lib/librte_hash/rte_cuckoo_hash.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  * Copyright(c) 2018 Arm Limited
   4  */
   5
   6 #include <string.h>
   7 #include <stdint.h>
   8 #include <errno.h>
   9 #include <stdio.h>
  10 #include <stdarg.h>
  11 #include <sys/queue.h>
  12
  13 #include <rte_common.h>
  14 #include <rte_memory.h>         /* for definition of RTE_CACHE_LINE_SIZE */
  15 #include <rte_log.h>
  16 #include <rte_prefetch.h>
  17 #include <rte_branch_prediction.h>
  18 #include <rte_malloc.h>
  19 #include <rte_eal.h>
  20 #include <rte_eal_memconfig.h>
  21 #include <rte_per_lcore.h>
  22 #include <rte_errno.h>
  23 #include <rte_string_fns.h>
  24 #include <rte_cpuflags.h>
  25 #include <rte_rwlock.h>
  26 #include <rte_spinlock.h>
  27 #include <rte_ring.h>
  28 #include <rte_compat.h>
  29 #include <rte_vect.h>
  30 #include <rte_tailq.h>
  31
  32 #include "rte_hash.h"
  33 #include "rte_cuckoo_hash.h"
  34
  35 #define FOR_EACH_BUCKET(CURRENT_BKT, START_BUCKET)                            \
  36         for (CURRENT_BKT = START_BUCKET;                                      \
  37                 CURRENT_BKT != NULL;                                          \
  38                 CURRENT_BKT = CURRENT_BKT->next)
  39
  40 TAILQ_HEAD(rte_hash_list, rte_tailq_entry);
  41
  42 static struct rte_tailq_elem rte_hash_tailq = {
  43         .name = "RTE_HASH",
  44 };
  45 EAL_REGISTER_TAILQ(rte_hash_tailq)
  46
  47 struct rte_hash *
  48 rte_hash_find_existing(const char *name)
  49 {
  50         struct rte_hash *h = NULL;
  51         struct rte_tailq_entry *te;
  52         struct rte_hash_list *hash_list;
  53
  54         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
  55
  56         rte_mcfg_tailq_read_lock();
  57         TAILQ_FOREACH(te, hash_list, next) {
  58                 h = (struct rte_hash *) te->data;
  59                 if (strncmp(name, h->name, RTE_HASH_NAMESIZE) == 0)
  60                         break;
  61         }
  62         rte_mcfg_tailq_read_unlock();
  63
  64         if (te == NULL) {
  65                 rte_errno = ENOENT;
  66                 return NULL;
  67         }
  68         return h;
  69 }
  70
  71 static inline struct rte_hash_bucket *
  72 rte_hash_get_last_bkt(struct rte_hash_bucket *lst_bkt)
  73 {
  74         while (lst_bkt->next != NULL)
  75                 lst_bkt = lst_bkt->next;
  76         return lst_bkt;
  77 }
  78
  79 void rte_hash_set_cmp_func(struct rte_hash *h, rte_hash_cmp_eq_t func)
  80 {
  81         h->cmp_jump_table_idx = KEY_CUSTOM;
  82         h->rte_hash_custom_cmp_eq = func;
  83 }
  84
  85 static inline int
  86 rte_hash_cmp_eq(const void *key1, const void *key2, const struct rte_hash *h)
  87 {
  88         if (h->cmp_jump_table_idx == KEY_CUSTOM)
  89                 return h->rte_hash_custom_cmp_eq(key1, key2, h->key_len);
  90         else
  91                 return cmp_jump_table[h->cmp_jump_table_idx](key1, key2, h->key_len);
  92 }
  93
  94 /*
  95  * We use higher 16 bits of hash as the signature value stored in table.
  96  * We use the lower bits for the primary bucket
  97  * location. Then we XOR primary bucket location and the signature
  98  * to get the secondary bucket location. This is same as
  99  * proposed in Bin Fan, et al's paper
 100  * "MemC3: Compact and Concurrent MemCache with Dumber Caching and
 101  * Smarter Hashing". The benefit to use
 102  * XOR is that one could derive the alternative bucket location
 103  * by only using the current bucket location and the signature.
