| 1 | /* |
| 2 | * rculfhash.c |
| 3 | * |
| 4 | * Userspace RCU library - Lock-Free Resizable RCU Hash Table |
| 5 | * |
| 6 | * Copyright 2010-2011 - 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 | /* |
| 24 | * Based on the following articles: |
| 25 | * - Ori Shalev and Nir Shavit. Split-ordered lists: Lock-free |
| 26 | * extensible hash tables. J. ACM 53, 3 (May 2006), 379-405. |
| 27 | * - Michael, M. M. High performance dynamic lock-free hash tables |
| 28 | * and list-based sets. In Proceedings of the fourteenth annual ACM |
| 29 | * symposium on Parallel algorithms and architectures, ACM Press, |
| 30 | * (2002), 73-82. |
| 31 | * |
| 32 | * Some specificities of this Lock-Free Resizable RCU Hash Table |
| 33 | * implementation: |
| 34 | * |
| 35 | * - RCU read-side critical section allows readers to perform hash |
| 36 | * table lookups and use the returned objects safely by delaying |
| 37 | * memory reclaim of a grace period. |
| 38 | * - Add and remove operations are lock-free, and do not need to |
| 39 | * allocate memory. They need to be executed within RCU read-side |
| 40 | * critical section to ensure the objects they read are valid and to |
| 41 | * deal with the cmpxchg ABA problem. |
| 42 | * - add and add_unique operations are supported. add_unique checks if |
| 43 | * the node key already exists in the hash table. It ensures no key |
| 44 | * duplicata exists. |
| 45 | * - The resize operation executes concurrently with add/remove/lookup. |
| 46 | * - Hash table nodes are contained within a split-ordered list. This |
| 47 | * list is ordered by incrementing reversed-bits-hash value. |
| 48 | * - An index of dummy nodes is kept. These dummy nodes are the hash |
| 49 | * table "buckets", and they are also chained together in the |
| 50 | * split-ordered list, which allows recursive expansion. |
| 51 | * - The resize operation for small tables only allows expanding the hash table. |
| 52 | * It is triggered automatically by detecting long chains in the add |
| 53 | * operation. |
| 54 | * - The resize operation for larger tables (and available through an |
| 55 | * API) allows both expanding and shrinking the hash table. |
| 56 | * - Per-CPU Split-counters are used to keep track of the number of |
| 57 | * nodes within the hash table for automatic resize triggering. |
| 58 | * - Resize operation initiated by long chain detection is executed by a |
| 59 | * call_rcu thread, which keeps lock-freedom of add and remove. |
| 60 | * - Resize operations are protected by a mutex. |
| 61 | * - The removal operation is split in two parts: first, a "removed" |
| 62 | * flag is set in the next pointer within the node to remove. Then, |
| 63 | * a "garbage collection" is performed in the bucket containing the |
| 64 | * removed node (from the start of the bucket up to the removed node). |
| 65 | * All encountered nodes with "removed" flag set in their next |
| 66 | * pointers are removed from the linked-list. If the cmpxchg used for |
| 67 | * removal fails (due to concurrent garbage-collection or concurrent |
| 68 | * add), we retry from the beginning of the bucket. This ensures that |
| 69 | * the node with "removed" flag set is removed from the hash table |
| 70 | * (not visible to lookups anymore) before the RCU read-side critical |
| 71 | * section held across removal ends. Furthermore, this ensures that |
| 72 | * the node with "removed" flag set is removed from the linked-list |
| 73 | * before its memory is reclaimed. Only the thread which removal |
| 74 | * successfully set the "removed" flag (with a cmpxchg) into a node's |
| 75 | * next pointer is considered to have succeeded its removal (and thus |
| 76 | * owns the node to reclaim). Because we garbage-collect starting from |
| 77 | * an invariant node (the start-of-bucket dummy node) up to the |
| 78 | * "removed" node (or find a reverse-hash that is higher), we are sure |
| 79 | * that a successful traversal of the chain leads to a chain that is |
| 80 | * present in the linked-list (the start node is never removed) and |
| 81 | * that is does not contain the "removed" node anymore, even if |
| 82 | * concurrent delete/add operations are changing the structure of the |
| 83 | * list concurrently. |
| 84 | * - The add operation performs gargage collection of buckets if it |
| 85 | * encounters nodes with removed flag set in the bucket where it wants |
| 86 | * to add its new node. This ensures lock-freedom of add operation by |
| 87 | * helping the remover unlink nodes from the list rather than to wait |
| 88 | * for it do to so. |
| 89 | * - A RCU "order table" indexed by log2(hash index) is copied and |
| 90 | * expanded by the resize operation. This order table allows finding |
| 91 | * the "dummy node" tables. |
| 92 | * - There is one dummy node table per hash index order. The size of |
| 93 | * each dummy node table is half the number of hashes contained in |
| 94 | * this order. |
| 95 | * - call_rcu is used to garbage-collect the old order table. |
| 96 | * - The per-order dummy node tables contain a compact version of the |
| 97 | * hash table nodes. These tables are invariant after they are |
| 98 | * populated into the hash table. |
| 99 | * |
| 100 | * A bit of ascii art explanation: |
| 101 | * |
| 102 | * Order index is the off-by-one compare to the actual power of 2 because |
| 103 | * we use index 0 to deal with the 0 special-case. |
| 104 | * |
| 105 | * This shows the nodes for a small table ordered by reversed bits: |
| 106 | * |
| 107 | * bits reverse |
| 108 | * 0 000 000 |
| 109 | * 4 100 001 |
| 110 | * 2 010 010 |
| 111 | * 6 110 011 |
| 112 | * 1 001 100 |
| 113 | * 5 101 101 |
| 114 | * 3 011 110 |
| 115 | * 7 111 111 |
| 116 | * |
| 117 | * This shows the nodes in order of non-reversed bits, linked by |
| 118 | * reversed-bit order. |
| 119 | * |
| 120 | * order bits reverse |
| 121 | * 0 0 000 000 |
| 122 | * | |
| 123 | * 1 | 1 001 100 <- <- |
| 124 | * | | | | |
| 125 | * 2 | | 2 010 010 | | |
| 126 | * | | | 3 011 110 | <- | |
| 127 | * | | | | | | | |
| 128 | * 3 -> | | | 4 100 001 | | |
| 129 | * -> | | 5 101 101 | |
| 130 | * -> | 6 110 011 |
| 131 | * -> 7 111 111 |
