| 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 <string.h> |
| 28 | #include <urcu/rcuja.h> |
| 29 | #include <urcu/compiler.h> |
| 30 | #include <urcu/arch.h> |
| 31 | #include <assert.h> |
| 32 | #include <urcu-pointer.h> |
| 33 | #include <urcu/uatomic.h> |
| 34 | #include <stdint.h> |
| 35 | |
| 36 | #include "rcuja-internal.h" |
| 37 | #include "bitfield.h" |
| 38 | |
| 39 | #ifndef abs |
| 40 | #define abs_int(a) ((int) (a) > 0 ? (int) (a) : -((int) (a))) |
| 41 | #endif |
| 42 | |
| 43 | enum cds_ja_type_class { |
| 44 | RCU_JA_LINEAR = 0, /* Type A */ |
| 45 | /* 32-bit: 1 to 25 children, 8 to 128 bytes */ |
| 46 | /* 64-bit: 1 to 28 children, 16 to 256 bytes */ |
| 47 | RCU_JA_POOL = 1, /* Type B */ |
| 48 | /* 32-bit: 26 to 100 children, 256 to 512 bytes */ |
| 49 | /* 64-bit: 29 to 112 children, 512 to 1024 bytes */ |
| 50 | RCU_JA_PIGEON = 2, /* Type C */ |
| 51 | /* 32-bit: 101 to 256 children, 1024 bytes */ |
| 52 | /* 64-bit: 113 to 256 children, 2048 bytes */ |
| 53 | /* Leaf nodes are implicit from their height in the tree */ |
| 54 | RCU_JA_NR_TYPES, |
| 55 | |
| 56 | RCU_JA_NULL, /* not an encoded type, but keeps code regular */ |
| 57 | }; |
| 58 | |
| 59 | struct cds_ja_type { |
| 60 | enum cds_ja_type_class type_class; |
| 61 | uint16_t min_child; /* minimum number of children: 1 to 256 */ |
| 62 | uint16_t max_child; /* maximum number of children: 1 to 256 */ |
| 63 | uint16_t max_linear_child; /* per-pool max nr. children: 1 to 256 */ |
| 64 | uint16_t order; /* node size is (1 << order), in bytes */ |
| 65 | uint16_t nr_pool_order; /* number of pools */ |
| 66 | uint16_t pool_size_order; /* pool size */ |
| 67 | }; |
| 68 | |
| 69 | /* |
| 70 | * Iteration on the array to find the right node size for the number of |
| 71 | * children stops when it reaches .max_child == 256 (this is the largest |
| 72 | * possible node size, which contains 256 children). |
| 73 | * The min_child overlaps with the previous max_child to provide an |
| 74 | * hysteresis loop to reallocation for patterns of cyclic add/removal |
| 75 | * within the same node. |
| 76 | * The node the index within the following arrays is represented on 3 |
| 77 | * bits. It identifies the node type, min/max number of children, and |
| 78 | * the size order. |
| 79 | * The max_child values for the RCU_JA_POOL below result from |
| 80 | * statistical approximation: over million populations, the max_child |
| 81 | * covers between 97% and 99% of the populations generated. Therefore, a |
| 82 | * fallback should exist to cover the rare extreme population unbalance |
| 83 | * cases, but it will not have a major impact on speed nor space |
| 84 | * consumption, since those are rare cases. |
| 85 | */ |
| 86 | |
| 87 | #if (CAA_BITS_PER_LONG < 64) |
| 88 | /* 32-bit pointers */ |
| 89 | enum { |
| 90 | ja_type_0_max_child = 1, |
| 91 | ja_type_1_max_child = 3, |
| 92 | ja_type_2_max_child = 6, |
| 93 | ja_type_3_max_child = 12, |
| 94 | ja_type_4_max_child = 25, |
| 95 | ja_type_5_max_child = 48, |
| 96 | ja_type_6_max_child = 92, |
| 97 | ja_type_7_max_child = 256, |
| 98 | ja_type_8_max_child = 0, /* NULL */ |
| 99 | }; |
| 100 | |
| 101 | enum { |
| 102 | ja_type_0_max_linear_child = 1, |
| 103 | ja_type_1_max_linear_child = 3, |
| 104 | ja_type_2_max_linear_child = 6, |
| 105 | ja_type_3_max_linear_child = 12, |
| 106 | ja_type_4_max_linear_child = 25, |
| 107 | ja_type_5_max_linear_child = 24, |
| 108 | ja_type_6_max_linear_child = 23, |
| 109 | }; |
| 110 | |
| 111 | enum { |
| 112 | ja_type_5_nr_pool_order = 1, |
| 113 | ja_type_6_nr_pool_order = 2, |
| 114 | }; |
| 115 | |
| 116 | const struct cds_ja_type ja_types[] = { |
| 117 | { .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, }, |
| 118 | { .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, }, |
| 119 | { .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, }, |
| 120 | { .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, }, |
| 121 | { .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, }, |
| 122 | |
| 123 | /* Pools may fill sooner than max_child */ |
| 124 | { .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, }, |
| 125 | { .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, }, |
| 126 | |
| 127 | /* |
| 128 | * Upon node removal below min_child, if child pool is filled |
| 129 | * beyond capacity, we roll back to pigeon. |
| 130 | */ |
| 131 | { .type_class = RCU_JA_PIGEON, .min_child = 89, .max_child = ja_type_7_max_child, .order = 10, }, |
| 132 | |
| 133 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, |
| 134 | }; |
| 135 | #else /* !(CAA_BITS_PER_LONG < 64) */ |
| 136 | /* 64-bit pointers */ |
| 137 | enum { |
| 138 | ja_type_0_max_child = 1, |
| 139 | ja_type_1_max_child = 3, |
| 140 | ja_type_2_max_child = 7, |
| 141 | ja_type_3_max_child = 14, |
| 142 | ja_type_4_max_child = 28, |
| 143 | ja_type_5_max_child = 54, |
| 144 | ja_type_6_max_child = 104, |
| 145 | ja_type_7_max_child = 256, |
| 146 | ja_type_8_max_child = 256, |
| 147 | }; |
| 148 | |
| 149 | enum { |
| 150 | ja_type_0_max_linear_child = 1, |
| 151 | ja_type_1_max_linear_child = 3, |
| 152 | ja_type_2_max_linear_child = 7, |
| 153 | ja_type_3_max_linear_child = 14, |
| 154 | ja_type_4_max_linear_child = 28, |
| 155 | ja_type_5_max_linear_child = 27, |
| 156 | ja_type_6_max_linear_child = 26, |
| 157 | }; |
| 158 | |
| 159 | enum { |
| 160 | ja_type_5_nr_pool_order = 1, |
| 161 | ja_type_6_nr_pool_order = 2, |
| 162 | }; |
| 163 | |
| 164 | const struct cds_ja_type ja_types[] = { |
| 165 | { .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, }, |
| 166 | { .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, }, |
| 167 | { .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, }, |
| 168 | { .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, }, |
| 169 | { .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, }, |
| 170 | |
| 171 | /* Pools may fill sooner than max_child. */ |
| 172 | { .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, }, |
| 173 | { .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, }, |
| 174 | |
| 175 | /* |
| 176 | * Upon node removal below min_child, if child pool is filled |
| 177 | * beyond capacity, we roll back to pigeon. |
| 178 | */ |
| 179 | { .type_class = RCU_JA_PIGEON, .min_child = 101, .max_child = ja_type_7_max_child, .order = 11, }, |
| 180 | |
| 181 | { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, |
| 182 | }; |
| 183 | #endif /* !(BITS_PER_LONG < 64) */ |
| 184 | |
| 185 | static inline __attribute__((unused)) |
| 186 | void static_array_size_check(void) |
| 187 | { |
| 188 | CAA_BUILD_BUG_ON(CAA_ARRAY_SIZE(ja_types) < JA_TYPE_MAX_NR); |
| 189 | } |
| 190 | |
| 191 | /* |
| 192 | * The cds_ja_node contains the compressed node data needed for |
| 193 | * read-side. For linear and pool node configurations, it starts with a |
| 194 | * byte counting the number of children in the node. Then, the |
| 195 | * node-specific data is placed. |
| 196 | * The node mutex, if any is needed, protecting concurrent updated of |
| 197 | * each node is placed in a separate hash table indexed by node address. |
| 198 | * For the pigeon configuration, the number of children is also kept in |
| 199 | * a separate hash table, indexed by node address, because it is only |
| 200 | * required for updates. |
| 201 | */ |
| 202 | |
| 203 | #define DECLARE_LINEAR_NODE(index) \ |
| 204 | struct { \ |
| 205 | uint8_t nr_child; \ |
| 206 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ |
| 207 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
| 208 | } |
| 209 | |
| 210 | #define DECLARE_POOL_NODE(index) \ |
| 211 | struct { \ |
| 212 | struct { \ |
| 213 | uint8_t nr_child; \ |
| 214 | uint8_t child_value[ja_type_## index ##_max_linear_child]; \ |
| 215 | struct cds_ja_inode_flag *child_ptr[ja_type_## index ##_max_linear_child]; \ |
| 216 | } linear[1U << ja_type_## index ##_nr_pool_order]; \ |
| 217 | } |
| 218 | |
| 219 | struct cds_ja_inode { |
| 220 | union { |
| 221 | /* Linear configuration */ |
| 222 | DECLARE_LINEAR_NODE(0) conf_0; |
| 223 | DECLARE_LINEAR_NODE(1) conf_1; |
| 224 | DECLARE_LINEAR_NODE(2) conf_2; |
| 225 | DECLARE_LINEAR_NODE(3) conf_3; |
| 226 | DECLARE_LINEAR_NODE(4) conf_4; |
| 227 | |
| 228 | /* Pool configuration */ |
| 229 | DECLARE_POOL_NODE(5) conf_5; |
| 230 | DECLARE_POOL_NODE(6) conf_6; |
| 231 | |
| 232 | /* Pigeon configuration */ |
| 233 | struct { |
| 234 | struct cds_ja_inode_flag *child[ja_type_7_max_child]; |
| 235 | } conf_7; |
| 236 | /* data aliasing nodes for computed accesses */ |
| 237 | uint8_t data[sizeof(struct cds_ja_inode_flag *) * ja_type_7_max_child]; |
| 238 | } u; |
| 239 | }; |
| 240 | |
| 241 | enum ja_recompact { |
| 242 | JA_RECOMPACT_ADD_SAME, |
| 243 | JA_RECOMPACT_ADD_NEXT, |
| 244 | JA_RECOMPACT_DEL, |
| 245 | }; |
| 246 | |
| 247 | static |
| 248 | unsigned long node_fallback_count_distribution[JA_ENTRY_PER_NODE]; |
| 249 | |
| 250 | static |
| 251 | struct cds_ja_inode *_ja_node_mask_ptr(struct cds_ja_inode_flag *node) |
| 252 | { |
| 253 | return (struct cds_ja_inode *) (((unsigned long) node) & JA_PTR_MASK); |
| 254 | } |
| 255 | |
| 256 | unsigned long ja_node_type(struct cds_ja_inode_flag *node) |
| 257 | { |
| 258 | unsigned long type; |
| 259 | |
| 260 | if (_ja_node_mask_ptr(node) == NULL) { |
| 261 | return NODE_INDEX_NULL; |
| 262 | } |
| 263 | type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK); |
| 264 | assert(type < (1UL << JA_TYPE_BITS)); |
| 265 | return type; |
| 266 | } |
| 267 | |
| 268 | struct cds_ja_inode *ja_node_ptr(struct cds_ja_inode_flag *node) |
| 269 | { |
| 270 | unsigned long type_index = ja_node_type(node); |
| 271 | const struct cds_ja_type *type; |
| 272 | |
| 273 | type = &ja_types[type_index]; |
| 274 | switch (type->type_class) { |
| 275 | case RCU_JA_LINEAR: |
| 276 | case RCU_JA_PIGEON: /* fall-through */ |
| 277 | case RCU_JA_NULL: /* fall-through */ |
| 278 | default: /* fall-through */ |
| 279 | return _ja_node_mask_ptr(node); |
| 280 | case RCU_JA_POOL: |
| 281 | switch (type->nr_pool_order) { |
| 282 | case 1: |
| 283 | return (struct cds_ja_inode *) (((unsigned long) node) & ~(JA_POOL_1D_MASK | JA_TYPE_MASK)); |
| 284 | case 2: |
| 285 | return (struct cds_ja_inode *) (((unsigned long) node) & ~(JA_POOL_2D_MASK | JA_POOL_1D_MASK | JA_TYPE_MASK)); |
| 286 | default: |
| 287 | assert(0); |
| 288 | } |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | struct cds_ja_inode *alloc_cds_ja_node(const struct cds_ja_type *ja_type) |
| 293 | { |
| 294 | size_t len = 1U << ja_type->order; |
| 295 | void *p; |
| 296 | int ret; |
| 297 | |
| 298 | ret = posix_memalign(&p, len, len); |