 104  */
 105 static inline uint16_t
 106 get_short_sig(const hash_sig_t hash)
 107 {
 108         return hash >> 16;
 109 }
 110
 111 static inline uint32_t
 112 get_prim_bucket_index(const struct rte_hash *h, const hash_sig_t hash)
 113 {
 114         return hash & h->bucket_bitmask;
 115 }
 116
 117 static inline uint32_t
 118 get_alt_bucket_index(const struct rte_hash *h,
 119                         uint32_t cur_bkt_idx, uint16_t sig)
 120 {
 121         return (cur_bkt_idx ^ sig) & h->bucket_bitmask;
 122 }
 123
 124 struct rte_hash *
 125 rte_hash_create(const struct rte_hash_parameters *params)
 126 {
 127         struct rte_hash *h = NULL;
 128         struct rte_tailq_entry *te = NULL;
 129         struct rte_hash_list *hash_list;
 130         struct rte_ring *r = NULL;
 131         struct rte_ring *r_ext = NULL;
 132         char hash_name[RTE_HASH_NAMESIZE];
 133         void *k = NULL;
 134         void *buckets = NULL;
 135         void *buckets_ext = NULL;
 136         char ring_name[RTE_RING_NAMESIZE];
 137         char ext_ring_name[RTE_RING_NAMESIZE];
 138         unsigned num_key_slots;
 139         unsigned i;
 140         unsigned int hw_trans_mem_support = 0, use_local_cache = 0;
 141         unsigned int ext_table_support = 0;
 142         unsigned int readwrite_concur_support = 0;
 143         unsigned int writer_takes_lock = 0;
 144         unsigned int no_free_on_del = 0;
 145         uint32_t *ext_bkt_to_free = NULL;
 146         uint32_t *tbl_chng_cnt = NULL;
 147         unsigned int readwrite_concur_lf_support = 0;
 148
 149         rte_hash_function default_hash_func = (rte_hash_function)rte_jhash;
 150
 151         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
 152
 153         if (params == NULL) {
 154                 RTE_LOG(ERR, HASH, "rte_hash_create has no parameters\n");
 155                 return NULL;
 156         }
 157
 158         /* Check for valid parameters */
 159         if ((params->entries > RTE_HASH_ENTRIES_MAX) ||
 160                         (params->entries < RTE_HASH_BUCKET_ENTRIES) ||
 161                         (params->key_len == 0)) {
 162                 rte_errno = EINVAL;
 163                 RTE_LOG(ERR, HASH, "rte_hash_create has invalid parameters\n");
 164                 return NULL;
 165         }
 166
 167         /* Validate correct usage of extra options */
 168         if ((params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY) &&
 169             (params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY_LF)) {
 170                 rte_errno = EINVAL;
 171                 RTE_LOG(ERR, HASH, "rte_hash_create: choose rw concurrency or "
 172                         "rw concurrency lock free\n");
 173                 return NULL;
 174         }
 175
 176         /* Check extra flags field to check extra options. */
 177         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_TRANS_MEM_SUPPORT)
 178                 hw_trans_mem_support = 1;
 179
 180         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) {
 181                 use_local_cache = 1;
 182                 writer_takes_lock = 1;
 183         }
 184
 185         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY) {
 186                 readwrite_concur_support = 1;
 187                 writer_takes_lock = 1;
 188         }
 189
 190         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_EXT_TABLE)
 191                 ext_table_support = 1;
 192
 193         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_NO_FREE_ON_DEL)
 194                 no_free_on_del = 1;
 195
 196         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY_LF) {
 197                 readwrite_concur_lf_support = 1;
 198                 /* Enable not freeing internal memory/index on delete */
 199                 no_free_on_del = 1;
 200         }
 201
 202         /* Store all keys and leave the first entry as a dummy entry for lookup_bulk */
 203         if (use_local_cache)
 204                 /*
 205                  * Increase number of slots by total number of indices
 206                  * that can be stored in the lcore caches
 207                  * except for the first cache
 208                  */
 209                 num_key_slots = params->entries + (RTE_MAX_LCORE - 1) *
 210                                         (LCORE_CACHE_SIZE - 1) + 1;
 211         else
 212                 num_key_slots = params->entries + 1;
 213
 214         snprintf(ring_name, sizeof(ring_name), "HT_%s", params->name);
 215         /* Create ring (Dummy slot index is not enqueued) */
 216         r = rte_ring_create(ring_name, rte_align32pow2(num_key_slots),
 217                         params->socket_id, 0);
 218         if (r == NULL) {
 219                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
 220                 goto err;
 221         }
 222
 223         const uint32_t num_buckets = rte_align32pow2(params->entries) /
 224                                                 RTE_HASH_BUCKET_ENTRIES;
 225
 226         /* Create ring for extendable buckets. */
 227         if (ext_table_support) {
 228                 snprintf(ext_ring_name, sizeof(ext_ring_name), "HT_EXT_%s",
 229                                                                 params->name);
 230                 r_ext = rte_ring_create(ext_ring_name,
 231                                 rte_align32pow2(num_buckets + 1),
 232                                 params->socket_id, 0);
 233
 234                 if (r_ext == NULL) {
 235                         RTE_LOG(ERR, HASH, "ext buckets memory allocation "
 236                                                                 "failed\n");
 237                         goto err;
 238                 }
 239         }
 240
 241         snprintf(hash_name, sizeof(hash_name), "HT_%s", params->name);
 242
 243         rte_mcfg_tailq_write_lock();
 244
 245         /* guarantee there's no existing: this is normally already checked
 246          * by ring creation above */
 247         TAILQ_FOREACH(te, hash_list, next) {
 248                 h = (struct rte_hash *) te->data;
 249                 if (strncmp(params->name, h->name, RTE_HASH_NAMESIZE) == 0)
 250                         break;