| 132 | */ |
| 133 | |
| 134 | #define _LGPL_SOURCE |
| 135 | #include <stdlib.h> |
| 136 | #include <errno.h> |
| 137 | #include <assert.h> |
| 138 | #include <stdio.h> |
| 139 | #include <stdint.h> |
| 140 | #include <string.h> |
| 141 | |
| 142 | #include "config.h" |
| 143 | #include <urcu.h> |
| 144 | #include <urcu-call-rcu.h> |
| 145 | #include <urcu/arch.h> |
| 146 | #include <urcu/uatomic.h> |
| 147 | #include <urcu/jhash.h> |
| 148 | #include <urcu/compiler.h> |
| 149 | #include <urcu/rculfhash.h> |
| 150 | #include <stdio.h> |
| 151 | #include <pthread.h> |
| 152 | |
| 153 | #ifdef DEBUG |
| 154 | #define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args) |
| 155 | #else |
| 156 | #define dbg_printf(fmt, args...) |
| 157 | #endif |
| 158 | |
| 159 | /* |
| 160 | * Per-CPU split-counters lazily update the global counter each 1024 |
| 161 | * addition/removal. It automatically keeps track of resize required. |
| 162 | * We use the bucket length as indicator for need to expand for small |
| 163 | * tables and machines lacking per-cpu data suppport. |
| 164 | */ |
| 165 | #define COUNT_COMMIT_ORDER 10 |
| 166 | #define CHAIN_LEN_TARGET 1 |
| 167 | #define CHAIN_LEN_RESIZE_THRESHOLD 3 |
| 168 | |
| 169 | /* |
| 170 | * Define the minimum table size. Protects against hash table resize overload |
| 171 | * when too many entries are added quickly before the resize can complete. |
| 172 | * This is especially the case if the table could be shrinked to a size of 1. |
| 173 | * TODO: we might want to make the add/remove operations help the resize to |
| 174 | * add or remove dummy nodes when a resize is ongoing to ensure upper-bound on |
| 175 | * chain length. |
| 176 | */ |
| 177 | #define MIN_TABLE_SIZE 128 |
| 178 | |
| 179 | #if (CAA_BITS_PER_LONG == 32) |
| 180 | #define MAX_TABLE_ORDER 32 |
| 181 | #else |
| 182 | #define MAX_TABLE_ORDER 64 |
| 183 | #endif |
| 184 | |
| 185 | #ifndef min |
| 186 | #define min(a, b) ((a) < (b) ? (a) : (b)) |
| 187 | #endif |
| 188 | |
| 189 | #ifndef max |
| 190 | #define max(a, b) ((a) > (b) ? (a) : (b)) |
| 191 | #endif |
| 192 | |
| 193 | /* |
| 194 | * The removed flag needs to be updated atomically with the pointer. |
| 195 | * The dummy flag does not require to be updated atomically with the |
| 196 | * pointer, but it is added as a pointer low bit flag to save space. |
| 197 | */ |
| 198 | #define REMOVED_FLAG (1UL << 0) |
| 199 | #define DUMMY_FLAG (1UL << 1) |
| 200 | #define FLAGS_MASK ((1UL << 2) - 1) |
| 201 | |
| 202 | struct ht_items_count { |
| 203 | unsigned long add, remove; |
| 204 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
| 205 | |
| 206 | struct rcu_level { |
| 207 | struct rcu_head head; |
| 208 | struct _cds_lfht_node nodes[0]; |
| 209 | }; |
| 210 | |
| 211 | struct rcu_table { |
| 212 | unsigned long size; /* always a power of 2, shared (RCU) */ |
| 213 | unsigned long resize_target; |
| 214 | int resize_initiated; |
| 215 | struct rcu_level *tbl[MAX_TABLE_ORDER]; |
| 216 | }; |
| 217 | |
| 218 | struct cds_lfht { |
| 219 | struct rcu_table t; |
| 220 | cds_lfht_hash_fct hash_fct; |
| 221 | cds_lfht_compare_fct compare_fct; |
| 222 | unsigned long hash_seed; |
| 223 | int flags; |
| 224 | /* |
| 225 | * We need to put the work threads offline (QSBR) when taking this |
| 226 | * mutex, because we use synchronize_rcu within this mutex critical |
| 227 | * section, which waits on read-side critical sections, and could |
| 228 | * therefore cause grace-period deadlock if we hold off RCU G.P. |
| 229 | * completion. |
| 230 | */ |
| 231 | pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ |
| 232 | unsigned int in_progress_resize, in_progress_destroy; |
| 233 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
| 234 | void (*func)(struct rcu_head *head)); |
| 235 | void (*cds_lfht_synchronize_rcu)(void); |
| 236 | void (*cds_lfht_rcu_read_lock)(void); |
| 237 | void (*cds_lfht_rcu_read_unlock)(void); |
| 238 | void (*cds_lfht_rcu_thread_offline)(void); |
| 239 | void (*cds_lfht_rcu_thread_online)(void); |
| 240 | unsigned long count; /* global approximate item count */ |
| 241 | struct ht_items_count *percpu_count; /* per-cpu item count */ |
| 242 | }; |
| 243 | |
| 244 | struct rcu_resize_work { |
| 245 | struct rcu_head head; |
| 246 | struct cds_lfht *ht; |
| 247 | }; |
| 248 | |
| 249 | /* |
| 250 | * Algorithm to reverse bits in a word by lookup table, extended to |
| 251 | * 64-bit words. |
| 252 | * Source: |
| 253 | * http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable |
| 254 | * Originally from Public Domain. |
| 255 | */ |
| 256 | |
| 257 | static const uint8_t BitReverseTable256[256] = |
| 258 | { |
| 259 | #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 |
| 260 | #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) |
| 261 | #define R6(n) R4(n), R4((n) + 2*4 ), R4((n) + 1*4 ), R4((n) + 3*4 ) |
| 262 | R6(0), R6(2), R6(1), R6(3) |
| 263 | }; |
| 264 | #undef R2 |
| 265 | #undef R4 |
| 266 | #undef R6 |
| 267 | |
| 268 | static |
| 269 | uint8_t bit_reverse_u8(uint8_t v) |
| 270 | { |
| 271 | return BitReverseTable256[v]; |
| 272 | } |
| 273 | |
| 274 | static __attribute__((unused)) |
| 275 | uint32_t bit_reverse_u32(uint32_t v) |
| 276 | { |
| 277 | return ((uint32_t) bit_reverse_u8(v) << 24) | |
| 278 | ((uint32_t) bit_reverse_u8(v >> 8) << 16) | |
| 279 | ((uint32_t) bit_reverse_u8(v >> 16) << 8) | |
| 280 | ((uint32_t) bit_reverse_u8(v >> 24)); |
| 281 | } |
| 282 | |
| 283 | static __attribute__((unused)) |
| 284 | uint64_t bit_reverse_u64(uint64_t v) |
| 285 | { |
| 286 | return ((uint64_t) bit_reverse_u8(v) << 56) | |
| 287 | ((uint64_t) bit_reverse_u8(v >> 8) << 48) | |
| 288 | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | |
| 289 | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | |
| 290 | ((uint64_t) bit_reverse_u8(v >> 32) << 24) | |
| 291 | ((uint64_t) bit_reverse_u8(v >> 40) << 16) | |
| 292 | ((uint64_t) bit_reverse_u8(v >> 48) << 8) | |
| 293 | ((uint64_t) bit_reverse_u8(v >> 56)); |
| 294 | } |
| 295 | |
| 296 | static |
| 297 | unsigned long bit_reverse_ulong(unsigned long v) |
| 298 | { |
| 299 | #if (CAA_BITS_PER_LONG == 32) |