| 299 | if (ret || !p) { |
| 300 | return NULL; |
| 301 | } |
| 302 | memset(p, 0, len); |
| 303 | return p; |
| 304 | } |
| 305 | |
| 306 | void free_cds_ja_node(struct cds_ja_inode *node) |
| 307 | { |
| 308 | free(node); |
| 309 | } |
| 310 | |
| 311 | #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) |
| 312 | #define JA_ALIGN(v, align) __JA_ALIGN_MASK(v, (typeof(v)) (align) - 1) |
| 313 | #define __JA_FLOOR_MASK(v, mask) ((v) & ~(mask)) |
| 314 | #define JA_FLOOR(v, align) __JA_FLOOR_MASK(v, (typeof(v)) (align) - 1) |
| 315 | |
| 316 | static |
| 317 | uint8_t *align_ptr_size(uint8_t *ptr) |
| 318 | { |
| 319 | return (uint8_t *) JA_ALIGN((unsigned long) ptr, sizeof(void *)); |
| 320 | } |
| 321 | |
| 322 | static |
| 323 | uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type, |
| 324 | struct cds_ja_inode *node) |
| 325 | { |
| 326 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 327 | return rcu_dereference(node->u.data[0]); |
| 328 | } |
| 329 | |
| 330 | /* |
| 331 | * The order in which values and pointers are does does not matter: if |
| 332 | * a value is missing, we return NULL. If a value is there, but its |
| 333 | * associated pointers is still NULL, we return NULL too. |
| 334 | */ |
| 335 | static |
| 336 | struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, |
| 337 | struct cds_ja_inode *node, |
| 338 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
| 339 | struct cds_ja_inode_flag **child_node_flag_v, |
| 340 | struct cds_ja_inode_flag ***node_flag_ptr, |
| 341 | uint8_t n) |
| 342 | { |
| 343 | uint8_t nr_child; |
| 344 | uint8_t *values; |
| 345 | struct cds_ja_inode_flag **pointers; |
| 346 | struct cds_ja_inode_flag *ptr; |
| 347 | unsigned int i; |
| 348 | |
| 349 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 350 | |
| 351 | nr_child = ja_linear_node_get_nr_child(type, node); |
| 352 | cmm_smp_rmb(); /* read nr_child before values and pointers */ |
| 353 | assert(nr_child <= type->max_linear_child); |
| 354 | assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child); |
| 355 | |
| 356 | values = &node->u.data[1]; |
| 357 | for (i = 0; i < nr_child; i++) { |
| 358 | if (CMM_LOAD_SHARED(values[i]) == n) |
| 359 | break; |
| 360 | } |
| 361 | if (i >= nr_child) { |
| 362 | if (caa_unlikely(node_flag_ptr)) |
| 363 | *node_flag_ptr = NULL; |
| 364 | return NULL; |
| 365 | } |
| 366 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
| 367 | ptr = rcu_dereference(pointers[i]); |
| 368 | if (caa_unlikely(child_node_flag_ptr) && ptr) |
| 369 | *child_node_flag_ptr = &pointers[i]; |
| 370 | if (caa_unlikely(child_node_flag_v) && ptr) |
| 371 | *child_node_flag_v = ptr; |
| 372 | if (caa_unlikely(node_flag_ptr)) |
| 373 | *node_flag_ptr = &pointers[i]; |
| 374 | return ptr; |
| 375 | } |
| 376 | |
| 377 | static |
| 378 | void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, |
| 379 | struct cds_ja_inode *node, |
| 380 | uint8_t i, |
| 381 | uint8_t *v, |
| 382 | struct cds_ja_inode_flag **iter) |
| 383 | { |
| 384 | uint8_t *values; |
| 385 | struct cds_ja_inode_flag **pointers; |
| 386 | |
| 387 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 388 | assert(i < ja_linear_node_get_nr_child(type, node)); |
| 389 | |
| 390 | values = &node->u.data[1]; |
| 391 | *v = values[i]; |
| 392 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
| 393 | *iter = pointers[i]; |
| 394 | } |
| 395 | |
| 396 | static |
| 397 | struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, |
| 398 | struct cds_ja_inode *node, |
| 399 | struct cds_ja_inode_flag *node_flag, |
| 400 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
| 401 | struct cds_ja_inode_flag **child_node_flag_v, |
| 402 | struct cds_ja_inode_flag ***node_flag_ptr, |
| 403 | uint8_t n) |
| 404 | { |
| 405 | struct cds_ja_inode *linear; |
| 406 | |
| 407 | assert(type->type_class == RCU_JA_POOL); |
| 408 | |
| 409 | switch (type->nr_pool_order) { |
| 410 | case 1: |
| 411 | { |
| 412 | unsigned long bitsel, index; |
| 413 | |
| 414 | bitsel = ja_node_pool_1d_bitsel(node_flag); |
| 415 | assert(bitsel < CHAR_BIT); |
| 416 | index = ((unsigned long) n >> bitsel) & 0x1; |
| 417 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; |
| 418 | break; |
| 419 | } |
| 420 | case 2: |
| 421 | { |
| 422 | unsigned long bitsel[2], index[2], rindex; |
| 423 | |
| 424 | ja_node_pool_2d_bitsel(node_flag, bitsel); |
| 425 | assert(bitsel[0] < CHAR_BIT); |
| 426 | assert(bitsel[1] < CHAR_BIT); |
| 427 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; |
| 428 | index[0] <<= 1; |
| 429 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; |
| 430 | rindex = index[0] | index[1]; |
| 431 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; |
| 432 | break; |
| 433 | } |
| 434 | default: |
| 435 | linear = NULL; |
| 436 | assert(0); |
| 437 | } |
| 438 | return ja_linear_node_get_nth(type, linear, child_node_flag_ptr, |
| 439 | child_node_flag_v, node_flag_ptr, n); |
| 440 | } |
| 441 | |
| 442 | static |
| 443 | struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, |
| 444 | struct cds_ja_inode *node, |
| 445 | uint8_t i) |
| 446 | { |
| 447 | assert(type->type_class == RCU_JA_POOL); |
| 448 | return (struct cds_ja_inode *) |
| 449 | &node->u.data[(unsigned int) i << type->pool_size_order]; |
| 450 | } |
| 451 | |
| 452 | static |
| 453 | struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, |
| 454 | struct cds_ja_inode *node, |
| 455 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
| 456 | struct cds_ja_inode_flag **child_node_flag_v, |
| 457 | struct cds_ja_inode_flag ***node_flag_ptr, |
| 458 | uint8_t n) |
| 459 | { |
| 460 | struct cds_ja_inode_flag **child_node_flag; |
| 461 | struct cds_ja_inode_flag *child_node_flag_read; |
| 462 | |
| 463 | assert(type->type_class == RCU_JA_PIGEON); |
| 464 | child_node_flag = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
| 465 | child_node_flag_read = rcu_dereference(*child_node_flag); |
| 466 | dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", |
| 467 | child_node_flag); |
| 468 | if (caa_unlikely(child_node_flag_ptr) && child_node_flag_read) |
| 469 | *child_node_flag_ptr = child_node_flag; |
| 470 | if (caa_unlikely(child_node_flag_v) && child_node_flag_read) |
| 471 | *child_node_flag_v = child_node_flag_read; |
| 472 | if (caa_unlikely(node_flag_ptr)) |
| 473 | *node_flag_ptr = child_node_flag; |
| 474 | return child_node_flag_read; |
| 475 | } |
| 476 | |
| 477 | static |
| 478 | struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *type, |
| 479 | struct cds_ja_inode *node, |
| 480 | uint8_t i) |
| 481 | { |
| 482 | return ja_pigeon_node_get_nth(type, node, NULL, NULL, NULL, i); |
| 483 | } |
| 484 | |
| 485 | /* |
| 486 | * ja_node_get_nth: get nth item from a node. |
| 487 | * node_flag is already rcu_dereference'd. |
| 488 | */ |
| 489 | static |
| 490 | struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, |
| 491 | struct cds_ja_inode_flag ***child_node_flag_ptr, |
| 492 | struct cds_ja_inode_flag **child_node_flag, |
| 493 | struct cds_ja_inode_flag ***node_flag_ptr, |
| 494 | uint8_t n) |
| 495 | { |
| 496 | unsigned int type_index; |
| 497 | struct cds_ja_inode *node; |
| 498 | const struct cds_ja_type *type; |
| 499 | |
| 500 | node = ja_node_ptr(node_flag); |
| 501 | assert(node != NULL); |
| 502 | type_index = ja_node_type(node_flag); |
| 503 | type = &ja_types[type_index]; |
| 504 | |
| 505 | switch (type->type_class) { |
| 506 | case RCU_JA_LINEAR: |
| 507 | return ja_linear_node_get_nth(type, node, |
| 508 | child_node_flag_ptr, child_node_flag, |
| 509 | node_flag_ptr, n); |
| 510 | case RCU_JA_POOL: |
| 511 | return ja_pool_node_get_nth(type, node, node_flag, |
| 512 | child_node_flag_ptr, child_node_flag, |
| 513 | node_flag_ptr, n); |
| 514 | case RCU_JA_PIGEON: |
| 515 | return ja_pigeon_node_get_nth(type, node, |
| 516 | child_node_flag_ptr, child_node_flag, |
| 517 | node_flag_ptr, n); |
| 518 | default: |
| 519 | assert(0); |
| 520 | return (void *) -1UL; |
| 521 | } |
| 522 | } |
| 523 | |
| 524 | static |
| 525 | int ja_linear_node_set_nth(const struct cds_ja_type *type, |
| 526 | struct cds_ja_inode *node, |
| 527 | struct cds_ja_shadow_node *shadow_node, |
| 528 | uint8_t n, |
| 529 | struct cds_ja_inode_flag *child_node_flag) |
| 530 | { |
| 531 | uint8_t nr_child; |
| 532 | uint8_t *values, *nr_child_ptr; |
| 533 | struct cds_ja_inode_flag **pointers; |
| 534 | unsigned int i, unused = 0; |
| 535 | |
| 536 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 537 | |
| 538 | nr_child_ptr = &node->u.data[0]; |
| 539 | dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr); |
| 540 | nr_child = *nr_child_ptr; |
| 541 | assert(nr_child <= type->max_linear_child); |
| 542 | |
| 543 | values = &node->u.data[1]; |
| 544 | pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); |
| 545 | /* Check if node value is already populated */ |
| 546 | for (i = 0; i < nr_child; i++) { |
| 547 | if (values[i] == n) { |
| 548 | if (pointers[i]) |
| 549 | return -EEXIST; |
| 550 | else |
| 551 | break; |
| 552 | } else { |
| 553 | if (!pointers[i]) |
| 554 | unused++; |
| 555 | } |
| 556 | } |
| 557 | if (i == nr_child && nr_child >= type->max_linear_child) { |
| 558 | if (unused) |
| 559 | return -ERANGE; /* recompact node */ |
| 560 | else |
| 561 | return -ENOSPC; /* No space left in this node type */ |
| 562 | } |
| 563 | |
| 564 | assert(pointers[i] == NULL); |
| 565 | rcu_assign_pointer(pointers[i], child_node_flag); |
| 566 | /* If we expanded the nr_child, increment it */ |
| 567 | if (i == nr_child) { |
| 568 | CMM_STORE_SHARED(values[nr_child], n); |
| 569 | /* write pointer and value before nr_child */ |
| 570 | cmm_smp_wmb(); |
| 571 | CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1); |
| 572 | } |
| 573 | shadow_node->nr_child++; |
| 574 | dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", |
| 575 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
| 576 | (unsigned int) shadow_node->nr_child, |
| 577 | node, shadow_node); |
| 578 | |
| 579 | return 0; |
| 580 | } |
| 581 | |
| 582 | static |
| 583 | int ja_pool_node_set_nth(const struct cds_ja_type *type, |
| 584 | struct cds_ja_inode *node, |
| 585 | struct cds_ja_inode_flag *node_flag, |
| 586 | struct cds_ja_shadow_node *shadow_node, |
| 587 | uint8_t n, |
| 588 | struct cds_ja_inode_flag *child_node_flag) |
| 589 | { |
| 590 | struct cds_ja_inode *linear; |
| 591 | |
| 592 | assert(type->type_class == RCU_JA_POOL); |
| 593 | |
| 594 | switch (type->nr_pool_order) { |
| 595 | case 1: |
| 596 | { |
| 597 | unsigned long bitsel, index; |
| 598 | |
| 599 | bitsel = ja_node_pool_1d_bitsel(node_flag); |