 251         }
 252         h = NULL;
 253         if (te != NULL) {
 254                 rte_errno = EEXIST;
 255                 te = NULL;
 256                 goto err_unlock;
 257         }
 258
 259         te = rte_zmalloc("HASH_TAILQ_ENTRY", sizeof(*te), 0);
 260         if (te == NULL) {
 261                 RTE_LOG(ERR, HASH, "tailq entry allocation failed\n");
 262                 goto err_unlock;
 263         }
 264
 265         h = (struct rte_hash *)rte_zmalloc_socket(hash_name, sizeof(struct rte_hash),
 266                                         RTE_CACHE_LINE_SIZE, params->socket_id);
 267
 268         if (h == NULL) {
 269                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
 270                 goto err_unlock;
 271         }
 272
 273         buckets = rte_zmalloc_socket(NULL,
 274                                 num_buckets * sizeof(struct rte_hash_bucket),
 275                                 RTE_CACHE_LINE_SIZE, params->socket_id);
 276
 277         if (buckets == NULL) {
 278                 RTE_LOG(ERR, HASH, "buckets memory allocation failed\n");
 279                 goto err_unlock;
 280         }
 281
 282         /* Allocate same number of extendable buckets */
 283         if (ext_table_support) {
 284                 buckets_ext = rte_zmalloc_socket(NULL,
 285                                 num_buckets * sizeof(struct rte_hash_bucket),
 286                                 RTE_CACHE_LINE_SIZE, params->socket_id);
 287                 if (buckets_ext == NULL) {
 288                         RTE_LOG(ERR, HASH, "ext buckets memory allocation "
 289                                                         "failed\n");
 290                         goto err_unlock;
 291                 }
 292                 /* Populate ext bkt ring. We reserve 0 similar to the
 293                  * key-data slot, just in case in future we want to
 294                  * use bucket index for the linked list and 0 means NULL
 295                  * for next bucket
 296                  */
 297                 for (i = 1; i <= num_buckets; i++)
 298                         rte_ring_sp_enqueue(r_ext, (void *)((uintptr_t) i));
 299
 300                 if (readwrite_concur_lf_support) {
 301                         ext_bkt_to_free = rte_zmalloc(NULL, sizeof(uint32_t) *
 302                                                                 num_key_slots, 0);
 303                         if (ext_bkt_to_free == NULL) {
 304                                 RTE_LOG(ERR, HASH, "ext bkt to free memory allocation "
 305                                                                 "failed\n");
 306                                 goto err_unlock;
 307                         }
 308                 }
 309         }
 310
 311         const uint32_t key_entry_size =
 312                 RTE_ALIGN(sizeof(struct rte_hash_key) + params->key_len,
 313                           KEY_ALIGNMENT);
 314         const uint64_t key_tbl_size = (uint64_t) key_entry_size * num_key_slots;
 315
 316         k = rte_zmalloc_socket(NULL, key_tbl_size,
 317                         RTE_CACHE_LINE_SIZE, params->socket_id);
 318
 319         if (k == NULL) {
 320                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
 321                 goto err_unlock;
 322         }
 323
 324         tbl_chng_cnt = rte_zmalloc_socket(NULL, sizeof(uint32_t),
 325                         RTE_CACHE_LINE_SIZE, params->socket_id);
 326
 327         if (tbl_chng_cnt == NULL) {
 328                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
 329                 goto err_unlock;
 330         }
 331
 332 /*
 333  * If x86 architecture is used, select appropriate compare function,
 334  * which may use x86 intrinsics, otherwise use memcmp
 335  */
 336 #if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
 337         /* Select function to compare keys */
 338         switch (params->key_len) {
 339         case 16:
 340                 h->cmp_jump_table_idx = KEY_16_BYTES;
 341                 break;
 342         case 32:
 343                 h->cmp_jump_table_idx = KEY_32_BYTES;
 344                 break;
 345         case 48:
 346                 h->cmp_jump_table_idx = KEY_48_BYTES;
 347                 break;
 348         case 64:
 349                 h->cmp_jump_table_idx = KEY_64_BYTES;
 350                 break;
 351         case 80:
 352                 h->cmp_jump_table_idx = KEY_80_BYTES;
 353                 break;
 354         case 96:
 355                 h->cmp_jump_table_idx = KEY_96_BYTES;
 356                 break;
 357         case 112:
 358                 h->cmp_jump_table_idx = KEY_112_BYTES;
 359                 break;
 360         case 128:
 361                 h->cmp_jump_table_idx = KEY_128_BYTES;
 362                 break;
 363         default:
 364                 /* If key is not multiple of 16, use generic memcmp */
 365                 h->cmp_jump_table_idx = KEY_OTHER_BYTES;
 366         }
 367 #else
 368         h->cmp_jump_table_idx = KEY_OTHER_BYTES;
 369 #endif
 370
 371         if (use_local_cache) {
 372                 h->local_free_slots = rte_zmalloc_socket(NULL,
 373                                 sizeof(struct lcore_cache) * RTE_MAX_LCORE,
 374                                 RTE_CACHE_LINE_SIZE, params->socket_id);
 375         }
 376
 377         /* Default hash function */
 378 #if defined(RTE_ARCH_X86)
 379         default_hash_func = (rte_hash_function)rte_hash_crc;
 380 #elif defined(RTE_ARCH_ARM64)
 381         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_CRC32))
 382                 default_hash_func = (rte_hash_function)rte_hash_crc;
 383 #endif
 384         /* Setup hash context */
 385         strlcpy(h->name, params->name, sizeof(h->name));
 386         h->entries = params->entries;
 387         h->key_len = params->key_len;
 388         h->key_entry_size = key_entry_size;
 389         h->hash_func_init_val = params->hash_func_init_val;
 390
 391         h->num_buckets = num_buckets;
 392         h->bucket_bitmask = h->num_buckets - 1;
 393         h->buckets = buckets;
 394         h->buckets_ext = buckets_ext;
 395         h->free_ext_bkts = r_ext;
 396         h->hash_func = (params->hash_func == NULL) ?