| 300 | return bit_reverse_u32(v); |
| 301 | #else |
| 302 | return bit_reverse_u64(v); |
| 303 | #endif |
| 304 | } |
| 305 | |
| 306 | /* |
| 307 | * fls: returns the position of the most significant bit. |
| 308 | * Returns 0 if no bit is set, else returns the position of the most |
| 309 | * significant bit (from 1 to 32 on 32-bit, from 1 to 64 on 64-bit). |
| 310 | */ |
| 311 | #if defined(__i386) || defined(__x86_64) |
| 312 | static inline |
| 313 | unsigned int fls_u32(uint32_t x) |
| 314 | { |
| 315 | int r; |
| 316 | |
| 317 | asm("bsrl %1,%0\n\t" |
| 318 | "jnz 1f\n\t" |
| 319 | "movl $-1,%0\n\t" |
| 320 | "1:\n\t" |
| 321 | : "=r" (r) : "rm" (x)); |
| 322 | return r + 1; |
| 323 | } |
| 324 | #define HAS_FLS_U32 |
| 325 | #endif |
| 326 | |
| 327 | #if defined(__x86_64) |
| 328 | static inline |
| 329 | unsigned int fls_u64(uint64_t x) |
| 330 | { |
| 331 | long r; |
| 332 | |
| 333 | asm("bsrq %1,%0\n\t" |
| 334 | "jnz 1f\n\t" |
| 335 | "movq $-1,%0\n\t" |
| 336 | "1:\n\t" |
| 337 | : "=r" (r) : "rm" (x)); |
| 338 | return r + 1; |
| 339 | } |
| 340 | #define HAS_FLS_U64 |
| 341 | #endif |
| 342 | |
| 343 | #ifndef HAS_FLS_U64 |
| 344 | static __attribute__((unused)) |
| 345 | unsigned int fls_u64(uint64_t x) |
| 346 | { |
| 347 | unsigned int r = 64; |
| 348 | |
| 349 | if (!x) |
| 350 | return 0; |
| 351 | |
| 352 | if (!(x & 0xFFFFFFFF00000000ULL)) { |
| 353 | x <<= 32; |
| 354 | r -= 32; |
| 355 | } |
| 356 | if (!(x & 0xFFFF000000000000ULL)) { |
| 357 | x <<= 16; |
| 358 | r -= 16; |
| 359 | } |
| 360 | if (!(x & 0xFF00000000000000ULL)) { |
| 361 | x <<= 8; |
| 362 | r -= 8; |
| 363 | } |
| 364 | if (!(x & 0xF000000000000000ULL)) { |
| 365 | x <<= 4; |
| 366 | r -= 4; |
| 367 | } |
| 368 | if (!(x & 0xC000000000000000ULL)) { |
| 369 | x <<= 2; |
| 370 | r -= 2; |
| 371 | } |
| 372 | if (!(x & 0x8000000000000000ULL)) { |
| 373 | x <<= 1; |
| 374 | r -= 1; |
| 375 | } |
| 376 | return r; |
| 377 | } |
| 378 | #endif |
| 379 | |
| 380 | #ifndef HAS_FLS_U32 |
| 381 | static __attribute__((unused)) |
| 382 | unsigned int fls_u32(uint32_t x) |
| 383 | { |
| 384 | unsigned int r = 32; |
| 385 | |
| 386 | if (!x) |
| 387 | return 0; |
| 388 | if (!(x & 0xFFFF0000U)) { |
| 389 | x <<= 16; |
| 390 | r -= 16; |
| 391 | } |
| 392 | if (!(x & 0xFF000000U)) { |
| 393 | x <<= 8; |
| 394 | r -= 8; |
| 395 | } |
| 396 | if (!(x & 0xF0000000U)) { |
| 397 | x <<= 4; |
| 398 | r -= 4; |
| 399 | } |
| 400 | if (!(x & 0xC0000000U)) { |
| 401 | x <<= 2; |
| 402 | r -= 2; |
| 403 | } |
| 404 | if (!(x & 0x80000000U)) { |
| 405 | x <<= 1; |
| 406 | r -= 1; |
| 407 | } |
| 408 | return r; |
| 409 | } |
| 410 | #endif |
| 411 | |
| 412 | unsigned int fls_ulong(unsigned long x) |
| 413 | { |
| 414 | #if (CAA_BITS_PER_lONG == 32) |
| 415 | return fls_u32(x); |
| 416 | #else |
| 417 | return fls_u64(x); |
| 418 | #endif |
| 419 | } |
| 420 | |
| 421 | int get_count_order_u32(uint32_t x) |
| 422 | { |
| 423 | int order; |
| 424 | |
| 425 | order = fls_u32(x) - 1; |
| 426 | if (x & (x - 1)) |
| 427 | order++; |
| 428 | return order; |
| 429 | } |
| 430 | |
| 431 | int get_count_order_ulong(unsigned long x) |
| 432 | { |
| 433 | int order; |
| 434 | |
| 435 | order = fls_ulong(x) - 1; |
| 436 | if (x & (x - 1)) |
| 437 | order++; |
| 438 | return order; |
| 439 | } |
| 440 | |
| 441 | #ifdef POISON_FREE |
| 442 | #define poison_free(ptr) \ |
| 443 | do { \ |
| 444 | memset(ptr, 0x42, sizeof(*(ptr))); \ |
| 445 | free(ptr); \ |
| 446 | } while (0) |
| 447 | #else |
| 448 | #define poison_free(ptr) free(ptr) |
| 449 | #endif |
| 450 | |
| 451 | static |
| 452 | void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth); |
| 453 | |
| 454 | /* |
| 455 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are |
| 456 | * available, then we support hash table item accounting. |
| 457 | * In the unfortunate event the number of CPUs reported would be |
| 458 | * inaccurate, we use modulo arithmetic on the number of CPUs we got. |
| 459 | */ |
| 460 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
| 461 | |
| 462 | static |
| 463 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
| 464 | unsigned long count); |
| 465 | |
| 466 | static long nr_cpus_mask = -1; |
| 467 | |
| 468 | static |
| 469 | struct ht_items_count *alloc_per_cpu_items_count(void) |
| 470 | { |
| 471 | struct ht_items_count *count; |
| 472 | |
| 473 | switch (nr_cpus_mask) { |
| 474 | case -2: |
| 475 | return NULL; |
| 476 | case -1: |
| 477 | { |
| 478 | long maxcpus; |
| 479 | |
| 480 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); |
| 481 | if (maxcpus <= 0) { |
| 482 | nr_cpus_mask = -2; |
| 483 | return NULL; |
| 484 | } |
| 485 | /* |
| 486 | * round up number of CPUs to next power of two, so we |
| 487 | * can use & for modulo. |
| 488 | */ |
| 489 | maxcpus = 1UL << get_count_order_ulong(maxcpus); |
| 490 | nr_cpus_mask = maxcpus - 1; |
| 491 | } |
| 492 | /* Fall-through */ |
| 493 | default: |
| 494 | return calloc(nr_cpus_mask + 1, sizeof(*count)); |
| 495 | } |
| 496 | } |
| 497 | |
| 498 | static |
| 499 | void free_per_cpu_items_count(struct ht_items_count *count) |
| 500 | { |
| 501 | poison_free(count); |
| 502 | } |
| 503 | |
| 504 | static |
| 505 | int ht_get_cpu(void) |
| 506 | { |
| 507 | int cpu; |
| 508 | |
| 509 | assert(nr_cpus_mask >= 0); |
| 510 | cpu = sched_getcpu(); |
| 511 | if (unlikely(cpu < 0)) |
| 512 | return cpu; |
| 513 | else |
| 514 | return cpu & nr_cpus_mask; |
| 515 | } |
| 516 | |
| 517 | static |
| 518 | void ht_count_add(struct cds_lfht *ht, unsigned long size) |
| 519 | { |
| 520 | unsigned long percpu_count; |
| 521 | int cpu; |
| 522 | |
| 523 | if (unlikely(!ht->percpu_count)) |
| 524 | return; |
| 525 | cpu = ht_get_cpu(); |
| 526 | if (unlikely(cpu < 0)) |
| 527 | return; |
| 528 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1); |
| 529 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
| 530 | unsigned long count; |
| 531 | |
| 532 | dbg_printf("add percpu %lu\n", percpu_count); |
| 533 | count = uatomic_add_return(&ht->count, |
| 534 | 1UL << COUNT_COMMIT_ORDER); |
| 535 | /* If power of 2 */ |
| 536 | if (!(count & (count - 1))) { |