| 600 | assert(bitsel < CHAR_BIT); |
| 601 | index = ((unsigned long) n >> bitsel) & 0x1; |
| 602 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; |
| 603 | break; |
| 604 | } |
| 605 | case 2: |
| 606 | { |
| 607 | unsigned long bitsel[2], index[2], rindex; |
| 608 | |
| 609 | ja_node_pool_2d_bitsel(node_flag, bitsel); |
| 610 | assert(bitsel[0] < CHAR_BIT); |
| 611 | assert(bitsel[1] < CHAR_BIT); |
| 612 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; |
| 613 | index[0] <<= 1; |
| 614 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; |
| 615 | rindex = index[0] | index[1]; |
| 616 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; |
| 617 | break; |
| 618 | } |
| 619 | default: |
| 620 | linear = NULL; |
| 621 | assert(0); |
| 622 | } |
| 623 | |
| 624 | return ja_linear_node_set_nth(type, linear, shadow_node, |
| 625 | n, child_node_flag); |
| 626 | } |
| 627 | |
| 628 | static |
| 629 | int ja_pigeon_node_set_nth(const struct cds_ja_type *type, |
| 630 | struct cds_ja_inode *node, |
| 631 | struct cds_ja_shadow_node *shadow_node, |
| 632 | uint8_t n, |
| 633 | struct cds_ja_inode_flag *child_node_flag) |
| 634 | { |
| 635 | struct cds_ja_inode_flag **ptr; |
| 636 | |
| 637 | assert(type->type_class == RCU_JA_PIGEON); |
| 638 | ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; |
| 639 | if (*ptr) |
| 640 | return -EEXIST; |
| 641 | rcu_assign_pointer(*ptr, child_node_flag); |
| 642 | shadow_node->nr_child++; |
| 643 | return 0; |
| 644 | } |
| 645 | |
| 646 | /* |
| 647 | * _ja_node_set_nth: set nth item within a node. Return an error |
| 648 | * (negative error value) if it is already there. |
| 649 | */ |
| 650 | static |
| 651 | int _ja_node_set_nth(const struct cds_ja_type *type, |
| 652 | struct cds_ja_inode *node, |
| 653 | struct cds_ja_inode_flag *node_flag, |
| 654 | struct cds_ja_shadow_node *shadow_node, |
| 655 | uint8_t n, |
| 656 | struct cds_ja_inode_flag *child_node_flag) |
| 657 | { |
| 658 | switch (type->type_class) { |
| 659 | case RCU_JA_LINEAR: |
| 660 | return ja_linear_node_set_nth(type, node, shadow_node, n, |
| 661 | child_node_flag); |
| 662 | case RCU_JA_POOL: |
| 663 | return ja_pool_node_set_nth(type, node, node_flag, shadow_node, n, |
| 664 | child_node_flag); |
| 665 | case RCU_JA_PIGEON: |
| 666 | return ja_pigeon_node_set_nth(type, node, shadow_node, n, |
| 667 | child_node_flag); |
| 668 | case RCU_JA_NULL: |
| 669 | return -ENOSPC; |
| 670 | default: |
| 671 | assert(0); |
| 672 | return -EINVAL; |
| 673 | } |
| 674 | |
| 675 | return 0; |
| 676 | } |
| 677 | |
| 678 | static |
| 679 | int ja_linear_node_clear_ptr(const struct cds_ja_type *type, |
| 680 | struct cds_ja_inode *node, |
| 681 | struct cds_ja_shadow_node *shadow_node, |
| 682 | struct cds_ja_inode_flag **node_flag_ptr) |
| 683 | { |
| 684 | uint8_t nr_child; |
| 685 | uint8_t *nr_child_ptr; |
| 686 | |
| 687 | assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); |
| 688 | |
| 689 | nr_child_ptr = &node->u.data[0]; |
| 690 | nr_child = *nr_child_ptr; |
| 691 | assert(nr_child <= type->max_linear_child); |
| 692 | |
| 693 | if (shadow_node->fallback_removal_count) { |
| 694 | shadow_node->fallback_removal_count--; |
| 695 | } else { |
| 696 | if (type->type_class == RCU_JA_LINEAR |
| 697 | && shadow_node->nr_child <= type->min_child) { |
| 698 | /* We need to try recompacting the node */ |
| 699 | return -EFBIG; |
| 700 | } |
| 701 | } |
| 702 | dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); |
| 703 | assert(*node_flag_ptr != NULL); |
| 704 | rcu_assign_pointer(*node_flag_ptr, NULL); |
| 705 | /* |
| 706 | * Value and nr_child are never changed (would cause ABA issue). |
| 707 | * Instead, we leave the pointer to NULL and recompact the node |
| 708 | * once in a while. It is allowed to set a NULL pointer to a new |
| 709 | * value without recompaction though. |
| 710 | * Only update the shadow node accounting. |
| 711 | */ |
| 712 | shadow_node->nr_child--; |
| 713 | dbg_printf("linear clear ptr: %u child, shadow: %u child, for node %p shadow %p\n", |
| 714 | (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr), |
| 715 | (unsigned int) shadow_node->nr_child, |
| 716 | node, shadow_node); |
| 717 | return 0; |
| 718 | } |
| 719 | |
| 720 | static |
| 721 | int ja_pool_node_clear_ptr(const struct cds_ja_type *type, |
| 722 | struct cds_ja_inode *node, |
| 723 | struct cds_ja_inode_flag *node_flag, |
| 724 | struct cds_ja_shadow_node *shadow_node, |
| 725 | struct cds_ja_inode_flag **node_flag_ptr, |
| 726 | uint8_t n) |
| 727 | { |
| 728 | struct cds_ja_inode *linear; |
| 729 | |
| 730 | assert(type->type_class == RCU_JA_POOL); |
| 731 | |
| 732 | if (shadow_node->fallback_removal_count) { |
| 733 | shadow_node->fallback_removal_count--; |
| 734 | } else { |
| 735 | /* We should try recompacting the node */ |
| 736 | if (shadow_node->nr_child <= type->min_child) |
| 737 | return -EFBIG; |
| 738 | } |
| 739 | |
| 740 | switch (type->nr_pool_order) { |
| 741 | case 1: |
| 742 | { |
| 743 | unsigned long bitsel, index; |
| 744 | |
| 745 | bitsel = ja_node_pool_1d_bitsel(node_flag); |
| 746 | assert(bitsel < CHAR_BIT); |
| 747 | index = ((unsigned long) n >> bitsel) & type->nr_pool_order; |
| 748 | linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; |
| 749 | break; |
| 750 | } |
| 751 | case 2: |
| 752 | { |
| 753 | unsigned long bitsel[2], index[2], rindex; |
| 754 | |
| 755 | ja_node_pool_2d_bitsel(node_flag, bitsel); |
| 756 | assert(bitsel[0] < CHAR_BIT); |
| 757 | assert(bitsel[1] < CHAR_BIT); |
| 758 | index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; |
| 759 | index[0] <<= 1; |
| 760 | index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; |
| 761 | rindex = index[0] | index[1]; |
| 762 | linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; |
| 763 | break; |
| 764 | } |
| 765 | default: |
| 766 | linear = NULL; |
| 767 | assert(0); |
| 768 | } |
| 769 | |
| 770 | return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); |
| 771 | } |
| 772 | |
| 773 | static |
| 774 | int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, |
| 775 | struct cds_ja_inode *node, |
| 776 | struct cds_ja_shadow_node *shadow_node, |
| 777 | struct cds_ja_inode_flag **node_flag_ptr) |
| 778 | { |
| 779 | assert(type->type_class == RCU_JA_PIGEON); |
| 780 | |
| 781 | if (shadow_node->fallback_removal_count) { |
| 782 | shadow_node->fallback_removal_count--; |
| 783 | } else { |
| 784 | /* We should try recompacting the node */ |
| 785 | if (shadow_node->nr_child <= type->min_child) |
| 786 | return -EFBIG; |
| 787 | } |
| 788 | dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); |
| 789 | rcu_assign_pointer(*node_flag_ptr, NULL); |
| 790 | shadow_node->nr_child--; |
| 791 | return 0; |
| 792 | } |
| 793 | |
| 794 | /* |
| 795 | * _ja_node_clear_ptr: clear ptr item within a node. Return an error |
| 796 | * (negative error value) if it is not found (-ENOENT). |
| 797 | */ |
| 798 | static |
| 799 | int _ja_node_clear_ptr(const struct cds_ja_type *type, |
| 800 | struct cds_ja_inode *node, |
| 801 | struct cds_ja_inode_flag *node_flag, |
| 802 | struct cds_ja_shadow_node *shadow_node, |
| 803 | struct cds_ja_inode_flag **node_flag_ptr, |
| 804 | uint8_t n) |
| 805 | { |
| 806 | switch (type->type_class) { |
| 807 | case RCU_JA_LINEAR: |
| 808 | return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
| 809 | case RCU_JA_POOL: |
| 810 | return ja_pool_node_clear_ptr(type, node, node_flag, shadow_node, node_flag_ptr, n); |
| 811 | case RCU_JA_PIGEON: |
| 812 | return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); |
| 813 | case RCU_JA_NULL: |
| 814 | return -ENOENT; |
| 815 | default: |
| 816 | assert(0); |
| 817 | return -EINVAL; |
| 818 | } |
| 819 | |
| 820 | return 0; |
| 821 | } |
| 822 | |
| 823 | /* |
| 824 | * Calculate bit distribution. Returns the bit (0 to 7) that splits the |
| 825 | * distribution in two sub-distributions containing as much elements one |
| 826 | * compared to the other. |
| 827 | */ |
| 828 | static |
| 829 | unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode, |
| 830 | struct cds_ja *ja, |
| 831 | unsigned int type_index, |
| 832 | const struct cds_ja_type *type, |
| 833 | struct cds_ja_inode *node, |
| 834 | struct cds_ja_shadow_node *shadow_node, |
| 835 | uint8_t n, |
| 836 | struct cds_ja_inode_flag *child_node_flag, |
| 837 | struct cds_ja_inode_flag **nullify_node_flag_ptr) |
| 838 | { |
| 839 | uint8_t nr_one[JA_BITS_PER_BYTE]; |
| 840 | unsigned int bitsel = 0, bit_i, overall_best_distance = UINT_MAX; |
| 841 | unsigned int distrib_nr_child = 0; |
| 842 | |
| 843 | memset(nr_one, 0, sizeof(nr_one)); |
| 844 | |
| 845 | switch (type->type_class) { |
| 846 | case RCU_JA_LINEAR: |
| 847 | { |
| 848 | uint8_t nr_child = |
| 849 | ja_linear_node_get_nr_child(type, node); |
| 850 | unsigned int i; |
| 851 | |
| 852 | for (i = 0; i < nr_child; i++) { |
| 853 | struct cds_ja_inode_flag *iter; |
| 854 | uint8_t v; |
| 855 | |
| 856 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); |
| 857 | if (!iter) |
| 858 | continue; |
| 859 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 860 | continue; |
| 861 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 862 | if (v & (1U << bit_i)) |
| 863 | nr_one[bit_i]++; |
| 864 | } |
| 865 | distrib_nr_child++; |
| 866 | } |
| 867 | break; |
| 868 | } |
| 869 | case RCU_JA_POOL: |
| 870 | { |
| 871 | unsigned int pool_nr; |
| 872 | |
| 873 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { |
| 874 | struct cds_ja_inode *pool = |
| 875 | ja_pool_node_get_ith_pool(type, |
| 876 | node, pool_nr); |
| 877 | uint8_t nr_child = |
| 878 | ja_linear_node_get_nr_child(type, pool); |
| 879 | unsigned int j; |
| 880 | |
| 881 | for (j = 0; j < nr_child; j++) { |
| 882 | struct cds_ja_inode_flag *iter; |
| 883 | uint8_t v; |
| 884 | |
| 885 | ja_linear_node_get_ith_pos(type, pool, |
| 886 | j, &v, &iter); |
| 887 | if (!iter) |
| 888 | continue; |
| 889 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 890 | continue; |
| 891 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 892 | if (v & (1U << bit_i)) |
| 893 | nr_one[bit_i]++; |
| 894 | } |
| 895 | distrib_nr_child++; |
| 896 | } |
| 897 | } |
| 898 | break; |
| 899 | } |
| 900 | case RCU_JA_PIGEON: |
| 901 | { |
| 902 | uint8_t nr_child; |
| 903 | unsigned int i; |
| 904 | |
| 905 | assert(mode == JA_RECOMPACT_DEL); |
| 906 | nr_child = shadow_node->nr_child; |
| 907 | for (i = 0; i < nr_child; i++) { |
| 908 | struct cds_ja_inode_flag *iter; |
| 909 | |
| 910 | iter = ja_pigeon_node_get_ith_pos(type, node, i); |
| 911 | if (!iter) |
| 912 | continue; |