 397                 default_hash_func : params->hash_func;
 398         h->key_store = k;
 399         h->free_slots = r;
 400         h->ext_bkt_to_free = ext_bkt_to_free;
 401         h->tbl_chng_cnt = tbl_chng_cnt;
 402         *h->tbl_chng_cnt = 0;
 403         h->hw_trans_mem_support = hw_trans_mem_support;
 404         h->use_local_cache = use_local_cache;
 405         h->readwrite_concur_support = readwrite_concur_support;
 406         h->ext_table_support = ext_table_support;
 407         h->writer_takes_lock = writer_takes_lock;
 408         h->no_free_on_del = no_free_on_del;
 409         h->readwrite_concur_lf_support = readwrite_concur_lf_support;
 410
 411 #if defined(RTE_ARCH_X86)
 412         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE2))
 413                 h->sig_cmp_fn = RTE_HASH_COMPARE_SSE;
 414         else
 415 #elif defined(RTE_ARCH_ARM64)
 416         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_NEON))
 417                 h->sig_cmp_fn = RTE_HASH_COMPARE_NEON;
 418         else
 419 #endif
 420                 h->sig_cmp_fn = RTE_HASH_COMPARE_SCALAR;
 421
 422         /* Writer threads need to take the lock when:
 423          * 1) RTE_HASH_EXTRA_FLAGS_RW_CONCURRENCY is enabled OR
 424          * 2) RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD is enabled
 425          */
 426         if (h->writer_takes_lock) {
 427                 h->readwrite_lock = rte_malloc(NULL, sizeof(rte_rwlock_t),
 428                                                 RTE_CACHE_LINE_SIZE);
 429                 if (h->readwrite_lock == NULL)
 430                         goto err_unlock;
 431
 432                 rte_rwlock_init(h->readwrite_lock);
 433         }
 434
 435         /* Populate free slots ring. Entry zero is reserved for key misses. */
 436         for (i = 1; i < num_key_slots; i++)
 437                 rte_ring_sp_enqueue(r, (void *)((uintptr_t) i));
 438
 439         te->data = (void *) h;
 440         TAILQ_INSERT_TAIL(hash_list, te, next);
 441         rte_mcfg_tailq_write_unlock();
 442
 443         return h;
 444 err_unlock:
 445         rte_mcfg_tailq_write_unlock();
 446 err:
 447         rte_ring_free(r);
 448         rte_ring_free(r_ext);
 449         rte_free(te);
 450         rte_free(h);
 451         rte_free(buckets);
 452         rte_free(buckets_ext);
 453         rte_free(k);
 454         rte_free(tbl_chng_cnt);
 455         rte_free(ext_bkt_to_free);
 456         return NULL;
 457 }
 458
 459 void
 460 rte_hash_free(struct rte_hash *h)
 461 {
 462         struct rte_tailq_entry *te;
 463         struct rte_hash_list *hash_list;
 464
 465         if (h == NULL)
 466                 return;
 467
 468         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
 469
 470         rte_mcfg_tailq_write_lock();
 471
 472         /* find out tailq entry */
 473         TAILQ_FOREACH(te, hash_list, next) {
 474                 if (te->data == (void *) h)
 475                         break;
 476         }
 477
 478         if (te == NULL) {
 479                 rte_mcfg_tailq_write_unlock();
 480                 return;
 481         }
 482
 483         TAILQ_REMOVE(hash_list, te, next);
 484
 485         rte_mcfg_tailq_write_unlock();
 486
 487         if (h->use_local_cache)
 488                 rte_free(h->local_free_slots);
 489         if (h->writer_takes_lock)
 490                 rte_free(h->readwrite_lock);
 491         rte_ring_free(h->free_slots);
 492         rte_ring_free(h->free_ext_bkts);
 493         rte_free(h->key_store);
 494         rte_free(h->buckets);
 495         rte_free(h->buckets_ext);
 496         rte_free(h->tbl_chng_cnt);
 497         rte_free(h->ext_bkt_to_free);
 498         rte_free(h);
 499         rte_free(te);
 500 }
 501
 502 hash_sig_t
 503 rte_hash_hash(const struct rte_hash *h, const void *key)
 504 {
 505         /* calc hash result by key */
 506         return h->hash_func(key, h->key_len, h->hash_func_init_val);
 507 }
 508
 509 int32_t
 510 rte_hash_count(const struct rte_hash *h)
 511 {
 512         uint32_t tot_ring_cnt, cached_cnt = 0;
 513         uint32_t i, ret;
 514
 515         if (h == NULL)
 516                 return -EINVAL;
 517
 518         if (h->use_local_cache) {
 519                 tot_ring_cnt = h->entries + (RTE_MAX_LCORE - 1) *
 520                                         (LCORE_CACHE_SIZE - 1);
 521                 for (i = 0; i < RTE_MAX_LCORE; i++)
 522                         cached_cnt += h->local_free_slots[i].len;
 523