| 537 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) |
| 538 | return; |
| 539 | dbg_printf("add set global %lu\n", count); |
| 540 | cds_lfht_resize_lazy_count(ht, size, |
| 541 | count >> (CHAIN_LEN_TARGET - 1)); |
| 542 | } |
| 543 | } |
| 544 | } |
| 545 | |
| 546 | static |
| 547 | void ht_count_remove(struct cds_lfht *ht, unsigned long size) |
| 548 | { |
| 549 | unsigned long percpu_count; |
| 550 | int cpu; |
| 551 | |
| 552 | if (unlikely(!ht->percpu_count)) |
| 553 | return; |
| 554 | cpu = ht_get_cpu(); |
| 555 | if (unlikely(cpu < 0)) |
| 556 | return; |
| 557 | percpu_count = uatomic_add_return(&ht->percpu_count[cpu].remove, -1); |
| 558 | if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { |
| 559 | unsigned long count; |
| 560 | |
| 561 | dbg_printf("remove percpu %lu\n", percpu_count); |
| 562 | count = uatomic_add_return(&ht->count, |
| 563 | -(1UL << COUNT_COMMIT_ORDER)); |
| 564 | /* If power of 2 */ |
| 565 | if (!(count & (count - 1))) { |
| 566 | if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) |
| 567 | return; |
| 568 | dbg_printf("remove set global %lu\n", count); |
| 569 | cds_lfht_resize_lazy_count(ht, size, |
| 570 | count >> (CHAIN_LEN_TARGET - 1)); |
| 571 | } |
| 572 | } |
| 573 | } |
| 574 | |
| 575 | #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ |
| 576 | |
| 577 | static const long nr_cpus_mask = -1; |
| 578 | |
| 579 | static |
| 580 | struct ht_items_count *alloc_per_cpu_items_count(void) |
| 581 | { |
| 582 | return NULL; |
| 583 | } |
| 584 | |
| 585 | static |
| 586 | void free_per_cpu_items_count(struct ht_items_count *count) |
| 587 | { |
| 588 | } |
| 589 | |
| 590 | static |
| 591 | void ht_count_add(struct cds_lfht *ht, unsigned long size) |
| 592 | { |
| 593 | } |
| 594 | |
| 595 | static |
| 596 | void ht_count_remove(struct cds_lfht *ht, unsigned long size) |
| 597 | { |
| 598 | } |
| 599 | |
| 600 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ |
| 601 | |
| 602 | |
| 603 | static |
| 604 | void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) |
| 605 | { |
| 606 | unsigned long count; |
| 607 | |
| 608 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
| 609 | return; |
| 610 | count = uatomic_read(&ht->count); |
| 611 | /* |
| 612 | * Use bucket-local length for small table expand and for |
| 613 | * environments lacking per-cpu data support. |
| 614 | */ |
| 615 | if (count >= (1UL << COUNT_COMMIT_ORDER)) |
| 616 | return; |
| 617 | if (chain_len > 100) |
| 618 | dbg_printf("WARNING: large chain length: %u.\n", |
| 619 | chain_len); |
| 620 | if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) |
| 621 | cds_lfht_resize_lazy(ht, size, |
| 622 | get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); |
| 623 | } |
| 624 | |
| 625 | static |
| 626 | struct cds_lfht_node *clear_flag(struct cds_lfht_node *node) |
| 627 | { |
| 628 | return (struct cds_lfht_node *) (((unsigned long) node) & ~FLAGS_MASK); |
| 629 | } |
| 630 | |
| 631 | static |
| 632 | int is_removed(struct cds_lfht_node *node) |
| 633 | { |
| 634 | return ((unsigned long) node) & REMOVED_FLAG; |
| 635 | } |
| 636 | |
| 637 | static |
| 638 | struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) |
| 639 | { |
| 640 | return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); |
| 641 | } |
| 642 | |
| 643 | static |
| 644 | int is_dummy(struct cds_lfht_node *node) |
| 645 | { |
| 646 | return ((unsigned long) node) & DUMMY_FLAG; |
| 647 | } |
| 648 | |
| 649 | static |
| 650 | struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node) |
| 651 | { |
| 652 | return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG); |
| 653 | } |
| 654 | |
| 655 | static |
| 656 | unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) |
| 657 | { |
| 658 | unsigned long old1, old2; |
| 659 | |
| 660 | old1 = uatomic_read(ptr); |
| 661 | do { |
| 662 | old2 = old1; |
| 663 | if (old2 >= v) |
| 664 | return old2; |
| 665 | } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); |
| 666 | return v; |
| 667 | } |
| 668 | |
| 669 | static |
| 670 | void cds_lfht_free_level(struct rcu_head *head) |
| 671 | { |
| 672 | struct rcu_level *l = |
| 673 | caa_container_of(head, struct rcu_level, head); |
| 674 | poison_free(l); |
| 675 | } |
| 676 | |
| 677 | /* |
| 678 | * Remove all logically deleted nodes from a bucket up to a certain node key. |
| 679 | */ |
| 680 | static |
| 681 | void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node) |
| 682 | { |
| 683 | struct cds_lfht_node *iter_prev, *iter, *next, *new_next; |
| 684 | |
| 685 | assert(!is_dummy(dummy)); |
| 686 | assert(!is_removed(dummy)); |
| 687 | assert(!is_dummy(node)); |
| 688 | assert(!is_removed(node)); |
| 689 | for (;;) { |
| 690 | iter_prev = dummy; |
| 691 | /* We can always skip the dummy node initially */ |
| 692 | iter = rcu_dereference(iter_prev->p.next); |
| 693 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); |
| 694 | /* |
| 695 | * We should never be called with dummy (start of chain) |
| 696 | * and logically removed node (end of path compression |
| 697 | * marker) being the actual same node. This would be a |
| 698 | * bug in the algorithm implementation. |
| 699 | */ |
| 700 | assert(dummy != node); |
| 701 | for (;;) { |
| 702 | if (unlikely(!clear_flag(iter))) |
| 703 | return; |
| 704 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
| 705 | return; |
| 706 | next = rcu_dereference(clear_flag(iter)->p.next); |
| 707 | if (likely(is_removed(next))) |
| 708 | break; |
| 709 | iter_prev = clear_flag(iter); |
| 710 | iter = next; |
| 711 | } |
| 712 | assert(!is_removed(iter)); |
| 713 | if (is_dummy(iter)) |
| 714 | new_next = flag_dummy(clear_flag(next)); |
| 715 | else |
| 716 | new_next = clear_flag(next); |
| 717 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); |
| 718 | } |
| 719 | } |
| 720 | |
| 721 | static |
| 722 | struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht, |
| 723 | unsigned long size, |
| 724 | struct cds_lfht_node *node, |
| 725 | int unique, int dummy) |
| 726 | { |
| 727 | struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, |
| 728 | *dummy_node; |
| 729 | struct _cds_lfht_node *lookup; |