| 913 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 914 | continue; |
| 915 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 916 | if (i & (1U << bit_i)) |
| 917 | nr_one[bit_i]++; |
| 918 | } |
| 919 | distrib_nr_child++; |
| 920 | } |
| 921 | break; |
| 922 | } |
| 923 | case RCU_JA_NULL: |
| 924 | assert(mode == JA_RECOMPACT_ADD_NEXT); |
| 925 | break; |
| 926 | default: |
| 927 | assert(0); |
| 928 | break; |
| 929 | } |
| 930 | |
| 931 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { |
| 932 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 933 | if (n & (1U << bit_i)) |
| 934 | nr_one[bit_i]++; |
| 935 | } |
| 936 | distrib_nr_child++; |
| 937 | } |
| 938 | |
| 939 | /* |
| 940 | * The best bit selector is that for which the number of ones is |
| 941 | * closest to half of the number of children in the |
| 942 | * distribution. We calculate the distance using the double of |
| 943 | * the sub-distribution sizes to eliminate truncation error. |
| 944 | */ |
| 945 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 946 | unsigned int distance_to_best; |
| 947 | |
| 948 | distance_to_best = abs_int((nr_one[bit_i] << 1U) - distrib_nr_child); |
| 949 | if (distance_to_best < overall_best_distance) { |
| 950 | overall_best_distance = distance_to_best; |
| 951 | bitsel = bit_i; |
| 952 | } |
| 953 | } |
| 954 | dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel); |
| 955 | return bitsel; |
| 956 | } |
| 957 | |
| 958 | /* |
| 959 | * Calculate bit distribution in two dimensions. Returns the two bits |
| 960 | * (each 0 to 7) that splits the distribution in four sub-distributions |
| 961 | * containing as much elements one compared to the other. |
| 962 | */ |
| 963 | static |
| 964 | void ja_node_sum_distribution_2d(enum ja_recompact mode, |
| 965 | struct cds_ja *ja, |
| 966 | unsigned int type_index, |
| 967 | const struct cds_ja_type *type, |
| 968 | struct cds_ja_inode *node, |
| 969 | struct cds_ja_shadow_node *shadow_node, |
| 970 | uint8_t n, |
| 971 | struct cds_ja_inode_flag *child_node_flag, |
| 972 | struct cds_ja_inode_flag **nullify_node_flag_ptr, |
| 973 | unsigned int *_bitsel) |
| 974 | { |
| 975 | uint8_t nr_2d_11[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], |
| 976 | nr_2d_10[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], |
| 977 | nr_2d_01[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], |
| 978 | nr_2d_00[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE]; |
| 979 | unsigned int bitsel[2] = { 0, 1 }; |
| 980 | unsigned int bit_i, bit_j; |
| 981 | int overall_best_distance = INT_MAX; |
| 982 | unsigned int distrib_nr_child = 0; |
| 983 | |
| 984 | memset(nr_2d_11, 0, sizeof(nr_2d_11)); |
| 985 | memset(nr_2d_10, 0, sizeof(nr_2d_10)); |
| 986 | memset(nr_2d_01, 0, sizeof(nr_2d_01)); |
| 987 | memset(nr_2d_00, 0, sizeof(nr_2d_00)); |
| 988 | |
| 989 | switch (type->type_class) { |
| 990 | case RCU_JA_LINEAR: |
| 991 | { |
| 992 | uint8_t nr_child = |
| 993 | ja_linear_node_get_nr_child(type, node); |
| 994 | unsigned int i; |
| 995 | |
| 996 | for (i = 0; i < nr_child; i++) { |
| 997 | struct cds_ja_inode_flag *iter; |
| 998 | uint8_t v; |
| 999 | |
| 1000 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); |
| 1001 | if (!iter) |
| 1002 | continue; |
| 1003 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 1004 | continue; |
| 1005 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 1006 | for (bit_j = 0; bit_j < bit_i; bit_j++) { |
| 1007 | if ((v & (1U << bit_i)) && (v & (1U << bit_j))) { |
| 1008 | nr_2d_11[bit_i][bit_j]++; |
| 1009 | } |
| 1010 | if ((v & (1U << bit_i)) && !(v & (1U << bit_j))) { |
| 1011 | nr_2d_10[bit_i][bit_j]++; |
| 1012 | } |
| 1013 | if (!(v & (1U << bit_i)) && (v & (1U << bit_j))) { |
| 1014 | nr_2d_01[bit_i][bit_j]++; |
| 1015 | } |
| 1016 | if (!(v & (1U << bit_i)) && !(v & (1U << bit_j))) { |
| 1017 | nr_2d_00[bit_i][bit_j]++; |
| 1018 | } |
| 1019 | } |
| 1020 | } |
| 1021 | distrib_nr_child++; |
| 1022 | } |
| 1023 | break; |
| 1024 | } |
| 1025 | case RCU_JA_POOL: |
| 1026 | { |
| 1027 | unsigned int pool_nr; |
| 1028 | |
| 1029 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { |
| 1030 | struct cds_ja_inode *pool = |
| 1031 | ja_pool_node_get_ith_pool(type, |
| 1032 | node, pool_nr); |
| 1033 | uint8_t nr_child = |
| 1034 | ja_linear_node_get_nr_child(type, pool); |
| 1035 | unsigned int j; |
| 1036 | |
| 1037 | for (j = 0; j < nr_child; j++) { |
| 1038 | struct cds_ja_inode_flag *iter; |
| 1039 | uint8_t v; |
| 1040 | |
| 1041 | ja_linear_node_get_ith_pos(type, pool, |
| 1042 | j, &v, &iter); |
| 1043 | if (!iter) |
| 1044 | continue; |
| 1045 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 1046 | continue; |
| 1047 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 1048 | for (bit_j = 0; bit_j < bit_i; bit_j++) { |
| 1049 | if ((v & (1U << bit_i)) && (v & (1U << bit_j))) { |
| 1050 | nr_2d_11[bit_i][bit_j]++; |
| 1051 | } |
| 1052 | if ((v & (1U << bit_i)) && !(v & (1U << bit_j))) { |
| 1053 | nr_2d_10[bit_i][bit_j]++; |
| 1054 | } |
| 1055 | if (!(v & (1U << bit_i)) && (v & (1U << bit_j))) { |
| 1056 | nr_2d_01[bit_i][bit_j]++; |
| 1057 | } |
| 1058 | if (!(v & (1U << bit_i)) && !(v & (1U << bit_j))) { |
| 1059 | nr_2d_00[bit_i][bit_j]++; |
| 1060 | } |
| 1061 | } |
| 1062 | } |
| 1063 | distrib_nr_child++; |
| 1064 | } |
| 1065 | } |
| 1066 | break; |
| 1067 | } |
| 1068 | case RCU_JA_PIGEON: |
| 1069 | { |
| 1070 | uint8_t nr_child; |
| 1071 | unsigned int i; |
| 1072 | |
| 1073 | assert(mode == JA_RECOMPACT_DEL); |
| 1074 | nr_child = shadow_node->nr_child; |
| 1075 | for (i = 0; i < nr_child; i++) { |
| 1076 | struct cds_ja_inode_flag *iter; |
| 1077 | |
| 1078 | iter = ja_pigeon_node_get_ith_pos(type, node, i); |
| 1079 | if (!iter) |
| 1080 | continue; |
| 1081 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 1082 | continue; |
| 1083 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 1084 | for (bit_j = 0; bit_j < bit_i; bit_j++) { |
| 1085 | if ((i & (1U << bit_i)) && (i & (1U << bit_j))) { |
| 1086 | nr_2d_11[bit_i][bit_j]++; |
| 1087 | } |
| 1088 | if ((i & (1U << bit_i)) && !(i & (1U << bit_j))) { |
| 1089 | nr_2d_10[bit_i][bit_j]++; |
| 1090 | } |
| 1091 | if (!(i & (1U << bit_i)) && (i & (1U << bit_j))) { |
| 1092 | nr_2d_01[bit_i][bit_j]++; |
| 1093 | } |
| 1094 | if (!(i & (1U << bit_i)) && !(i & (1U << bit_j))) { |
| 1095 | nr_2d_00[bit_i][bit_j]++; |
| 1096 | } |
| 1097 | } |
| 1098 | } |
| 1099 | distrib_nr_child++; |
| 1100 | } |
| 1101 | break; |
| 1102 | } |
| 1103 | case RCU_JA_NULL: |
| 1104 | assert(mode == JA_RECOMPACT_ADD_NEXT); |
| 1105 | break; |
| 1106 | default: |
| 1107 | assert(0); |
| 1108 | break; |
| 1109 | } |
| 1110 | |
| 1111 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { |
| 1112 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 1113 | for (bit_j = 0; bit_j < bit_i; bit_j++) { |
| 1114 | if ((n & (1U << bit_i)) && (n & (1U << bit_j))) { |
| 1115 | nr_2d_11[bit_i][bit_j]++; |
| 1116 | } |
| 1117 | if ((n & (1U << bit_i)) && !(n & (1U << bit_j))) { |
| 1118 | nr_2d_10[bit_i][bit_j]++; |
| 1119 | } |
| 1120 | if (!(n & (1U << bit_i)) && (n & (1U << bit_j))) { |
| 1121 | nr_2d_01[bit_i][bit_j]++; |
| 1122 | } |
| 1123 | if (!(n & (1U << bit_i)) && !(n & (1U << bit_j))) { |
| 1124 | nr_2d_00[bit_i][bit_j]++; |
| 1125 | } |
| 1126 | } |
| 1127 | } |
| 1128 | distrib_nr_child++; |
| 1129 | } |
| 1130 | |
| 1131 | /* |
| 1132 | * The best bit selector is that for which the number of nodes |
| 1133 | * in each sub-class is closest to one-fourth of the number of |
| 1134 | * children in the distribution. We calculate the distance using |
| 1135 | * 4 times the size of the sub-distribution to eliminate |
| 1136 | * truncation error. |
| 1137 | */ |
| 1138 | for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { |
| 1139 | for (bit_j = 0; bit_j < bit_i; bit_j++) { |
| 1140 | int distance_to_best[4]; |
| 1141 | |
| 1142 | distance_to_best[0] = (nr_2d_11[bit_i][bit_j] << 2U) - distrib_nr_child; |
| 1143 | distance_to_best[1] = (nr_2d_10[bit_i][bit_j] << 2U) - distrib_nr_child; |
| 1144 | distance_to_best[2] = (nr_2d_01[bit_i][bit_j] << 2U) - distrib_nr_child; |
| 1145 | distance_to_best[3] = (nr_2d_00[bit_i][bit_j] << 2U) - distrib_nr_child; |
| 1146 | |
| 1147 | /* Consider worse distance above best */ |
| 1148 | if (distance_to_best[1] > 0 && distance_to_best[1] > distance_to_best[0]) |
| 1149 | distance_to_best[0] = distance_to_best[1]; |
| 1150 | if (distance_to_best[2] > 0 && distance_to_best[2] > distance_to_best[0]) |
| 1151 | distance_to_best[0] = distance_to_best[2]; |
| 1152 | if (distance_to_best[3] > 0 && distance_to_best[3] > distance_to_best[0]) |
| 1153 | distance_to_best[0] = distance_to_best[3]; |
| 1154 | |
| 1155 | /* |
| 1156 | * If our worse distance is better than overall, |
| 1157 | * we become new best candidate. |
| 1158 | */ |
| 1159 | if (distance_to_best[0] < overall_best_distance) { |
| 1160 | overall_best_distance = distance_to_best[0]; |
| 1161 | bitsel[0] = bit_i; |
| 1162 | bitsel[1] = bit_j; |
| 1163 | } |
| 1164 | } |
| 1165 | } |
| 1166 | |
| 1167 | dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel[0], bitsel[1]); |
| 1168 | |
| 1169 | /* Return our bit selection */ |
| 1170 | _bitsel[0] = bitsel[0]; |
| 1171 | _bitsel[1] = bitsel[1]; |
| 1172 | } |
| 1173 | |
| 1174 | /* |
| 1175 | * ja_node_recompact_add: recompact a node, adding a new child. |
| 1176 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
| 1177 | * error value otherwise. |
| 1178 | */ |
| 1179 | static |
| 1180 | int ja_node_recompact(enum ja_recompact mode, |
| 1181 | struct cds_ja *ja, |
| 1182 | unsigned int old_type_index, |
| 1183 | const struct cds_ja_type *old_type, |
| 1184 | struct cds_ja_inode *old_node, |
| 1185 | struct cds_ja_shadow_node *shadow_node, |
| 1186 | struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, |
| 1187 | struct cds_ja_inode_flag *child_node_flag, |
| 1188 | struct cds_ja_inode_flag **nullify_node_flag_ptr) |
| 1189 | { |
| 1190 | unsigned int new_type_index; |
| 1191 | struct cds_ja_inode *new_node; |
| 1192 | struct cds_ja_shadow_node *new_shadow_node = NULL; |
| 1193 | const struct cds_ja_type *new_type; |
| 1194 | struct cds_ja_inode_flag *new_node_flag, *old_node_flag; |
| 1195 | int ret; |
| 1196 | int fallback = 0; |
| 1197 | |
| 1198 | old_node_flag = *old_node_flag_ptr; |
| 1199 | |
| 1200 | switch (mode) { |
| 1201 | case JA_RECOMPACT_ADD_SAME: |
| 1202 | if (old_type->type_class == RCU_JA_POOL) { |
| 1203 | /* |
| 1204 | * For pool type, try redistributing |
| 1205 | * into a different distribution of same |
| 1206 | * size if we have not reached limits. |