 524                 ret = tot_ring_cnt - rte_ring_count(h->free_slots) -
 525                                                                 cached_cnt;
 526         } else {
 527                 tot_ring_cnt = h->entries;
 528                 ret = tot_ring_cnt - rte_ring_count(h->free_slots);
 529         }
 530         return ret;
 531 }
 532
 533 /* Read write locks implemented using rte_rwlock */
 534 static inline void
 535 __hash_rw_writer_lock(const struct rte_hash *h)
 536 {
 537         if (h->writer_takes_lock && h->hw_trans_mem_support)
 538                 rte_rwlock_write_lock_tm(h->readwrite_lock);
 539         else if (h->writer_takes_lock)
 540                 rte_rwlock_write_lock(h->readwrite_lock);
 541 }
 542
 543 static inline void
 544 __hash_rw_reader_lock(const struct rte_hash *h)
 545 {
 546         if (h->readwrite_concur_support && h->hw_trans_mem_support)
 547                 rte_rwlock_read_lock_tm(h->readwrite_lock);
 548         else if (h->readwrite_concur_support)
 549                 rte_rwlock_read_lock(h->readwrite_lock);
 550 }
 551
 552 static inline void
 553 __hash_rw_writer_unlock(const struct rte_hash *h)
 554 {
 555         if (h->writer_takes_lock && h->hw_trans_mem_support)
 556                 rte_rwlock_write_unlock_tm(h->readwrite_lock);
 557         else if (h->writer_takes_lock)
 558                 rte_rwlock_write_unlock(h->readwrite_lock);
 559 }
 560
 561 static inline void
 562 __hash_rw_reader_unlock(const struct rte_hash *h)
 563 {
 564         if (h->readwrite_concur_support && h->hw_trans_mem_support)
 565                 rte_rwlock_read_unlock_tm(h->readwrite_lock);
 566         else if (h->readwrite_concur_support)
 567                 rte_rwlock_read_unlock(h->readwrite_lock);
 568 }
 569
 570 void
 571 rte_hash_reset(struct rte_hash *h)
 572 {
 573         void *ptr;
 574         uint32_t tot_ring_cnt, i;
 575
 576         if (h == NULL)
 577                 return;
 578
 579         __hash_rw_writer_lock(h);
 580         memset(h->buckets, 0, h->num_buckets * sizeof(struct rte_hash_bucket));
 581         memset(h->key_store, 0, h->key_entry_size * (h->entries + 1));
 582         *h->tbl_chng_cnt = 0;
 583
 584         /* clear the free ring */
 585         while (rte_ring_dequeue(h->free_slots, &ptr) == 0)
 586                 continue;
 587
 588         /* clear free extendable bucket ring and memory */
 589         if (h->ext_table_support) {
 590                 memset(h->buckets_ext, 0, h->num_buckets *
 591                                                 sizeof(struct rte_hash_bucket));
 592                 while (rte_ring_dequeue(h->free_ext_bkts, &ptr) == 0)
 593                         continue;
 594         }
 595
 596         /* Repopulate the free slots ring. Entry zero is reserved for key misses */
 597         if (h->use_local_cache)
 598                 tot_ring_cnt = h->entries + (RTE_MAX_LCORE - 1) *
 599                                         (LCORE_CACHE_SIZE - 1);
 600         else
 601                 tot_ring_cnt = h->entries;
 602
 603         for (i = 1; i < tot_ring_cnt + 1; i++)
 604                 rte_ring_sp_enqueue(h->free_slots, (void *)((uintptr_t) i));
 605
 606         /* Repopulate the free ext bkt ring. */
 607         if (h->ext_table_support) {
 608                 for (i = 1; i <= h->num_buckets; i++)
 609                         rte_ring_sp_enqueue(h->free_ext_bkts,
 610                                                 (void *)((uintptr_t) i));
 611         }
 612
 613         if (h->use_local_cache) {
 614                 /* Reset local caches per lcore */
 615                 for (i = 0; i < RTE_MAX_LCORE; i++)
 616                         h->local_free_slots[i].len = 0;
 617         }
 618         __hash_rw_writer_unlock(h);
 619 }
 620
 621 /*
 622  * Function called to enqueue back an index in the cache/ring,
 623  * as slot has not being used and it can be used in the
 624  * next addition attempt.
 625  */
 626 static inline void
 627 enqueue_slot_back(const struct rte_hash *h,
 628                 struct lcore_cache *cached_free_slots,
 629                 void *slot_id)
 630 {
 631         if (h->use_local_cache) {
 632                 cached_free_slots->objs[cached_free_slots->len] = slot_id;
 633                 cached_free_slots->len++;
 634         } else
 635                 rte_ring_sp_enqueue(h->free_slots, slot_id);
 636 }
 637
 638 /* Search a key from bucket and update its data.
 639  * Writer holds the lock before calling this.