| 730 | unsigned long hash, index, order; |
| 731 | |
| 732 | assert(!is_dummy(node)); |
| 733 | assert(!is_removed(node)); |
| 734 | if (!size) { |
| 735 | assert(dummy); |
| 736 | node->p.next = flag_dummy(NULL); |
| 737 | return node; /* Initial first add (head) */ |
| 738 | } |
| 739 | hash = bit_reverse_ulong(node->p.reverse_hash); |
| 740 | for (;;) { |
| 741 | uint32_t chain_len = 0; |
| 742 | |
| 743 | /* |
| 744 | * iter_prev points to the non-removed node prior to the |
| 745 | * insert location. |
| 746 | */ |
| 747 | index = hash & (size - 1); |
| 748 | order = get_count_order_ulong(index + 1); |
| 749 | lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)]; |
| 750 | iter_prev = (struct cds_lfht_node *) lookup; |
| 751 | /* We can always skip the dummy node initially */ |
| 752 | iter = rcu_dereference(iter_prev->p.next); |
| 753 | assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); |
| 754 | for (;;) { |
| 755 | /* TODO: check if removed */ |
| 756 | if (unlikely(!clear_flag(iter))) |
| 757 | goto insert; |
| 758 | /* TODO: check if removed */ |
| 759 | if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) |
| 760 | goto insert; |
| 761 | next = rcu_dereference(clear_flag(iter)->p.next); |
| 762 | if (unlikely(is_removed(next))) |
| 763 | goto gc_node; |
| 764 | if (unique |
| 765 | && !is_dummy(next) |
| 766 | && !ht->compare_fct(node->key, node->key_len, |
| 767 | clear_flag(iter)->key, |
| 768 | clear_flag(iter)->key_len)) |
| 769 | return clear_flag(iter); |
| 770 | /* Only account for identical reverse hash once */ |
| 771 | if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash |
| 772 | && !is_dummy(next)) |
| 773 | check_resize(ht, size, ++chain_len); |
| 774 | iter_prev = clear_flag(iter); |
| 775 | iter = next; |
| 776 | } |
| 777 | insert: |
| 778 | assert(node != clear_flag(iter)); |
| 779 | assert(!is_removed(iter_prev)); |
| 780 | assert(!is_removed(iter)); |
| 781 | assert(iter_prev != node); |
| 782 | if (!dummy) |
| 783 | node->p.next = clear_flag(iter); |
| 784 | else |
| 785 | node->p.next = flag_dummy(clear_flag(iter)); |
| 786 | if (is_dummy(iter)) |
| 787 | new_node = flag_dummy(node); |
| 788 | else |
| 789 | new_node = node; |
| 790 | if (uatomic_cmpxchg(&iter_prev->p.next, iter, |
| 791 | new_node) != iter) |
| 792 | continue; /* retry */ |
| 793 | else |
| 794 | goto gc_end; |
| 795 | gc_node: |
| 796 | assert(!is_removed(iter)); |
| 797 | if (is_dummy(iter)) |
| 798 | new_next = flag_dummy(clear_flag(next)); |
| 799 | else |
| 800 | new_next = clear_flag(next); |
| 801 | (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); |
| 802 | /* retry */ |
| 803 | } |
| 804 | gc_end: |
| 805 | /* Garbage collect logically removed nodes in the bucket */ |
| 806 | index = hash & (size - 1); |
| 807 | order = get_count_order_ulong(index + 1); |
| 808 | lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))]; |
| 809 | dummy_node = (struct cds_lfht_node *) lookup; |
| 810 | _cds_lfht_gc_bucket(dummy_node, node); |
| 811 | return node; |
| 812 | } |
| 813 | |
| 814 | static |
| 815 | int _cds_lfht_remove(struct cds_lfht *ht, unsigned long size, |
| 816 | struct cds_lfht_node *node, |
| 817 | int dummy_removal) |
| 818 | { |
| 819 | struct cds_lfht_node *dummy, *next, *old; |
| 820 | struct _cds_lfht_node *lookup; |
| 821 | int flagged = 0; |
| 822 | unsigned long hash, index, order; |
| 823 | |
| 824 | /* logically delete the node */ |
| 825 | assert(!is_dummy(node)); |
| 826 | assert(!is_removed(node)); |
| 827 | old = rcu_dereference(node->p.next); |
| 828 | do { |
| 829 | next = old; |
| 830 | if (unlikely(is_removed(next))) |
| 831 | goto end; |
| 832 | if (dummy_removal) |
| 833 | assert(is_dummy(next)); |
| 834 | else |
| 835 | assert(!is_dummy(next)); |
| 836 | old = uatomic_cmpxchg(&node->p.next, next, |
| 837 | flag_removed(next)); |
| 838 | } while (old != next); |
| 839 | |
| 840 | /* We performed the (logical) deletion. */ |
| 841 | flagged = 1; |
| 842 | |
| 843 | /* |
| 844 | * Ensure that the node is not visible to readers anymore: lookup for |
| 845 | * the node, and remove it (along with any other logically removed node) |
| 846 | * if found. |
| 847 | */ |
| 848 | hash = bit_reverse_ulong(node->p.reverse_hash); |
| 849 | assert(size > 0); |
| 850 | index = hash & (size - 1); |
| 851 | order = get_count_order_ulong(index + 1); |
| 852 | lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))]; |
| 853 | dummy = (struct cds_lfht_node *) lookup; |
| 854 | _cds_lfht_gc_bucket(dummy, node); |
| 855 | end: |
| 856 | /* |
| 857 | * Only the flagging action indicated that we (and no other) |
| 858 | * removed the node from the hash. |
| 859 | */ |
| 860 | if (flagged) { |
| 861 | assert(is_removed(rcu_dereference(node->p.next))); |
| 862 | return 0; |
| 863 | } else |
| 864 | return -ENOENT; |
| 865 | } |
| 866 | |
| 867 | static |
| 868 | void init_table_hash(struct cds_lfht *ht, unsigned long i, |
| 869 | unsigned long len) |
| 870 | { |
| 871 | unsigned long j; |
| 872 | |
| 873 | for (j = 0; j < len; j++) { |
| 874 | struct cds_lfht_node *new_node = |
| 875 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; |
| 876 | |
| 877 | dbg_printf("init hash entry: i %lu j %lu hash %lu\n", |
| 878 | i, j, !i ? 0 : (1UL << (i - 1)) + j); |
| 879 | new_node->p.reverse_hash = |
| 880 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); |
| 881 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
| 882 | break; |
| 883 | } |
| 884 | } |
| 885 | |
| 886 | static |
| 887 | void init_table_link(struct cds_lfht *ht, unsigned long i, unsigned long len) |
| 888 | { |
| 889 | unsigned long j; |
| 890 | |
| 891 | ht->cds_lfht_rcu_thread_online(); |
| 892 | ht->cds_lfht_rcu_read_lock(); |
| 893 | for (j = 0; j < len; j++) { |
| 894 | struct cds_lfht_node *new_node = |
| 895 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; |
| 896 | |
| 897 | dbg_printf("init link: i %lu j %lu hash %lu\n", |
| 898 | i, j, !i ? 0 : (1UL << (i - 1)) + j); |
| 899 | (void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)), |
| 900 | new_node, 0, 1); |
| 901 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
| 902 | break; |