| 1207 | */ |
| 1208 | if (shadow_node->nr_child + 1 > old_type->max_child) { |
| 1209 | new_type_index = old_type_index + 1; |
| 1210 | } else if (shadow_node->nr_child + 1 < old_type->min_child) { |
| 1211 | new_type_index = old_type_index - 1; |
| 1212 | } else { |
| 1213 | new_type_index = old_type_index; |
| 1214 | } |
| 1215 | } else { |
| 1216 | new_type_index = old_type_index; |
| 1217 | } |
| 1218 | break; |
| 1219 | case JA_RECOMPACT_ADD_NEXT: |
| 1220 | if (!shadow_node || old_type_index == NODE_INDEX_NULL) { |
| 1221 | new_type_index = 0; |
| 1222 | } else { |
| 1223 | if (old_type->type_class == RCU_JA_POOL) { |
| 1224 | /* |
| 1225 | * For pool type, try redistributing |
| 1226 | * into a different distribution of same |
| 1227 | * size if we have not reached limits. |
| 1228 | */ |
| 1229 | if (shadow_node->nr_child + 1 > old_type->max_child) { |
| 1230 | new_type_index = old_type_index + 1; |
| 1231 | } else { |
| 1232 | new_type_index = old_type_index; |
| 1233 | } |
| 1234 | } else { |
| 1235 | new_type_index = old_type_index + 1; |
| 1236 | } |
| 1237 | } |
| 1238 | break; |
| 1239 | case JA_RECOMPACT_DEL: |
| 1240 | if (old_type_index == 0) { |
| 1241 | new_type_index = NODE_INDEX_NULL; |
| 1242 | } else { |
| 1243 | if (old_type->type_class == RCU_JA_POOL) { |
| 1244 | /* |
| 1245 | * For pool type, try redistributing |
| 1246 | * into a different distribution of same |
| 1247 | * size if we have not reached limits. |
| 1248 | */ |
| 1249 | if (shadow_node->nr_child - 1 < old_type->min_child) { |
| 1250 | new_type_index = old_type_index - 1; |
| 1251 | } else { |
| 1252 | new_type_index = old_type_index; |
| 1253 | } |
| 1254 | } else { |
| 1255 | new_type_index = old_type_index - 1; |
| 1256 | } |
| 1257 | } |
| 1258 | break; |
| 1259 | default: |
| 1260 | assert(0); |
| 1261 | } |
| 1262 | |
| 1263 | retry: /* for fallback */ |
| 1264 | dbg_printf("Recompact from type %d to type %d\n", |
| 1265 | old_type_index, new_type_index); |
| 1266 | new_type = &ja_types[new_type_index]; |
| 1267 | if (new_type_index != NODE_INDEX_NULL) { |
| 1268 | new_node = alloc_cds_ja_node(new_type); |
| 1269 | if (!new_node) |
| 1270 | return -ENOMEM; |
| 1271 | |
| 1272 | if (new_type->type_class == RCU_JA_POOL) { |
| 1273 | switch (new_type->nr_pool_order) { |
| 1274 | case 1: |
| 1275 | { |
| 1276 | unsigned int node_distrib_bitsel; |
| 1277 | |
| 1278 | node_distrib_bitsel = |
| 1279 | ja_node_sum_distribution_1d(mode, ja, |
| 1280 | old_type_index, old_type, |
| 1281 | old_node, shadow_node, |
| 1282 | n, child_node_flag, |
| 1283 | nullify_node_flag_ptr); |
| 1284 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); |
| 1285 | new_node_flag = ja_node_flag_pool_1d(new_node, |
| 1286 | new_type_index, node_distrib_bitsel); |
| 1287 | break; |
| 1288 | } |
| 1289 | case 2: |
| 1290 | { |
| 1291 | unsigned int node_distrib_bitsel[2]; |
| 1292 | |
| 1293 | ja_node_sum_distribution_2d(mode, ja, |
| 1294 | old_type_index, old_type, |
| 1295 | old_node, shadow_node, |
| 1296 | n, child_node_flag, |
| 1297 | nullify_node_flag_ptr, |
| 1298 | node_distrib_bitsel); |
| 1299 | assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); |
| 1300 | assert(!((unsigned long) new_node & JA_POOL_2D_MASK)); |
| 1301 | new_node_flag = ja_node_flag_pool_2d(new_node, |
| 1302 | new_type_index, node_distrib_bitsel); |
| 1303 | break; |
| 1304 | } |
| 1305 | default: |
| 1306 | assert(0); |
| 1307 | } |
| 1308 | } else { |
| 1309 | new_node_flag = ja_node_flag(new_node, new_type_index); |
| 1310 | } |
| 1311 | |
| 1312 | dbg_printf("Recompact inherit lock from %p\n", shadow_node); |
| 1313 | new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja); |
| 1314 | if (!new_shadow_node) { |
| 1315 | free(new_node); |
| 1316 | return -ENOMEM; |
| 1317 | } |
| 1318 | if (fallback) |
| 1319 | new_shadow_node->fallback_removal_count = |
| 1320 | JA_FALLBACK_REMOVAL_COUNT; |
| 1321 | } else { |
| 1322 | new_node = NULL; |
| 1323 | new_node_flag = NULL; |
| 1324 | } |
| 1325 | |
| 1326 | assert(mode != JA_RECOMPACT_ADD_NEXT || old_type->type_class != RCU_JA_PIGEON); |
| 1327 | |
| 1328 | if (new_type_index == NODE_INDEX_NULL) |
| 1329 | goto skip_copy; |
| 1330 | |
| 1331 | switch (old_type->type_class) { |
| 1332 | case RCU_JA_LINEAR: |
| 1333 | { |
| 1334 | uint8_t nr_child = |
| 1335 | ja_linear_node_get_nr_child(old_type, old_node); |
| 1336 | unsigned int i; |
| 1337 | |
| 1338 | for (i = 0; i < nr_child; i++) { |
| 1339 | struct cds_ja_inode_flag *iter; |
| 1340 | uint8_t v; |
| 1341 | |
| 1342 | ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); |
| 1343 | if (!iter) |
| 1344 | continue; |
| 1345 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 1346 | continue; |
| 1347 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
| 1348 | new_shadow_node, |
| 1349 | v, iter); |
| 1350 | if (new_type->type_class == RCU_JA_POOL && ret) { |
| 1351 | goto fallback_toosmall; |
| 1352 | } |
| 1353 | assert(!ret); |
| 1354 | } |
| 1355 | break; |
| 1356 | } |
| 1357 | case RCU_JA_POOL: |
| 1358 | { |
| 1359 | unsigned int pool_nr; |
| 1360 | |
| 1361 | for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { |
| 1362 | struct cds_ja_inode *pool = |
| 1363 | ja_pool_node_get_ith_pool(old_type, |
| 1364 | old_node, pool_nr); |
| 1365 | uint8_t nr_child = |
| 1366 | ja_linear_node_get_nr_child(old_type, pool); |
| 1367 | unsigned int j; |
| 1368 | |
| 1369 | for (j = 0; j < nr_child; j++) { |
| 1370 | struct cds_ja_inode_flag *iter; |
| 1371 | uint8_t v; |
| 1372 | |
| 1373 | ja_linear_node_get_ith_pos(old_type, pool, |
| 1374 | j, &v, &iter); |
| 1375 | if (!iter) |
| 1376 | continue; |
| 1377 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 1378 | continue; |
| 1379 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
| 1380 | new_shadow_node, |
| 1381 | v, iter); |
| 1382 | if (new_type->type_class == RCU_JA_POOL |
| 1383 | && ret) { |
| 1384 | goto fallback_toosmall; |
| 1385 | } |
| 1386 | assert(!ret); |
| 1387 | } |
| 1388 | } |
| 1389 | break; |
| 1390 | } |
| 1391 | case RCU_JA_NULL: |
| 1392 | assert(mode == JA_RECOMPACT_ADD_NEXT); |
| 1393 | break; |
| 1394 | case RCU_JA_PIGEON: |
| 1395 | { |
| 1396 | uint8_t nr_child; |
| 1397 | unsigned int i; |
| 1398 | |
| 1399 | assert(mode == JA_RECOMPACT_DEL); |
| 1400 | nr_child = shadow_node->nr_child; |
| 1401 | for (i = 0; i < nr_child; i++) { |
| 1402 | struct cds_ja_inode_flag *iter; |
| 1403 | |
| 1404 | iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); |
| 1405 | if (!iter) |
| 1406 | continue; |
| 1407 | if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) |
| 1408 | continue; |
| 1409 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
| 1410 | new_shadow_node, |
| 1411 | i, iter); |
| 1412 | if (new_type->type_class == RCU_JA_POOL && ret) { |
| 1413 | goto fallback_toosmall; |
| 1414 | } |
| 1415 | assert(!ret); |
| 1416 | } |
| 1417 | break; |
| 1418 | } |
| 1419 | default: |
| 1420 | assert(0); |
| 1421 | ret = -EINVAL; |
| 1422 | goto end; |
| 1423 | } |
| 1424 | skip_copy: |
| 1425 | |
| 1426 | if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { |
| 1427 | /* add node */ |
| 1428 | ret = _ja_node_set_nth(new_type, new_node, new_node_flag, |
| 1429 | new_shadow_node, |
| 1430 | n, child_node_flag); |
| 1431 | if (new_type->type_class == RCU_JA_POOL && ret) { |
| 1432 | goto fallback_toosmall; |
| 1433 | } |
| 1434 | assert(!ret); |
| 1435 | } |
| 1436 | |
| 1437 | if (fallback) { |
| 1438 | dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n", |
| 1439 | new_shadow_node->nr_child, old_type_index, mode == JA_RECOMPACT_ADD_NEXT ? "add_next" : |
| 1440 | (mode == JA_RECOMPACT_DEL ? "del" : "add_same")); |
| 1441 | uatomic_inc(&node_fallback_count_distribution[new_shadow_node->nr_child]); |
| 1442 | } |
| 1443 | |
| 1444 | /* Return pointer to new recompacted node through old_node_flag_ptr */ |
| 1445 | *old_node_flag_ptr = new_node_flag; |
| 1446 | if (old_node) { |
| 1447 | int flags; |
| 1448 | |
| 1449 | flags = RCUJA_SHADOW_CLEAR_FREE_NODE; |
| 1450 | /* |
| 1451 | * It is OK to free the lock associated with a node |
| 1452 | * going to NULL, since we are holding the parent lock. |
| 1453 | * This synchronizes removal with re-add of that node. |
| 1454 | */ |
| 1455 | if (new_type_index == NODE_INDEX_NULL) |
| 1456 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; |
| 1457 | ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, |
| 1458 | flags); |
| 1459 | assert(!ret); |
| 1460 | } |
| 1461 | |
| 1462 | ret = 0; |
| 1463 | end: |
| 1464 | return ret; |
| 1465 | |
| 1466 | fallback_toosmall: |
| 1467 | /* fallback if next pool is too small */ |
| 1468 | assert(new_shadow_node); |
| 1469 | ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, |
| 1470 | RCUJA_SHADOW_CLEAR_FREE_NODE); |
| 1471 | assert(!ret); |
| 1472 | |
| 1473 | switch (mode) { |
| 1474 | case JA_RECOMPACT_ADD_SAME: |
| 1475 | /* |
| 1476 | * JA_RECOMPACT_ADD_SAME is only triggered if a linear |
| 1477 | * node within a pool has unused entries. It should |
| 1478 | * therefore _never_ be too small. |
| 1479 | */ |
| 1480 | assert(0); |
| 1481 | |
| 1482 | /* Fall-through */ |
| 1483 | case JA_RECOMPACT_ADD_NEXT: |
| 1484 | { |
| 1485 | const struct cds_ja_type *next_type; |
| 1486 | |
| 1487 | /* |
| 1488 | * Recompaction attempt on add failed. Should only |
| 1489 | * happen if target node type is pool. Caused by |
| 1490 | * hard-to-split distribution. Recompact using the next |
| 1491 | * distribution size. |
| 1492 | */ |
| 1493 | assert(new_type->type_class == RCU_JA_POOL); |
| 1494 | next_type = &ja_types[new_type_index + 1]; |
| 1495 | /* |
| 1496 | * Try going to the next pool size if our population |
| 1497 | * fits within its range. This is not flagged as a |
| 1498 | * fallback. |
| 1499 | */ |
| 1500 | if (shadow_node->nr_child + 1 >= next_type->min_child |
| 1501 | && shadow_node->nr_child + 1 <= next_type->max_child) { |
| 1502 | new_type_index++; |
| 1503 | goto retry; |
| 1504 | } else { |
| 1505 | new_type_index++; |
| 1506 | dbg_printf("Add fallback to type %d\n", new_type_index); |
| 1507 | uatomic_inc(&ja->nr_fallback); |
| 1508 | fallback = 1; |
| 1509 | goto retry; |
| 1510 | } |
| 1511 | break; |
| 1512 | } |
| 1513 | case JA_RECOMPACT_DEL: |
| 1514 | /* |
| 1515 | * Recompaction attempt on delete failed. Should only |
| 1516 | * happen if target node type is pool. This is caused by |
| 1517 | * a hard-to-split distribution. Recompact on same node |
| 1518 | * size, but flag current node as "fallback" to ensure |
| 1519 | * we don't attempt recompaction before some activity |