 640  */
 641 static inline int32_t
 642 search_and_update(const struct rte_hash *h, void *data, const void *key,
 643         struct rte_hash_bucket *bkt, uint16_t sig)
 644 {
 645         int i;
 646         struct rte_hash_key *k, *keys = h->key_store;
 647
 648         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 649                 if (bkt->sig_current[i] == sig) {
 650                         k = (struct rte_hash_key *) ((char *)keys +
 651                                         bkt->key_idx[i] * h->key_entry_size);
 652                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 653                                 /* The store to application data at *data
 654                                  * should not leak after the store to pdata
 655                                  * in the key store. i.e. pdata is the guard
 656                                  * variable. Release the application data
 657                                  * to the readers.
 658                                  */
 659                                 __atomic_store_n(&k->pdata,
 660                                         data,
 661                                         __ATOMIC_RELEASE);
 662                                 /*
 663                                  * Return index where key is stored,
 664                                  * subtracting the first dummy index
 665                                  */
 666                                 return bkt->key_idx[i] - 1;
 667                         }
 668                 }
 669         }
 670         return -1;
 671 }
 672
 673 /* Only tries to insert at one bucket (@prim_bkt) without trying to push
 674  * buckets around.
 675  * return 1 if matching existing key, return 0 if succeeds, return -1 for no
 676  * empty entry.
 677  */
 678 static inline int32_t
 679 rte_hash_cuckoo_insert_mw(const struct rte_hash *h,
 680                 struct rte_hash_bucket *prim_bkt,
 681                 struct rte_hash_bucket *sec_bkt,
 682                 const struct rte_hash_key *key, void *data,
 683                 uint16_t sig, uint32_t new_idx,
 684                 int32_t *ret_val)
 685 {
 686         unsigned int i;
 687         struct rte_hash_bucket *cur_bkt;
 688         int32_t ret;
 689
 690         __hash_rw_writer_lock(h);
 691         /* Check if key was inserted after last check but before this
 692          * protected region in case of inserting duplicated keys.
 693          */
 694         ret = search_and_update(h, data, key, prim_bkt, sig);
 695         if (ret != -1) {
 696                 __hash_rw_writer_unlock(h);
 697                 *ret_val = ret;
 698                 return 1;
 699         }
 700
 701         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
 702                 ret = search_and_update(h, data, key, cur_bkt, sig);
 703                 if (ret != -1) {
 704                         __hash_rw_writer_unlock(h);
 705                         *ret_val = ret;
 706                         return 1;
 707                 }
 708         }
 709
 710         /* Insert new entry if there is room in the primary
 711          * bucket.
 712          */
 713         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 714                 /* Check if slot is available */
 715                 if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
 716                         prim_bkt->sig_current[i] = sig;
 717                         /* Store to signature and key should not
 718                          * leak after the store to key_idx. i.e.
 719                          * key_idx is the guard variable for signature
 720                          * and key.
 721                          */
 722                         __atomic_store_n(&prim_bkt->key_idx[i],
 723                                          new_idx,
 724                                          __ATOMIC_RELEASE);
 725                         break;
 726                 }
 727         }
 728         __hash_rw_writer_unlock(h);
 729
 730         if (i != RTE_HASH_BUCKET_ENTRIES)
 731                 return 0;
 732
 733         /* no empty entry */
 734         return -1;
 735 }
 736
 737 /* Shift buckets along provided cuckoo_path (@leaf and @leaf_slot) and fill
 738  * the path head with new entry (sig, alt_hash, new_idx)
 739  * return 1 if matched key found, return -1 if cuckoo path invalided and fail,
 740  * return 0 if succeeds.
 741  */
 742 static inline int
 743 rte_hash_cuckoo_move_insert_mw(const struct rte_hash *h,
 744                         struct rte_hash_bucket *bkt,
 745                         struct rte_hash_bucket *alt_bkt,
 746                         const struct rte_hash_key *key, void *data,
 747                         struct queue_node *leaf, uint32_t leaf_slot,
 748                         uint16_t sig, uint32_t new_idx,
 749                         int32_t *ret_val)
 750 {
 751         uint32_t prev_alt_bkt_idx;
 752         struct rte_hash_bucket *cur_bkt;
 753         struct queue_node *prev_node, *curr_node = leaf;
 754         struct rte_hash_bucket *prev_bkt, *curr_bkt = leaf->bkt;
 755         uint32_t prev_slot, curr_slot = leaf_slot;
 756         int32_t ret;
 757
 758         __hash_rw_writer_lock(h);
 759
 760         /* In case empty slot was gone before entering protected region */
 761         if (curr_bkt->key_idx[curr_slot] != EMPTY_SLOT) {
 762                 __hash_rw_writer_unlock(h);
 763                 return -1;
 764         }
 765
 766         /* Check if key was inserted after last check but before this
 767          * protected region.