| 903 | } |
| 904 | ht->cds_lfht_rcu_read_unlock(); |
| 905 | ht->cds_lfht_rcu_thread_offline(); |
| 906 | } |
| 907 | |
| 908 | /* |
| 909 | * Holding RCU read lock to protect _cds_lfht_add against memory |
| 910 | * reclaim that could be performed by other call_rcu worker threads (ABA |
| 911 | * problem). |
| 912 | */ |
| 913 | static |
| 914 | void init_table(struct cds_lfht *ht, |
| 915 | unsigned long first_order, unsigned long len_order) |
| 916 | { |
| 917 | unsigned long i, end_order; |
| 918 | |
| 919 | dbg_printf("init table: first_order %lu end_order %lu\n", |
| 920 | first_order, first_order + len_order); |
| 921 | end_order = first_order + len_order; |
| 922 | for (i = first_order; i < end_order; i++) { |
| 923 | unsigned long len; |
| 924 | |
| 925 | len = !i ? 1 : 1UL << (i - 1); |
| 926 | dbg_printf("init order %lu len: %lu\n", i, len); |
| 927 | ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level) |
| 928 | + (len * sizeof(struct _cds_lfht_node))); |
| 929 | |
| 930 | /* Set all dummy nodes reverse hash values for a level */ |
| 931 | init_table_hash(ht, i, len); |
| 932 | |
| 933 | /* |
| 934 | * Link all dummy nodes into the table. Concurrent |
| 935 | * add/remove are helping us. |
| 936 | */ |
| 937 | init_table_link(ht, i, len); |
| 938 | |
| 939 | /* |
| 940 | * Update table size (after init for now, because no |
| 941 | * concurrent updater help (TODO)). |
| 942 | */ |
| 943 | cmm_smp_wmb(); /* populate data before RCU size */ |
| 944 | CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i)); |
| 945 | dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i)); |
| 946 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
| 947 | break; |
| 948 | } |
| 949 | } |
| 950 | |
| 951 | static |
| 952 | void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len) |
| 953 | { |
| 954 | unsigned long j; |
| 955 | |
| 956 | ht->cds_lfht_rcu_thread_online(); |
| 957 | ht->cds_lfht_rcu_read_lock(); |
| 958 | for (j = 0; j < len; j++) { |
| 959 | struct cds_lfht_node *fini_node = |
| 960 | (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; |
| 961 | |
| 962 | dbg_printf("remove entry: i %lu j %lu hash %lu\n", |
| 963 | i, j, !i ? 0 : (1UL << (i - 1)) + j); |
| 964 | fini_node->p.reverse_hash = |
| 965 | bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); |
| 966 | (void) _cds_lfht_remove(ht, !i ? 0 : (1UL << (i - 1)), |
| 967 | fini_node, 1); |
| 968 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
| 969 | break; |
| 970 | } |
| 971 | ht->cds_lfht_rcu_read_unlock(); |
| 972 | ht->cds_lfht_rcu_thread_offline(); |
| 973 | } |
| 974 | |
| 975 | /* |
| 976 | * Holding RCU read lock to protect _cds_lfht_remove against memory |
| 977 | * reclaim that could be performed by other call_rcu worker threads (ABA |
| 978 | * problem). |
| 979 | */ |
| 980 | static |
| 981 | void fini_table(struct cds_lfht *ht, |
| 982 | unsigned long first_order, unsigned long len_order) |
| 983 | { |
| 984 | long i, end_order; |
| 985 | |
| 986 | dbg_printf("fini table: first_order %lu end_order %lu\n", |
| 987 | first_order, first_order + len_order); |
| 988 | end_order = first_order + len_order; |
| 989 | assert(first_order > 0); |
| 990 | assert(ht->t.size == (1UL << (first_order - 1))); |
| 991 | for (i = end_order - 1; i >= first_order; i--) { |
| 992 | unsigned long len; |
| 993 | |
| 994 | len = !i ? 1 : 1UL << (i - 1); |
| 995 | dbg_printf("fini order %lu len: %lu\n", i, len); |
| 996 | |
| 997 | /* |
| 998 | * Set "removed" flag in dummy nodes about to be removed. |
| 999 | * Unlink all now-logically-removed dummy node pointers. |
| 1000 | * Concurrent add/remove operation are helping us doing |
| 1001 | * the gc. |
| 1002 | */ |
| 1003 | remove_table(ht, i, len); |
| 1004 | |
| 1005 | ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level); |
| 1006 | |
| 1007 | dbg_printf("fini new size: %lu\n", 1UL << i); |
| 1008 | if (CMM_LOAD_SHARED(ht->in_progress_destroy)) |
| 1009 | break; |
| 1010 | } |
| 1011 | } |
| 1012 | |
| 1013 | struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct, |
| 1014 | cds_lfht_compare_fct compare_fct, |
| 1015 | unsigned long hash_seed, |
| 1016 | unsigned long init_size, |
| 1017 | int flags, |
| 1018 | void (*cds_lfht_call_rcu)(struct rcu_head *head, |
| 1019 | void (*func)(struct rcu_head *head)), |
| 1020 | void (*cds_lfht_synchronize_rcu)(void), |
| 1021 | void (*cds_lfht_rcu_read_lock)(void), |
| 1022 | void (*cds_lfht_rcu_read_unlock)(void), |
| 1023 | void (*cds_lfht_rcu_thread_offline)(void), |
| 1024 | void (*cds_lfht_rcu_thread_online)(void)) |
| 1025 | { |
| 1026 | struct cds_lfht *ht; |
| 1027 | unsigned long order; |
| 1028 | |
| 1029 | /* init_size must be power of two */ |
| 1030 | if (init_size && (init_size & (init_size - 1))) |
| 1031 | return NULL; |
| 1032 | ht = calloc(1, sizeof(struct cds_lfht)); |
| 1033 | ht->hash_fct = hash_fct; |
| 1034 | ht->compare_fct = compare_fct; |
| 1035 | ht->hash_seed = hash_seed; |
| 1036 | ht->cds_lfht_call_rcu = cds_lfht_call_rcu; |
| 1037 | ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu; |
| 1038 | ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock; |
| 1039 | ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock; |
| 1040 | ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline; |
| 1041 | ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online; |
| 1042 | ht->percpu_count = alloc_per_cpu_items_count(); |
| 1043 | /* this mutex should not nest in read-side C.S. */ |
| 1044 | pthread_mutex_init(&ht->resize_mutex, NULL); |
| 1045 | order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)) + 1; |
| 1046 | ht->flags = flags; |
| 1047 | ht->cds_lfht_rcu_thread_offline(); |
| 1048 | pthread_mutex_lock(&ht->resize_mutex); |
| 1049 | init_table(ht, 0, order); |
| 1050 | pthread_mutex_unlock(&ht->resize_mutex); |
| 1051 | ht->cds_lfht_rcu_thread_online(); |
| 1052 | return ht; |
| 1053 | } |
| 1054 | |
| 1055 | struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len) |
| 1056 | { |
| 1057 | struct cds_lfht_node *node, *next; |
| 1058 | struct _cds_lfht_node *lookup; |
| 1059 | unsigned long hash, reverse_hash, index, order, size; |
| 1060 | |