| 1520 | * has reshuffled our node. |
| 1521 | */ |
| 1522 | assert(new_type->type_class == RCU_JA_POOL); |
| 1523 | new_type_index = old_type_index; |
| 1524 | dbg_printf("Delete fallback keeping type %d\n", new_type_index); |
| 1525 | uatomic_inc(&ja->nr_fallback); |
| 1526 | fallback = 1; |
| 1527 | goto retry; |
| 1528 | default: |
| 1529 | assert(0); |
| 1530 | return -EINVAL; |
| 1531 | } |
| 1532 | |
| 1533 | /* |
| 1534 | * Last resort fallback: pigeon. |
| 1535 | */ |
| 1536 | new_type_index = (1UL << JA_TYPE_BITS) - 1; |
| 1537 | dbg_printf("Fallback to type %d\n", new_type_index); |
| 1538 | uatomic_inc(&ja->nr_fallback); |
| 1539 | fallback = 1; |
| 1540 | goto retry; |
| 1541 | } |
| 1542 | |
| 1543 | /* |
| 1544 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
| 1545 | * error value otherwise. |
| 1546 | */ |
| 1547 | static |
| 1548 | int ja_node_set_nth(struct cds_ja *ja, |
| 1549 | struct cds_ja_inode_flag **node_flag, uint8_t n, |
| 1550 | struct cds_ja_inode_flag *child_node_flag, |
| 1551 | struct cds_ja_shadow_node *shadow_node) |
| 1552 | { |
| 1553 | int ret; |
| 1554 | unsigned int type_index; |
| 1555 | const struct cds_ja_type *type; |
| 1556 | struct cds_ja_inode *node; |
| 1557 | |
| 1558 | dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n", |
| 1559 | (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); |
| 1560 | |
| 1561 | node = ja_node_ptr(*node_flag); |
| 1562 | type_index = ja_node_type(*node_flag); |
| 1563 | type = &ja_types[type_index]; |
| 1564 | ret = _ja_node_set_nth(type, node, *node_flag, shadow_node, |
| 1565 | n, child_node_flag); |
| 1566 | switch (ret) { |
| 1567 | case -ENOSPC: |
| 1568 | /* Not enough space in node, need to recompact to next type. */ |
| 1569 | ret = ja_node_recompact(JA_RECOMPACT_ADD_NEXT, ja, type_index, type, node, |
| 1570 | shadow_node, node_flag, n, child_node_flag, NULL); |
| 1571 | break; |
| 1572 | case -ERANGE: |
| 1573 | /* Node needs to be recompacted. */ |
| 1574 | ret = ja_node_recompact(JA_RECOMPACT_ADD_SAME, ja, type_index, type, node, |
| 1575 | shadow_node, node_flag, n, child_node_flag, NULL); |
| 1576 | break; |
| 1577 | } |
| 1578 | return ret; |
| 1579 | } |
| 1580 | |
| 1581 | /* |
| 1582 | * Return 0 on success, -EAGAIN if need to retry, or other negative |
| 1583 | * error value otherwise. |
| 1584 | */ |
| 1585 | static |
| 1586 | int ja_node_clear_ptr(struct cds_ja *ja, |
| 1587 | struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ |
| 1588 | struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ |
| 1589 | struct cds_ja_shadow_node *shadow_node, /* of parent */ |
| 1590 | uint8_t n) |
| 1591 | { |
| 1592 | int ret; |
| 1593 | unsigned int type_index; |
| 1594 | const struct cds_ja_type *type; |
| 1595 | struct cds_ja_inode *node; |
| 1596 | |
| 1597 | dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", |
| 1598 | ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); |
| 1599 | |
| 1600 | node = ja_node_ptr(*parent_node_flag_ptr); |
| 1601 | type_index = ja_node_type(*parent_node_flag_ptr); |
| 1602 | type = &ja_types[type_index]; |
| 1603 | ret = _ja_node_clear_ptr(type, node, *parent_node_flag_ptr, shadow_node, node_flag_ptr, n); |
| 1604 | if (ret == -EFBIG) { |
| 1605 | /* Should try recompaction. */ |
| 1606 | ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, |
| 1607 | shadow_node, parent_node_flag_ptr, n, NULL, |
| 1608 | node_flag_ptr); |
| 1609 | } |
| 1610 | return ret; |
| 1611 | } |
| 1612 | |
| 1613 | struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key) |
| 1614 | { |
| 1615 | unsigned int tree_depth, i; |
| 1616 | struct cds_ja_inode_flag *node_flag; |
| 1617 | struct cds_hlist_head head = { NULL }; |
| 1618 | |
| 1619 | if (caa_unlikely(key > ja->key_max)) |
| 1620 | return head; |
| 1621 | tree_depth = ja->tree_depth; |
| 1622 | node_flag = rcu_dereference(ja->root); |
| 1623 | |
| 1624 | /* level 0: root node */ |
| 1625 | if (!ja_node_ptr(node_flag)) |
| 1626 | return head; |
| 1627 | |
| 1628 | for (i = 1; i < tree_depth; i++) { |
| 1629 | uint8_t iter_key; |
| 1630 | |
| 1631 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
| 1632 | node_flag = ja_node_get_nth(node_flag, NULL, NULL, NULL, |
| 1633 | iter_key); |
| 1634 | dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", |
| 1635 | (unsigned int) iter_key, node_flag); |
| 1636 | if (!ja_node_ptr(node_flag)) |
| 1637 | return head; |
| 1638 | } |
| 1639 | |
| 1640 | /* Last level lookup succeded. We got an actual match. */ |
| 1641 | head.next = (struct cds_hlist_node *) node_flag; |
| 1642 | return head; |
| 1643 | } |
| 1644 | |
| 1645 | /* |
| 1646 | * We reached an unpopulated node. Create it and the children we need, |
| 1647 | * and then attach the entire branch to the current node. This may |
| 1648 | * trigger recompaction of the current node. Locks needed: node lock |
| 1649 | * (for add), and, possibly, parent node lock (to update pointer due to |
| 1650 | * node recompaction). |
| 1651 | * |
| 1652 | * First take node lock, check if recompaction is needed, then take |
| 1653 | * parent lock (if needed). Then we can proceed to create the new |
| 1654 | * branch. Publish the new branch, and release locks. |
| 1655 | * TODO: we currently always take the parent lock even when not needed. |
| 1656 | */ |
| 1657 | static |
| 1658 | int ja_attach_node(struct cds_ja *ja, |
| 1659 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
| 1660 | struct cds_ja_inode_flag *attach_node_flag, |
| 1661 | struct cds_ja_inode_flag **node_flag_ptr, |
| 1662 | struct cds_ja_inode_flag *node_flag, |
| 1663 | struct cds_ja_inode_flag *parent_node_flag, |
| 1664 | uint64_t key, |
| 1665 | unsigned int level, |
| 1666 | struct cds_ja_node *child_node) |
| 1667 | { |
| 1668 | struct cds_ja_shadow_node *shadow_node = NULL, |
| 1669 | *parent_shadow_node = NULL; |
| 1670 | struct cds_ja_inode *node = ja_node_ptr(node_flag); |
| 1671 | struct cds_ja_inode *parent_node = ja_node_ptr(parent_node_flag); |
| 1672 | struct cds_hlist_head head; |
| 1673 | struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag; |
| 1674 | int ret, i; |
| 1675 | struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH]; |
| 1676 | int nr_created_nodes = 0; |
| 1677 | |
| 1678 | dbg_printf("Attach node at level %u (node %p, node_flag %p)\n", |
| 1679 | level, node, node_flag); |
| 1680 | |
| 1681 | assert(node); |
| 1682 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, node_flag); |
| 1683 | if (!shadow_node) { |
| 1684 | ret = -EAGAIN; |
| 1685 | goto end; |
| 1686 | } |
| 1687 | if (parent_node) { |
| 1688 | parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
| 1689 | parent_node_flag); |
| 1690 | if (!parent_shadow_node) { |
| 1691 | ret = -EAGAIN; |
| 1692 | goto unlock_shadow; |
| 1693 | } |
| 1694 | } |
| 1695 | |
| 1696 | if (node_flag_ptr && ja_node_ptr(*node_flag_ptr)) { |
| 1697 | /* |
| 1698 | * Target node has been updated between RCU lookup and |
| 1699 | * lock acquisition. We need to re-try lookup and |
| 1700 | * attach. |
| 1701 | */ |
| 1702 | ret = -EAGAIN; |
| 1703 | goto unlock_parent; |
| 1704 | } |
| 1705 | |
| 1706 | if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != |
| 1707 | ja_node_ptr(attach_node_flag)) { |
| 1708 | /* |
| 1709 | * Target node has been updated between RCU lookup and |
| 1710 | * lock acquisition. We need to re-try lookup and |
| 1711 | * attach. |
| 1712 | */ |
| 1713 | ret = -EAGAIN; |
| 1714 | goto unlock_parent; |
| 1715 | } |
| 1716 | |
| 1717 | /* Create new branch, starting from bottom */ |
| 1718 | CDS_INIT_HLIST_HEAD(&head); |
| 1719 | cds_hlist_add_head_rcu(&child_node->list, &head); |
| 1720 | iter_node_flag = (struct cds_ja_inode_flag *) head.next; |
| 1721 | |
| 1722 | for (i = ja->tree_depth; i > (int) level; i--) { |
| 1723 | uint8_t iter_key; |
| 1724 | |
| 1725 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i))); |
| 1726 | dbg_printf("branch creation level %d, key %u\n", |
| 1727 | i - 1, (unsigned int) iter_key); |
| 1728 | iter_dest_node_flag = NULL; |
| 1729 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, |
| 1730 | iter_key, |
| 1731 | iter_node_flag, |
| 1732 | NULL); |
| 1733 | if (ret) |
| 1734 | goto check_error; |
| 1735 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; |
| 1736 | iter_node_flag = iter_dest_node_flag; |
| 1737 | } |
| 1738 | |
| 1739 | if (level > 1) { |
| 1740 | uint8_t iter_key; |
| 1741 | |
| 1742 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); |
| 1743 | /* We need to use set_nth on the previous level. */ |
| 1744 | iter_dest_node_flag = node_flag; |
| 1745 | ret = ja_node_set_nth(ja, &iter_dest_node_flag, |
| 1746 | iter_key, |
| 1747 | iter_node_flag, |
| 1748 | shadow_node); |
| 1749 | if (ret) |
| 1750 | goto check_error; |
| 1751 | created_nodes[nr_created_nodes++] = iter_dest_node_flag; |
| 1752 | iter_node_flag = iter_dest_node_flag; |
| 1753 | } |
| 1754 | |
| 1755 | /* Publish new branch */ |
| 1756 | dbg_printf("Publish branch %p, replacing %p\n", |
| 1757 | iter_node_flag, *attach_node_flag_ptr); |
| 1758 | rcu_assign_pointer(*attach_node_flag_ptr, iter_node_flag); |
| 1759 | |
| 1760 | /* Success */ |
| 1761 | ret = 0; |
| 1762 | |
| 1763 | check_error: |
| 1764 | if (ret) { |
| 1765 | for (i = 0; i < nr_created_nodes; i++) { |
| 1766 | int tmpret; |
| 1767 | int flags; |
| 1768 | |
| 1769 | flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; |
| 1770 | if (i) |
| 1771 | flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; |
| 1772 | tmpret = rcuja_shadow_clear(ja->ht, |
| 1773 | created_nodes[i], |
| 1774 | NULL, |
| 1775 | flags); |
| 1776 | assert(!tmpret); |
| 1777 | } |
| 1778 | } |
| 1779 | unlock_parent: |
| 1780 | if (parent_shadow_node) |
| 1781 | rcuja_shadow_unlock(parent_shadow_node); |
| 1782 | unlock_shadow: |
| 1783 | if (shadow_node) |
| 1784 | rcuja_shadow_unlock(shadow_node); |
| 1785 | end: |
| 1786 | return ret; |
| 1787 | } |
| 1788 | |
| 1789 | /* |
| 1790 | * Lock the parent containing the hlist head pointer, and add node to list of |
| 1791 | * duplicates. Failure can happen if concurrent update changes the |
| 1792 | * parent before we get the lock. We return -EAGAIN in that case. |
| 1793 | * Return 0 on success, negative error value on failure. |
| 1794 | */ |
| 1795 | static |
| 1796 | int ja_chain_node(struct cds_ja *ja, |
| 1797 | struct cds_ja_inode_flag *parent_node_flag, |
| 1798 | struct cds_ja_inode_flag **node_flag_ptr, |
| 1799 | struct cds_ja_inode_flag *node_flag, |
| 1800 | struct cds_hlist_head *head, |
| 1801 | struct cds_ja_node *node) |
| 1802 | { |
| 1803 | struct cds_ja_shadow_node *shadow_node; |
| 1804 | int ret = 0; |