 768          */
 769         ret = search_and_update(h, data, key, bkt, sig);
 770         if (ret != -1) {
 771                 __hash_rw_writer_unlock(h);
 772                 *ret_val = ret;
 773                 return 1;
 774         }
 775
 776         FOR_EACH_BUCKET(cur_bkt, alt_bkt) {
 777                 ret = search_and_update(h, data, key, cur_bkt, sig);
 778                 if (ret != -1) {
 779                         __hash_rw_writer_unlock(h);
 780                         *ret_val = ret;
 781                         return 1;
 782                 }
 783         }
 784
 785         while (likely(curr_node->prev != NULL)) {
 786                 prev_node = curr_node->prev;
 787                 prev_bkt = prev_node->bkt;
 788                 prev_slot = curr_node->prev_slot;
 789
 790                 prev_alt_bkt_idx = get_alt_bucket_index(h,
 791                                         prev_node->cur_bkt_idx,
 792                                         prev_bkt->sig_current[prev_slot]);
 793
 794                 if (unlikely(&h->buckets[prev_alt_bkt_idx]
 795                                 != curr_bkt)) {
 796                         /* revert it to empty, otherwise duplicated keys */
 797                         __atomic_store_n(&curr_bkt->key_idx[curr_slot],
 798                                 EMPTY_SLOT,
 799                                 __ATOMIC_RELEASE);
 800                         __hash_rw_writer_unlock(h);
 801                         return -1;
 802                 }
 803
 804                 if (h->readwrite_concur_lf_support) {
 805                         /* Inform the previous move. The current move need
 806                          * not be informed now as the current bucket entry
 807                          * is present in both primary and secondary.
 808                          * Since there is one writer, load acquires on
 809                          * tbl_chng_cnt are not required.
 810                          */
 811                         __atomic_store_n(h->tbl_chng_cnt,
 812                                          *h->tbl_chng_cnt + 1,
 813                                          __ATOMIC_RELEASE);
 814                         /* The store to sig_current should not
 815                          * move above the store to tbl_chng_cnt.
 816                          */
 817                         __atomic_thread_fence(__ATOMIC_RELEASE);
 818                 }
 819
 820                 /* Need to swap current/alt sig to allow later
 821                  * Cuckoo insert to move elements back to its
 822                  * primary bucket if available
 823                  */
 824                 curr_bkt->sig_current[curr_slot] =
 825                         prev_bkt->sig_current[prev_slot];
 826                 /* Release the updated bucket entry */
 827                 __atomic_store_n(&curr_bkt->key_idx[curr_slot],
 828                         prev_bkt->key_idx[prev_slot],
 829                         __ATOMIC_RELEASE);
 830
 831                 curr_slot = prev_slot;
 832                 curr_node = prev_node;
 833                 curr_bkt = curr_node->bkt;
 834         }
 835
 836         if (h->readwrite_concur_lf_support) {
 837                 /* Inform the previous move. The current move need
 838                  * not be informed now as the current bucket entry
 839                  * is present in both primary and secondary.
 840                  * Since there is one writer, load acquires on
 841                  * tbl_chng_cnt are not required.
 842                  */
 843                 __atomic_store_n(h->tbl_chng_cnt,
 844                                  *h->tbl_chng_cnt + 1,
 845                                  __ATOMIC_RELEASE);
 846                 /* The store to sig_current should not
 847                  * move above the store to tbl_chng_cnt.
 848                  */
 849                 __atomic_thread_fence(__ATOMIC_RELEASE);
 850         }
 851
 852         curr_bkt->sig_current[curr_slot] = sig;
 853         /* Release the new bucket entry */
 854         __atomic_store_n(&curr_bkt->key_idx[curr_slot],
 855                          new_idx,
 856                          __ATOMIC_RELEASE);
 857
 858         __hash_rw_writer_unlock(h);
 859
 860         return 0;
 861
 862 }
 863
 864 /*
 865  * Make space for new key, using bfs Cuckoo Search and Multi-Writer safe
 866  * Cuckoo
 867  */
 868 static inline int
 869 rte_hash_cuckoo_make_space_mw(const struct rte_hash *h,
 870                         struct rte_hash_bucket *bkt,
 871                         struct rte_hash_bucket *sec_bkt,
 872                         const struct rte_hash_key *key, void *data,
 873                         uint16_t sig, uint32_t bucket_idx,
 874                         uint32_t new_idx, int32_t *ret_val)
 875 {
 876         unsigned int i;
 877         struct queue_node queue[RTE_HASH_BFS_QUEUE_MAX_LEN];
 878         struct queue_node *tail, *head;
 879         struct rte_hash_bucket *curr_bkt, *alt_bkt;
 880         uint32_t cur_idx, alt_idx;
 881
 882         tail = queue;
 883         head = queue + 1;
 884         tail->bkt = bkt;
 885         tail->prev = NULL;
 886         tail->prev_slot = -1;
 887         tail->cur_bkt_idx = bucket_idx;
 888
 889         /* Cuckoo bfs Search */
 890         while (likely(tail != head && head <
 891                                         queue + RTE_HASH_BFS_QUEUE_MAX_LEN -
 892                                         RTE_HASH_BUCKET_ENTRIES)) {
 893                 curr_bkt = tail->bkt;
 894                 cur_idx = tail->cur_bkt_idx;
 895                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 896                         if (curr_bkt->key_idx[i] == EMPTY_SLOT) {