| 1061 | hash = ht->hash_fct(key, key_len, ht->hash_seed); |
| 1062 | reverse_hash = bit_reverse_ulong(hash); |
| 1063 | |
| 1064 | size = rcu_dereference(ht->t.size); |
| 1065 | index = hash & (size - 1); |
| 1066 | order = get_count_order_ulong(index + 1); |
| 1067 | lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)]; |
| 1068 | dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", |
| 1069 | hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1))); |
| 1070 | node = (struct cds_lfht_node *) lookup; |
| 1071 | for (;;) { |
| 1072 | if (unlikely(!node)) |
| 1073 | break; |
| 1074 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
| 1075 | node = NULL; |
| 1076 | break; |
| 1077 | } |
| 1078 | next = rcu_dereference(node->p.next); |
| 1079 | if (likely(!is_removed(next)) |
| 1080 | && !is_dummy(next) |
| 1081 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
| 1082 | break; |
| 1083 | } |
| 1084 | node = clear_flag(next); |
| 1085 | } |
| 1086 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); |
| 1087 | return node; |
| 1088 | } |
| 1089 | |
| 1090 | struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht, |
| 1091 | struct cds_lfht_node *node) |
| 1092 | { |
| 1093 | struct cds_lfht_node *next; |
| 1094 | unsigned long reverse_hash; |
| 1095 | void *key; |
| 1096 | size_t key_len; |
| 1097 | |
| 1098 | reverse_hash = node->p.reverse_hash; |
| 1099 | key = node->key; |
| 1100 | key_len = node->key_len; |
| 1101 | next = rcu_dereference(node->p.next); |
| 1102 | node = clear_flag(next); |
| 1103 | |
| 1104 | for (;;) { |
| 1105 | if (unlikely(!node)) |
| 1106 | break; |
| 1107 | if (unlikely(node->p.reverse_hash > reverse_hash)) { |
| 1108 | node = NULL; |
| 1109 | break; |
| 1110 | } |
| 1111 | next = rcu_dereference(node->p.next); |
| 1112 | if (likely(!is_removed(next)) |
| 1113 | && !is_dummy(next) |
| 1114 | && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { |
| 1115 | break; |
| 1116 | } |
| 1117 | node = clear_flag(next); |
| 1118 | } |
| 1119 | assert(!node || !is_dummy(rcu_dereference(node->p.next))); |
| 1120 | return node; |
| 1121 | } |
| 1122 | |
| 1123 | void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) |
| 1124 | { |
| 1125 | unsigned long hash, size; |
| 1126 | |
| 1127 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
| 1128 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
| 1129 | |
| 1130 | size = rcu_dereference(ht->t.size); |
| 1131 | (void) _cds_lfht_add(ht, size, node, 0, 0); |
| 1132 | ht_count_add(ht, size); |
| 1133 | } |
| 1134 | |
| 1135 | struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, |
| 1136 | struct cds_lfht_node *node) |
| 1137 | { |
| 1138 | unsigned long hash, size; |
| 1139 | struct cds_lfht_node *ret; |
| 1140 | |
| 1141 | hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); |
| 1142 | node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); |
| 1143 | |
| 1144 | size = rcu_dereference(ht->t.size); |
| 1145 | ret = _cds_lfht_add(ht, size, node, 1, 0); |
| 1146 | if (ret != node) |
| 1147 | ht_count_add(ht, size); |
| 1148 | return ret; |
| 1149 | } |
| 1150 | |
| 1151 | int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node) |
| 1152 | { |
| 1153 | unsigned long size; |
| 1154 | int ret; |
| 1155 | |
| 1156 | size = rcu_dereference(ht->t.size); |
| 1157 | ret = _cds_lfht_remove(ht, size, node, 0); |
| 1158 | if (!ret) |
| 1159 | ht_count_remove(ht, size); |
| 1160 | return ret; |
| 1161 | } |
| 1162 | |
| 1163 | static |
| 1164 | int cds_lfht_delete_dummy(struct cds_lfht *ht) |
| 1165 | { |
| 1166 | struct cds_lfht_node *node; |
| 1167 | struct _cds_lfht_node *lookup; |
| 1168 | unsigned long order, i, size; |
| 1169 | |
| 1170 | /* Check that the table is empty */ |
| 1171 | lookup = &ht->t.tbl[0]->nodes[0]; |
| 1172 | node = (struct cds_lfht_node *) lookup; |
| 1173 | do { |
| 1174 | node = clear_flag(node)->p.next; |
| 1175 | if (!is_dummy(node)) |
| 1176 | return -EPERM; |
| 1177 | assert(!is_removed(node)); |
| 1178 | } while (clear_flag(node)); |
| 1179 | /* |
| 1180 | * size accessed without rcu_dereference because hash table is |
| 1181 | * being destroyed. |
| 1182 | */ |
| 1183 | size = ht->t.size; |
| 1184 | /* Internal sanity check: all nodes left should be dummy */ |
| 1185 | for (order = 0; order < get_count_order_ulong(size) + 1; order++) { |
| 1186 | unsigned long len; |
| 1187 | |
| 1188 | len = !order ? 1 : 1UL << (order - 1); |
| 1189 | for (i = 0; i < len; i++) { |
| 1190 | dbg_printf("delete order %lu i %lu hash %lu\n", |
| 1191 | order, i, |
| 1192 | bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash)); |
| 1193 | assert(is_dummy(ht->t.tbl[order]->nodes[i].next)); |
| 1194 | } |
| 1195 | poison_free(ht->t.tbl[order]); |
| 1196 | } |
| 1197 | return 0; |
| 1198 | } |
| 1199 | |
| 1200 | /* |
| 1201 | * Should only be called when no more concurrent readers nor writers can |
| 1202 | * possibly access the table. |
| 1203 | */ |
| 1204 | int cds_lfht_destroy(struct cds_lfht *ht) |
| 1205 | { |
| 1206 | int ret; |
| 1207 | |
| 1208 | /* Wait for in-flight resize operations to complete */ |
| 1209 | CMM_STORE_SHARED(ht->in_progress_destroy, 1); |
| 1210 | while (uatomic_read(&ht->in_progress_resize)) |
| 1211 | poll(NULL, 0, 100); /* wait for 100ms */ |
| 1212 | ret = cds_lfht_delete_dummy(ht); |
| 1213 | if (ret) |
| 1214 | return ret; |
| 1215 | free_per_cpu_items_count(ht->percpu_count); |
| 1216 | poison_free(ht); |
| 1217 | return ret; |
| 1218 | } |
| 1219 | |
| 1220 | void cds_lfht_count_nodes(struct cds_lfht *ht, |
| 1221 | unsigned long *count, |
| 1222 | unsigned long *removed) |
| 1223 | { |
| 1224 | struct cds_lfht_node *node, *next; |
| 1225 | struct _cds_lfht_node *lookup; |
| 1226 | unsigned long nr_dummy = 0; |
| 1227 | |
| 1228 | *count = 0; |
| 1229 | *removed = 0; |
| 1230 | |
| 1231 | /* Count non-dummy nodes in the table */ |
| 1232 | lookup = &ht->t.tbl[0]->nodes[0]; |
| 1233 | node = (struct cds_lfht_node *) lookup; |
| 1234 | do { |
| 1235 | next = rcu_dereference(node->p.next); |
| 1236 | if (is_removed(next)) { |
| 1237 | assert(!is_dummy(next)); |
| 1238 | (*removed)++; |
| 1239 | } else if (!is_dummy(next)) |