| 1805 | |
| 1806 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
| 1807 | if (!shadow_node) { |
| 1808 | return -EAGAIN; |
| 1809 | } |
| 1810 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
| 1811 | ret = -EAGAIN; |
| 1812 | goto end; |
| 1813 | } |
| 1814 | cds_hlist_add_head_rcu(&node->list, head); |
| 1815 | end: |
| 1816 | rcuja_shadow_unlock(shadow_node); |
| 1817 | return ret; |
| 1818 | } |
| 1819 | |
| 1820 | int cds_ja_add(struct cds_ja *ja, uint64_t key, |
| 1821 | struct cds_ja_node *new_node) |
| 1822 | { |
| 1823 | unsigned int tree_depth, i; |
| 1824 | struct cds_ja_inode_flag **attach_node_flag_ptr, |
| 1825 | **node_flag_ptr; |
| 1826 | struct cds_ja_inode_flag *node_flag, |
| 1827 | *parent_node_flag, |
| 1828 | *parent2_node_flag, |
| 1829 | *attach_node_flag; |
| 1830 | int ret; |
| 1831 | |
| 1832 | if (caa_unlikely(key > ja->key_max)) { |
| 1833 | return -EINVAL; |
| 1834 | } |
| 1835 | tree_depth = ja->tree_depth; |
| 1836 | |
| 1837 | retry: |
| 1838 | dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", |
| 1839 | key, new_node); |
| 1840 | parent2_node_flag = NULL; |
| 1841 | parent_node_flag = |
| 1842 | (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ |
| 1843 | attach_node_flag_ptr = &ja->root; |
| 1844 | attach_node_flag = rcu_dereference(ja->root); |
| 1845 | node_flag_ptr = &ja->root; |
| 1846 | node_flag = rcu_dereference(ja->root); |
| 1847 | |
| 1848 | /* Iterate on all internal levels */ |
| 1849 | for (i = 1; i < tree_depth; i++) { |
| 1850 | uint8_t iter_key; |
| 1851 | |
| 1852 | dbg_printf("cds_ja_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
| 1853 | attach_node_flag_ptr, node_flag_ptr, node_flag); |
| 1854 | if (!ja_node_ptr(node_flag)) { |
| 1855 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
| 1856 | attach_node_flag, |
| 1857 | node_flag_ptr, |
| 1858 | parent_node_flag, |
| 1859 | parent2_node_flag, |
| 1860 | key, i, new_node); |
| 1861 | if (ret == -EAGAIN || ret == -EEXIST) |
| 1862 | goto retry; |
| 1863 | else |
| 1864 | goto end; |
| 1865 | } |
| 1866 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
| 1867 | parent2_node_flag = parent_node_flag; |
| 1868 | parent_node_flag = node_flag; |
| 1869 | node_flag = ja_node_get_nth(node_flag, |
| 1870 | &attach_node_flag_ptr, |
| 1871 | &attach_node_flag, |
| 1872 | &node_flag_ptr, |
| 1873 | iter_key); |
| 1874 | dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n", |
| 1875 | (unsigned int) iter_key, node_flag, |
| 1876 | attach_node_flag_ptr, |
| 1877 | node_flag_ptr); |
| 1878 | } |
| 1879 | |
| 1880 | /* |
| 1881 | * We reached bottom of tree, simply add node to last internal |
| 1882 | * level, or chain it if key is already present. |
| 1883 | */ |
| 1884 | if (!ja_node_ptr(node_flag)) { |
| 1885 | dbg_printf("cds_ja_add attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", |
| 1886 | attach_node_flag_ptr, node_flag_ptr, node_flag); |
| 1887 | ret = ja_attach_node(ja, attach_node_flag_ptr, |
| 1888 | attach_node_flag, |
| 1889 | node_flag_ptr, parent_node_flag, |
| 1890 | parent2_node_flag, key, i, new_node); |
| 1891 | } else { |
| 1892 | ret = ja_chain_node(ja, |
| 1893 | parent_node_flag, |
| 1894 | node_flag_ptr, |
| 1895 | node_flag, |
| 1896 | (struct cds_hlist_head *) attach_node_flag_ptr, |
| 1897 | new_node); |
| 1898 | } |
| 1899 | if (ret == -EAGAIN || ret == -EEXIST) |
| 1900 | goto retry; |
| 1901 | end: |
| 1902 | return ret; |
| 1903 | } |
| 1904 | |
| 1905 | /* |
| 1906 | * Note: there is no need to lookup the pointer address associated with |
| 1907 | * each node's nth item after taking the lock: it's already been done by |
| 1908 | * cds_ja_del while holding the rcu read-side lock, and our node rules |
| 1909 | * ensure that when a match value -> pointer is found in a node, it is |
| 1910 | * _NEVER_ changed for that node without recompaction, and recompaction |
| 1911 | * reallocates the node. |
| 1912 | * However, when a child is removed from "linear" nodes, its pointer |
| 1913 | * is set to NULL. We therefore check, while holding the locks, if this |
| 1914 | * pointer is NULL, and return -ENOENT to the caller if it is the case. |
| 1915 | */ |
| 1916 | static |
| 1917 | int ja_detach_node(struct cds_ja *ja, |
| 1918 | struct cds_ja_inode_flag **snapshot, |
| 1919 | struct cds_ja_inode_flag ***snapshot_ptr, |
| 1920 | uint8_t *snapshot_n, |
| 1921 | int nr_snapshot, |
| 1922 | uint64_t key, |
| 1923 | struct cds_ja_node *node) |
| 1924 | { |
| 1925 | struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; |
| 1926 | struct cds_ja_inode_flag **node_flag_ptr = NULL, |
| 1927 | *parent_node_flag = NULL, |
| 1928 | **parent_node_flag_ptr = NULL; |
| 1929 | struct cds_ja_inode_flag *iter_node_flag; |
| 1930 | int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; |
| 1931 | uint8_t n = 0; |
| 1932 | |
| 1933 | assert(nr_snapshot == ja->tree_depth + 1); |
| 1934 | |
| 1935 | /* |
| 1936 | * From the last internal level node going up, get the node |
| 1937 | * lock, check if the node has only one child left. If it is the |
| 1938 | * case, we continue iterating upward. When we reach a node |
| 1939 | * which has more that one child left, we lock the parent, and |
| 1940 | * proceed to the node deletion (removing its children too). |
| 1941 | */ |
| 1942 | for (i = nr_snapshot - 2; i >= 1; i--) { |
| 1943 | struct cds_ja_shadow_node *shadow_node; |
| 1944 | |
| 1945 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
| 1946 | snapshot[i]); |
| 1947 | if (!shadow_node) { |
| 1948 | ret = -EAGAIN; |
| 1949 | goto end; |
| 1950 | } |
| 1951 | shadow_nodes[nr_shadow++] = shadow_node; |
| 1952 | |
| 1953 | /* |
| 1954 | * Check if node has been removed between RCU |
| 1955 | * lookup and lock acquisition. |
| 1956 | */ |
| 1957 | assert(snapshot_ptr[i + 1]); |
| 1958 | if (ja_node_ptr(*snapshot_ptr[i + 1]) |
| 1959 | != ja_node_ptr(snapshot[i + 1])) { |
| 1960 | ret = -ENOENT; |
| 1961 | goto end; |
| 1962 | } |
| 1963 | |
| 1964 | assert(shadow_node->nr_child > 0); |
| 1965 | if (shadow_node->nr_child == 1 && i > 1) |
| 1966 | nr_clear++; |
| 1967 | nr_branch++; |
| 1968 | if (shadow_node->nr_child > 1 || i == 1) { |
| 1969 | /* Lock parent and break */ |
| 1970 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
| 1971 | snapshot[i - 1]); |
| 1972 | if (!shadow_node) { |
| 1973 | ret = -EAGAIN; |
| 1974 | goto end; |
| 1975 | } |
| 1976 | shadow_nodes[nr_shadow++] = shadow_node; |
| 1977 | |
| 1978 | /* |
| 1979 | * Check if node has been removed between RCU |
| 1980 | * lookup and lock acquisition. |
| 1981 | */ |
| 1982 | assert(snapshot_ptr[i]); |
| 1983 | if (ja_node_ptr(*snapshot_ptr[i]) |
| 1984 | != ja_node_ptr(snapshot[i])) { |
| 1985 | ret = -ENOENT; |
| 1986 | goto end; |
| 1987 | } |
| 1988 | |
| 1989 | node_flag_ptr = snapshot_ptr[i + 1]; |
| 1990 | n = snapshot_n[i + 1]; |
| 1991 | parent_node_flag_ptr = snapshot_ptr[i]; |
| 1992 | parent_node_flag = snapshot[i]; |
| 1993 | |
| 1994 | if (i > 1) { |
| 1995 | /* |
| 1996 | * Lock parent's parent, in case we need |
| 1997 | * to recompact parent. |
| 1998 | */ |
| 1999 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, |
| 2000 | snapshot[i - 2]); |
| 2001 | if (!shadow_node) { |
| 2002 | ret = -EAGAIN; |
| 2003 | goto end; |
| 2004 | } |
| 2005 | shadow_nodes[nr_shadow++] = shadow_node; |
| 2006 | |
| 2007 | /* |
| 2008 | * Check if node has been removed between RCU |
| 2009 | * lookup and lock acquisition. |
| 2010 | */ |
| 2011 | assert(snapshot_ptr[i - 1]); |
| 2012 | if (ja_node_ptr(*snapshot_ptr[i - 1]) |
| 2013 | != ja_node_ptr(snapshot[i - 1])) { |
| 2014 | ret = -ENOENT; |
| 2015 | goto end; |
| 2016 | } |
| 2017 | } |
| 2018 | |
| 2019 | break; |
| 2020 | } |
| 2021 | } |
| 2022 | |
| 2023 | /* |
| 2024 | * At this point, we want to delete all nodes that are about to |
| 2025 | * be removed from shadow_nodes (except the last one, which is |
| 2026 | * either the root or the parent of the upmost node with 1 |
| 2027 | * child). OK to free lock here, because RCU read lock is held, |
| 2028 | * and free only performed in call_rcu. |
| 2029 | */ |
| 2030 | |
| 2031 | for (i = 0; i < nr_clear; i++) { |
| 2032 | ret = rcuja_shadow_clear(ja->ht, |
| 2033 | shadow_nodes[i]->node_flag, |
| 2034 | shadow_nodes[i], |
| 2035 | RCUJA_SHADOW_CLEAR_FREE_NODE |
| 2036 | | RCUJA_SHADOW_CLEAR_FREE_LOCK); |
| 2037 | assert(!ret); |
| 2038 | } |
| 2039 | |
| 2040 | iter_node_flag = parent_node_flag; |
| 2041 | /* Remove from parent */ |
| 2042 | ret = ja_node_clear_ptr(ja, |
| 2043 | node_flag_ptr, /* Pointer to location to nullify */ |
| 2044 | &iter_node_flag, /* Old new parent ptr in its parent */ |
| 2045 | shadow_nodes[nr_branch - 1], /* of parent */ |
| 2046 | n); |
| 2047 | if (ret) |
| 2048 | goto end; |
| 2049 | |
| 2050 | dbg_printf("ja_detach_node: publish %p instead of %p\n", |
| 2051 | iter_node_flag, *parent_node_flag_ptr); |
| 2052 | /* Update address of parent ptr in its parent */ |
| 2053 | rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); |
| 2054 | |
| 2055 | end: |
| 2056 | for (i = 0; i < nr_shadow; i++) |
| 2057 | rcuja_shadow_unlock(shadow_nodes[i]); |
| 2058 | return ret; |
| 2059 | } |
| 2060 | |
| 2061 | static |
| 2062 | int ja_unchain_node(struct cds_ja *ja, |
| 2063 | struct cds_ja_inode_flag *parent_node_flag, |
| 2064 | struct cds_ja_inode_flag **node_flag_ptr, |
| 2065 | struct cds_ja_inode_flag *node_flag, |
| 2066 | struct cds_ja_node *node) |
| 2067 | { |
| 2068 | struct cds_ja_shadow_node *shadow_node; |
| 2069 | struct cds_hlist_node *hlist_node; |
| 2070 | struct cds_hlist_head hlist_head; |
| 2071 | int ret = 0, count = 0, found = 0; |
| 2072 | |
| 2073 | shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); |
| 2074 | if (!shadow_node) |
| 2075 | return -EAGAIN; |
| 2076 | if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { |
| 2077 | ret = -EAGAIN; |
| 2078 | goto end; |
| 2079 | } |
| 2080 | hlist_head.next = (struct cds_hlist_node *) ja_node_ptr(node_flag); |
| 2081 | /* |
| 2082 | * Retry if another thread removed all but one of duplicates |
| 2083 | * since check (this check was performed without lock). |
| 2084 | * Ensure that the node we are about to remove is still in the |
| 2085 | * list (while holding lock). |
| 2086 | */ |
| 2087 | cds_hlist_for_each_rcu(hlist_node, &hlist_head) { |
| 2088 | if (count == 0) { |
| 2089 | /* FIXME: currently a work-around */ |
| 2090 | hlist_node->prev = (struct cds_hlist_node *) node_flag_ptr; |
| 2091 | } |
| 2092 | count++; |
| 2093 | if (hlist_node == &node->list) |
| 2094 | found++; |
| 2095 | } |
| 2096 | assert(found <= 1); |
| 2097 | if (!found || count == 1) { |
