_all_children(struct cds_ja_shadow_node *shadow_node, |
| 1620 | struct cds_ja_inode_flag *node_flag, |
| 1621 | void (*free_node_cb)(struct rcu_head *head)) |
| 1622 | { |
| 1623 | const struct rcu_flavor_struct *flavor; |
| 1624 | unsigned int type_index; |
| 1625 | struct cds_ja_inode *node; |
| 1626 | const struct cds_ja_type *type; |
| 1627 | |
| 1628 | flavor = cds_lfht_rcu_flavor(shadow_node->ja->ht); |
| 1629 | node = ja_node_ptr(node_flag); |
| 1630 | assert(node != NULL); |
| 1631 | type_index = ja_node_type(node_flag); |
| 1632 | type = &ja_types[type_index]; |
| 1633 | |
| 1634 | switch (type->type_class) { |
| 1635 | case RCU_JA_LINEAR: |
| 1636 | { |
| 1637 | uint8_t nr_child = |
| 1638 | ja_linear_node_get_nr_child(type, node); |
| 1639 | unsigned int i; |
| 1640 | |
| 1641 | for (i = 0; i < nr_child; i++) { |
| 1642 | struct cds_ja_inode_flag *iter; |
| 1643 | struct cds_hlist_head head; |
| 1644 | struct cds_ja_node *entry; |
| 1645 | struct cds_hlist_node *pos; |
| 1646 | uint8_t v; |
| 1647 | |
| 1648 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); |
| 1649 | if (!iter) |
| 1650 | continue; |
| 1651 | head.next = (struct cds_hlist_node *) iter; |
| 1652 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { |
| 1653 | flavor->update_call_rcu(&entry->head, free_node_cb); |
| 1654 | } |
| 1655 | } |
| 1656 | break; |
| 1657 | } |
| 1658 | case RCU_JA_POOL: |
| 1659 | { |
| 1660 | unsigned int pool_nr; |
| 1661 | |
| 1662 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { |
| 1663 | struct cds_ja_inode *pool = |
| 1664 | ja_pool_node_get_ith_pool(type, node, pool_nr); |
| 1665 | uint8_t nr_child = |
| 1666 | ja_linear_node_get_nr_child(type, pool); |
| 1667 | unsigned int j; |
| 1668 | |
| 1669 | for (j = 0; j < nr_child; j++) { |
| 1670 | struct cds_ja_inode_flag *iter; |
| 1671 | struct cds_hlist_head head; |
| 1672 | struct cds_ja_node *entry; |
| 1673 | struct cds_hlist_node *pos; |
| 1674 | uint8_t v; |
| 1675 | |
| 1676 | ja_linear_node_get_ith_pos(type, node, j, &v, &iter); |
| 1677 | if (!iter) |
| 1678 | continue; |
| 1679 | head.next = (struct cds_hlist_node *) iter; |
| 1680 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { |
| 1681 | flavor->update_call_rcu(&entry->head, free_node_cb); |
| 1682 | } |
| 1683 | } |
| 1684 | } |
| 1685 | break; |
| 1686 | } |
| 1687 | case RCU_JA_NULL: |
| 1688 | break; |
| 1689 | case RCU_JA_PIGEON: |
| 1690 | { |
| 1691 | uint8_t nr_child; |
| 1692 | unsigned int i; |
| 1693 | |
| 1694 | nr_child = shadow_node->nr_child; |
| 1695 | for (i = 0; i < nr_child; i++) { |
| 1696 | struct cds_ja_inode_flag *iter; |
| 1697 | struct cds_hlist_head head; |
| 1698 | struct cds_ja_node *entry; |
| 1699 | struct cds_hlist_node *pos; |
| 1700 | |
| 1701 | iter = ja_pigeon_node_get_ith_pos(type, node, i); |
| 1702 | if (!iter) |
| 1703 | continue; |
| 1704 | head.next = (struct cds_hlist_node *) iter; |
| 1705 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { |
| 1706 | flavor->update_call_rcu(&entry->head, free_node_cb); |
| 1707 | } |
| 1708 | } |
| 1709 | break; |
| 1710 | } |
| 1711 | default: |
| 1712 | assert(0); |
| 1713 | } |
| 1714 | } |
| 1715 | |
| 1716 | /* |
| 1717 | * There should be no more concurrent add to the judy array while it is |
| 1718 | * being destroyed (ensured by the caller). |
| 1719 | */ |
| 1720 | int cds_ja_destroy(struct cds_ja *ja, |
| 1721 | void (*free_node_cb)(struct rcu_head *head)) |
| 1722 | { |
| 1723 | int ret; |
| 1724 | |
| 1725 | rcuja_shadow_prune(ja->ht, |
| 1726 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK, |
| 1727 | free_node_cb); |
| 1728 | ret = rcuja_delete_ht(ja->ht); |
| 1729 | if (ret) |
| 1730 | return ret; |
| 1731 | if (uatomic_read(&ja->nr_fallback)) |
| 1732 | fprintf(stderr, |
| 1733 | "[warning] RCU Judy Array used %lu fallback node(s)\n", |
| 1734 | uatomic_read(&ja->nr_fallback)); |
| 1735 | free(ja); |
| 1736 | return 0; |
| 1737 | } |