| 1240 | (*count)++; |
| 1241 | else |
| 1242 | (nr_dummy)++; |
| 1243 | node = clear_flag(next); |
| 1244 | } while (node); |
| 1245 | dbg_printf("number of dummy nodes: %lu\n", nr_dummy); |
| 1246 | } |
| 1247 | |
| 1248 | /* called with resize mutex held */ |
| 1249 | static |
| 1250 | void _do_cds_lfht_grow(struct cds_lfht *ht, |
| 1251 | unsigned long old_size, unsigned long new_size) |
| 1252 | { |
| 1253 | unsigned long old_order, new_order; |
| 1254 | |
| 1255 | old_order = get_count_order_ulong(old_size) + 1; |
| 1256 | new_order = get_count_order_ulong(new_size) + 1; |
| 1257 | printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
| 1258 | old_size, old_order, new_size, new_order); |
| 1259 | assert(new_size > old_size); |
| 1260 | init_table(ht, old_order, new_order - old_order); |
| 1261 | } |
| 1262 | |
| 1263 | /* called with resize mutex held */ |
| 1264 | static |
| 1265 | void _do_cds_lfht_shrink(struct cds_lfht *ht, |
| 1266 | unsigned long old_size, unsigned long new_size) |
| 1267 | { |
| 1268 | unsigned long old_order, new_order; |
| 1269 | |
| 1270 | new_size = max(new_size, MIN_TABLE_SIZE); |
| 1271 | old_order = get_count_order_ulong(old_size) + 1; |
| 1272 | new_order = get_count_order_ulong(new_size) + 1; |
| 1273 | printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", |
| 1274 | old_size, old_order, new_size, new_order); |
| 1275 | assert(new_size < old_size); |
| 1276 | |
| 1277 | cmm_smp_wmb(); /* populate data before RCU size */ |
| 1278 | CMM_STORE_SHARED(ht->t.size, new_size); |
| 1279 | |
| 1280 | /* |
| 1281 | * We need to wait for all add operations to reach Q.S. (and |
| 1282 | * thus use the new table for lookups) before we can start |
| 1283 | * releasing the old dummy nodes. Otherwise their lookup will |
| 1284 | * return a logically removed node as insert position. |
| 1285 | */ |
| 1286 | ht->cds_lfht_synchronize_rcu(); |
| 1287 | |
| 1288 | /* Remove and unlink all dummy nodes to remove. */ |
| 1289 | fini_table(ht, new_order, old_order - new_order); |
| 1290 | } |
| 1291 | |
| 1292 | |
| 1293 | /* called with resize mutex held */ |
| 1294 | static |
| 1295 | void _do_cds_lfht_resize(struct cds_lfht *ht) |
| 1296 | { |
| 1297 | unsigned long new_size, old_size; |
| 1298 | |
| 1299 | /* |
| 1300 | * Resize table, re-do if the target size has changed under us. |
| 1301 | */ |
| 1302 | do { |
| 1303 | ht->t.resize_initiated = 1; |
| 1304 | old_size = ht->t.size; |
| 1305 | new_size = CMM_LOAD_SHARED(ht->t.resize_target); |
| 1306 | if (old_size < new_size) |
| 1307 | _do_cds_lfht_grow(ht, old_size, new_size); |
| 1308 | else if (old_size > new_size) |
| 1309 | _do_cds_lfht_shrink(ht, old_size, new_size); |
| 1310 | ht->t.resize_initiated = 0; |
| 1311 | /* write resize_initiated before read resize_target */ |
| 1312 | cmm_smp_mb(); |
| 1313 | } while (new_size != CMM_LOAD_SHARED(ht->t.resize_target)); |
| 1314 | } |
| 1315 | |
| 1316 | static |
| 1317 | unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size, |
| 1318 | int growth_order) |
| 1319 | { |
| 1320 | return _uatomic_max(&ht->t.resize_target, |
| 1321 | size << growth_order); |
| 1322 | } |
| 1323 | |
| 1324 | static |
| 1325 | void resize_target_update_count(struct cds_lfht *ht, |
| 1326 | unsigned long count) |
| 1327 | { |
| 1328 | count = max(count, MIN_TABLE_SIZE); |
| 1329 | uatomic_set(&ht->t.resize_target, count); |
| 1330 | } |
| 1331 | |
| 1332 | void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) |
| 1333 | { |
| 1334 | resize_target_update_count(ht, new_size); |
| 1335 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
| 1336 | ht->cds_lfht_rcu_thread_offline(); |
| 1337 | pthread_mutex_lock(&ht->resize_mutex); |
| 1338 | _do_cds_lfht_resize(ht); |
| 1339 | pthread_mutex_unlock(&ht->resize_mutex); |
| 1340 | ht->cds_lfht_rcu_thread_online(); |
| 1341 | } |
| 1342 | |
| 1343 | static |
| 1344 | void do_resize_cb(struct rcu_head *head) |
| 1345 | { |
| 1346 | struct rcu_resize_work *work = |
| 1347 | caa_container_of(head, struct rcu_resize_work, head); |
| 1348 | struct cds_lfht *ht = work->ht; |
| 1349 | |
| 1350 | ht->cds_lfht_rcu_thread_offline(); |
| 1351 | pthread_mutex_lock(&ht->resize_mutex); |
| 1352 | _do_cds_lfht_resize(ht); |
| 1353 | pthread_mutex_unlock(&ht->resize_mutex); |
| 1354 | ht->cds_lfht_rcu_thread_online(); |
| 1355 | poison_free(work); |
| 1356 | cmm_smp_mb(); /* finish resize before decrement */ |
| 1357 | uatomic_dec(&ht->in_progress_resize); |
| 1358 | } |
| 1359 | |
| 1360 | static |
| 1361 | void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth) |
| 1362 | { |
| 1363 | struct rcu_resize_work *work; |
| 1364 | unsigned long target_size; |
| 1365 | |
| 1366 | target_size = resize_target_update(ht, size, growth); |
| 1367 | /* Store resize_target before read resize_initiated */ |
| 1368 | cmm_smp_mb(); |
| 1369 | if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) { |
| 1370 | uatomic_inc(&ht->in_progress_resize); |
| 1371 | cmm_smp_mb(); /* increment resize count before calling it */ |
| 1372 | work = malloc(sizeof(*work)); |
| 1373 | work->ht = ht; |
| 1374 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); |
| 1375 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
| 1376 | } |
| 1377 | } |
| 1378 | |
| 1379 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
| 1380 | |
| 1381 | static |
| 1382 | void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, |
| 1383 | unsigned long count) |
| 1384 | { |
| 1385 | struct rcu_resize_work *work; |
| 1386 | |
| 1387 | if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) |
| 1388 | return; |
| 1389 | resize_target_update_count(ht, count); |
| 1390 | /* Store resize_target before read resize_initiated */ |
| 1391 | cmm_smp_mb(); |
| 1392 | if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) { |
| 1393 | uatomic_inc(&ht->in_progress_resize); |
| 1394 | cmm_smp_mb(); /* increment resize count before calling it */ |
| 1395 | work = malloc(sizeof(*work)); |
| 1396 | work->ht = ht; |
| 1397 | ht->cds_lfht_call_rcu(&work->head, do_resize_cb); |
| 1398 | CMM_STORE_SHARED(ht->t.resize_initiated, 1); |
| 1399 | } |
| 1400 | } |
| 1401 | |
| 1402 | #endif |