| 2098 | ret = -EAGAIN; |
| 2099 | goto end; |
| 2100 | } |
| 2101 | cds_hlist_del_rcu(&node->list); |
| 2102 | /* |
| 2103 | * Validate that we indeed removed the node from linked list. |
| 2104 | */ |
| 2105 | assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node); |
| 2106 | end: |
| 2107 | rcuja_shadow_unlock(shadow_node); |
| 2108 | return ret; |
| 2109 | } |
| 2110 | |
| 2111 | /* |
| 2112 | * Called with RCU read lock held. |
| 2113 | */ |
| 2114 | int cds_ja_del(struct cds_ja *ja, uint64_t key, |
| 2115 | struct cds_ja_node *node) |
| 2116 | { |
| 2117 | unsigned int tree_depth, i; |
| 2118 | struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; |
| 2119 | struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; |
| 2120 | uint8_t snapshot_n[JA_MAX_DEPTH]; |
| 2121 | struct cds_ja_inode_flag *node_flag; |
| 2122 | struct cds_ja_inode_flag **prev_node_flag_ptr, |
| 2123 | **node_flag_ptr; |
| 2124 | int nr_snapshot; |
| 2125 | int ret; |
| 2126 | |
| 2127 | if (caa_unlikely(key > ja->key_max)) |
| 2128 | return -EINVAL; |
| 2129 | tree_depth = ja->tree_depth; |
| 2130 | |
| 2131 | retry: |
| 2132 | nr_snapshot = 0; |
| 2133 | dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", |
| 2134 | key, node); |
| 2135 | |
| 2136 | /* snapshot for level 0 is only for shadow node lookup */ |
| 2137 | snapshot_n[0] = 0; |
| 2138 | snapshot_n[1] = 0; |
| 2139 | snapshot_ptr[nr_snapshot] = NULL; |
| 2140 | snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; |
| 2141 | node_flag = rcu_dereference(ja->root); |
| 2142 | prev_node_flag_ptr = &ja->root; |
| 2143 | node_flag_ptr = &ja->root; |
| 2144 | |
| 2145 | /* Iterate on all internal levels */ |
| 2146 | for (i = 1; i < tree_depth; i++) { |
| 2147 | uint8_t iter_key; |
| 2148 | |
| 2149 | dbg_printf("cds_ja_del iter node_flag %p\n", |
| 2150 | node_flag); |
| 2151 | if (!ja_node_ptr(node_flag)) { |
| 2152 | return -ENOENT; |
| 2153 | } |
| 2154 | iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); |
| 2155 | snapshot_n[nr_snapshot + 1] = iter_key; |
| 2156 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
| 2157 | snapshot[nr_snapshot++] = node_flag; |
| 2158 | node_flag = ja_node_get_nth(node_flag, |
| 2159 | &prev_node_flag_ptr, |
| 2160 | NULL, |
| 2161 | &node_flag_ptr, |
| 2162 | iter_key); |
| 2163 | dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", |
| 2164 | (unsigned int) iter_key, node_flag, |
| 2165 | prev_node_flag_ptr); |
| 2166 | } |
| 2167 | /* |
| 2168 | * We reached bottom of tree, try to find the node we are trying |
| 2169 | * to remove. Fail if we cannot find it. |
| 2170 | */ |
| 2171 | if (!ja_node_ptr(node_flag)) { |
| 2172 | dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", |
| 2173 | key); |
| 2174 | return -ENOENT; |
| 2175 | } else { |
| 2176 | struct cds_hlist_head hlist_head; |
| 2177 | struct cds_hlist_node *hlist_node; |
| 2178 | struct cds_ja_node *entry, *match = NULL; |
| 2179 | int count = 0; |
| 2180 | |
| 2181 | hlist_head.next = |
| 2182 | (struct cds_hlist_node *) ja_node_ptr(node_flag); |
| 2183 | cds_hlist_for_each_entry_rcu(entry, |
| 2184 | hlist_node, |
| 2185 | &hlist_head, |
| 2186 | list) { |
| 2187 | dbg_printf("cds_ja_del: compare %p with entry %p\n", node, entry); |
| 2188 | if (entry == node) |
| 2189 | match = entry; |
| 2190 | count++; |
| 2191 | } |
| 2192 | if (!match) { |
| 2193 | dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); |
| 2194 | return -ENOENT; |
| 2195 | } |
| 2196 | assert(count > 0); |
| 2197 | if (count == 1) { |
| 2198 | /* |
| 2199 | * Removing last of duplicates. Last snapshot |
| 2200 | * does not have a shadow node (external leafs). |
| 2201 | */ |
| 2202 | snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; |
| 2203 | snapshot[nr_snapshot++] = node_flag; |
| 2204 | ret = ja_detach_node(ja, snapshot, snapshot_ptr, |
| 2205 | snapshot_n, nr_snapshot, key, node); |
| 2206 | } else { |
| 2207 | ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], |
| 2208 | node_flag_ptr, node_flag, match); |
| 2209 | } |
| 2210 | } |
| 2211 | /* |
| 2212 | * Explanation of -ENOENT handling: caused by concurrent delete |
| 2213 | * between RCU lookup and actual removal. Need to re-do the |
| 2214 | * lookup and removal attempt. |
| 2215 | */ |
| 2216 | if (ret == -EAGAIN || ret == -ENOENT) |
| 2217 | goto retry; |
| 2218 | return ret; |
| 2219 | } |
| 2220 | |
| 2221 | struct cds_ja *_cds_ja_new(unsigned int key_bits, |
| 2222 | const struct rcu_flavor_struct *flavor) |
| 2223 | { |
| 2224 | struct cds_ja *ja; |
| 2225 | int ret; |
| 2226 | struct cds_ja_shadow_node *root_shadow_node; |
| 2227 | |
| 2228 | ja = calloc(sizeof(*ja), 1); |
| 2229 | if (!ja) |
| 2230 | goto ja_error; |
| 2231 | |
| 2232 | switch (key_bits) { |
| 2233 | case 8: |
| 2234 | case 16: |
| 2235 | case 24: |
| 2236 | case 32: |
| 2237 | case 40: |
| 2238 | case 48: |
| 2239 | case 56: |
| 2240 | ja->key_max = (1ULL << key_bits) - 1; |
| 2241 | break; |
| 2242 | case 64: |
| 2243 | ja->key_max = UINT64_MAX; |
| 2244 | break; |
| 2245 | default: |
| 2246 | goto check_error; |
| 2247 | } |
| 2248 | |
| 2249 | /* ja->root is NULL */ |
| 2250 | /* tree_depth 0 is for pointer to root node */ |
| 2251 | ja->tree_depth = (key_bits >> JA_LOG2_BITS_PER_BYTE) + 1; |
| 2252 | assert(ja->tree_depth <= JA_MAX_DEPTH); |
| 2253 | ja->ht = rcuja_create_ht(flavor); |
| 2254 | if (!ja->ht) |
| 2255 | goto ht_error; |
| 2256 | |
| 2257 | /* |
| 2258 | * Note: we should not free this node until judy array destroy. |
| 2259 | */ |
| 2260 | root_shadow_node = rcuja_shadow_set(ja->ht, |
| 2261 | (struct cds_ja_inode_flag *) &ja->root, |
| 2262 | NULL, ja); |
| 2263 | if (!root_shadow_node) { |
| 2264 | ret = -ENOMEM; |
| 2265 | goto ht_node_error; |
| 2266 | } |
| 2267 | root_shadow_node->level = 0; |
| 2268 | |
| 2269 | return ja; |
| 2270 | |
| 2271 | ht_node_error: |
| 2272 | ret = rcuja_delete_ht(ja->ht); |
| 2273 | assert(!ret); |
| 2274 | ht_error: |
| 2275 | check_error: |
| 2276 | free(ja); |
| 2277 | ja_error: |
| 2278 | return NULL; |
| 2279 | } |
| 2280 | |
| 2281 | /* |
| 2282 | * Called from RCU read-side CS. |
| 2283 | */ |
| 2284 | __attribute__((visibility("protected"))) |
| 2285 | void rcuja_free_all_children(struct cds_ja_shadow_node *shadow_node, |
| 2286 | struct cds_ja_inode_flag *node_flag, |
| 2287 | void (*free_node_cb)(struct rcu_head *head)) |
| 2288 | { |
| 2289 | const struct rcu_flavor_struct *flavor; |
| 2290 | unsigned int type_index; |
| 2291 | struct cds_ja_inode *node; |
| 2292 | const struct cds_ja_type *type; |
| 2293 | |
| 2294 | flavor = cds_lfht_rcu_flavor(shadow_node->ja->ht); |
| 2295 | node = ja_node_ptr(node_flag); |
| 2296 | assert(node != NULL); |
| 2297 | type_index = ja_node_type(node_flag); |
| 2298 | type = &ja_types[type_index]; |
| 2299 | |
| 2300 | switch (type->type_class) { |
| 2301 | case RCU_JA_LINEAR: |
| 2302 | { |
| 2303 | uint8_t nr_child = |
| 2304 | ja_linear_node_get_nr_child(type, node); |
| 2305 | unsigned int i; |
| 2306 | |
| 2307 | for (i = 0; i < nr_child; i++) { |
| 2308 | struct cds_ja_inode_flag *iter; |
| 2309 | struct cds_hlist_head head; |
| 2310 | struct cds_ja_node *entry; |
| 2311 | struct cds_hlist_node *pos; |
| 2312 | uint8_t v; |
| 2313 | |
| 2314 | ja_linear_node_get_ith_pos(type, node, i, &v, &iter); |
| 2315 | if (!iter) |
| 2316 | continue; |
| 2317 | head.next = (struct cds_hlist_node *) iter; |
| 2318 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { |
| 2319 | flavor->update_call_rcu(&entry->head, free_node_cb); |
| 2320 | } |
| 2321 | } |
| 2322 | break; |
| 2323 | } |
| 2324 | case RCU_JA_POOL: |
| 2325 | { |
| 2326 | unsigned int pool_nr; |
| 2327 | |
| 2328 | for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { |
| 2329 | struct cds_ja_inode *pool = |
| 2330 | ja_pool_node_get_ith_pool(type, node, pool_nr); |
| 2331 | uint8_t nr_child = |
| 2332 | ja_linear_node_get_nr_child(type, pool); |
| 2333 | unsigned int j; |
| 2334 | |
| 2335 | for (j = 0; j < nr_child; j++) { |
| 2336 | struct cds_ja_inode_flag *iter; |
| 2337 | struct cds_hlist_head head; |
| 2338 | struct cds_ja_node *entry; |
| 2339 | struct cds_hlist_node *pos; |
| 2340 | uint8_t v; |
| 2341 | |
| 2342 | ja_linear_node_get_ith_pos(type, node, j, &v, &iter); |
| 2343 | if (!iter) |
| 2344 | continue; |
| 2345 | head.next = (struct cds_hlist_node *) iter; |
| 2346 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { |
| 2347 | flavor->update_call_rcu(&entry->head, free_node_cb); |
| 2348 | } |
| 2349 | } |
| 2350 | } |
| 2351 | break; |
| 2352 | } |
| 2353 | case RCU_JA_NULL: |
| 2354 | break; |
| 2355 | case RCU_JA_PIGEON: |
| 2356 | { |
| 2357 | uint8_t nr_child; |
| 2358 | unsigned int i; |
| 2359 | |
| 2360 | nr_child = shadow_node->nr_child; |
| 2361 | for (i = 0; i < nr_child; i++) { |
| 2362 | struct cds_ja_inode_flag *iter; |
| 2363 | struct cds_hlist_head head; |
| 2364 | struct cds_ja_node *entry; |
| 2365 | struct cds_hlist_node *pos; |
| 2366 | |
| 2367 | iter = ja_pigeon_node_get_ith_pos(type, node, i); |
| 2368 | if (!iter) |
| 2369 | continue; |
| 2370 | head.next = (struct cds_hlist_node *) iter; |
| 2371 | cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { |
| 2372 | flavor->update_call_rcu(&entry->head, free_node_cb); |
| 2373 | } |
| 2374 | } |
| 2375 | break; |
| 2376 | } |
| 2377 | default: |
| 2378 | assert(0); |
| 2379 | } |
| 2380 | } |
| 2381 | |
| 2382 | static |
| 2383 | void print_debug_fallback_distribution(void) |
| 2384 | { |
| 2385 | int i; |
| 2386 | |
| 2387 | fprintf(stderr, "Fallback node distribution:\n"); |
| 2388 | for (i = 0; i < JA_ENTRY_PER_NODE; i++) { |
| 2389 | if (!node_fallback_count_distribution[i]) |
| 2390 | continue; |
| 2391 | fprintf(stderr, " %3u: %4lu\n", |
| 2392 | i, node_fallback_count_distribution[i]); |
| 2393 | } |
| 2394 | } |
| 2395 | |
| 2396 | /* |
| 2397 | * There should be no more concurrent add to the judy array while it is |
| 2398 | * being destroyed (ensured by the caller). |
| 2399 | */ |
| 2400 | int cds_ja_destroy(struct cds_ja *ja, |
| 2401 | void (*free_node_cb)(struct rcu_head *head)) |
| 2402 | { |
| 2403 | int ret; |
| 2404 | |
| 2405 | rcuja_shadow_prune(ja->ht, |
| 2406 | RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK, |
| 2407 | free_node_cb); |
| 2408 | ret = rcuja_delete_ht(ja->ht); |
| 2409 | if (ret) |
| 2410 | return ret; |
| 2411 | if (uatomic_read(&ja->nr_fallback)) |
| 2412 | fprintf(stderr, |
| 2413 | "[warning] RCU Judy Array used %lu fallback node(s)\n", |
| 2414 | uatomic_read(&ja->nr_fallback)); |
| 2415 | print_debug_fallback_distribution(); |
| 2416 | free(ja); |
| 2417 | return 0; |
| 2418 | } |