 897                                 int32_t ret = rte_hash_cuckoo_move_insert_mw(h,
 898                                                 bkt, sec_bkt, key, data,
 899                                                 tail, i, sig,
 900                                                 new_idx, ret_val);
 901                                 if (likely(ret != -1))
 902                                         return ret;
 903                         }
 904
 905                         /* Enqueue new node and keep prev node info */
 906                         alt_idx = get_alt_bucket_index(h, cur_idx,
 907                                                 curr_bkt->sig_current[i]);
 908                         alt_bkt = &(h->buckets[alt_idx]);
 909                         head->bkt = alt_bkt;
 910                         head->cur_bkt_idx = alt_idx;
 911                         head->prev = tail;
 912                         head->prev_slot = i;
 913                         head++;
 914                 }
 915                 tail++;
 916         }
 917
 918         return -ENOSPC;
 919 }
 920
 921 static inline int32_t
 922 __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
 923                                                 hash_sig_t sig, void *data)
 924 {
 925         uint16_t short_sig;
 926         uint32_t prim_bucket_idx, sec_bucket_idx;
 927         struct rte_hash_bucket *prim_bkt, *sec_bkt, *cur_bkt;
 928         struct rte_hash_key *new_k, *keys = h->key_store;
 929         void *slot_id = NULL;
 930         void *ext_bkt_id = NULL;
 931         uint32_t new_idx, bkt_id;
 932         int ret;
 933         unsigned n_slots;
 934         unsigned lcore_id;
 935         unsigned int i;
 936         struct lcore_cache *cached_free_slots = NULL;
 937         int32_t ret_val;
 938         struct rte_hash_bucket *last;
 939
 940         short_sig = get_short_sig(sig);
 941         prim_bucket_idx = get_prim_bucket_index(h, sig);
 942         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
 943         prim_bkt = &h->buckets[prim_bucket_idx];
 944         sec_bkt = &h->buckets[sec_bucket_idx];
 945         rte_prefetch0(prim_bkt);
 946         rte_prefetch0(sec_bkt);
 947
 948         /* Check if key is already inserted in primary location */
 949         __hash_rw_writer_lock(h);
 950         ret = search_and_update(h, data, key, prim_bkt, short_sig);
 951         if (ret != -1) {
 952                 __hash_rw_writer_unlock(h);
 953                 return ret;
 954         }
 955
 956         /* Check if key is already inserted in secondary location */
 957         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
 958                 ret = search_and_update(h, data, key, cur_bkt, short_sig);
 959                 if (ret != -1) {
 960                         __hash_rw_writer_unlock(h);
 961                         return ret;
 962                 }
 963         }
 964
 965         __hash_rw_writer_unlock(h);
 966
 967         /* Did not find a match, so get a new slot for storing the new key */
 968         if (h->use_local_cache) {
 969                 lcore_id = rte_lcore_id();
 970                 cached_free_slots = &h->local_free_slots[lcore_id];
 971                 /* Try to get a free slot from the local cache */
 972                 if (cached_free_slots->len == 0) {
 973                         /* Need to get another burst of free slots from global ring */
 974                         n_slots = rte_ring_mc_dequeue_burst(h->free_slots,
 975                                         cached_free_slots->objs,
 976                                         LCORE_CACHE_SIZE, NULL);
 977                         if (n_slots == 0) {
 978                                 return -ENOSPC;
 979                         }
 980
 981                         cached_free_slots->len += n_slots;
 982                 }
 983
 984                 /* Get a free slot from the local cache */
 985                 cached_free_slots->len--;
 986                 slot_id = cached_free_slots->objs[cached_free_slots->len];
 987         } else {
 988                 if (rte_ring_sc_dequeue(h->free_slots, &slot_id) != 0) {
 989                         return -ENOSPC;
 990                 }
 991         }
 992
 993         new_k = RTE_PTR_ADD(keys, (uintptr_t)slot_id * h->key_entry_size);
 994         new_idx = (uint32_t)((uintptr_t) slot_id);
 995         /* The store to application data (by the application) at *data should
 996          * not leak after the store of pdata in the key store. i.e. pdata is
 997          * the guard variable. Release the application data to the readers.
 998          */
 999         __atomic_store_n(&new_k->pdata,
1000                 data,
1001                 __ATOMIC_RELEASE);
1002         /* Copy key */
1003         memcpy(new_k->key, key, h->key_len);
1004
1005         /* Find an empty slot and insert */
1006         ret = rte_hash_cuckoo_insert_mw(h, prim_bkt, sec_bkt, key, data,
1007                                         short_sig, new_idx, &ret_val);
1008         if (ret == 0)
1009                 return new_idx - 1;
1010         else if (ret == 1) {
1011                 enqueue_slot_back(h, cached_free_slots, slot_id);
1012                 return ret_val;
1013         }
1014
1015         /* Primary bucket full, need to make space for new entry */
1016         ret = rte_hash_cuckoo_make_space_mw(h, prim_bkt, sec_bkt, key, data,
1017                                 short_sig, prim_bucket_idx, new_idx, &ret_val);
1018         if (ret == 0)
1019                 return new_idx - 1;
1020         else if (ret == 1) {
1021                 enqueue_slot_back(h, cached_free_slots, slot_id);
1022                 return ret_val;
1023         }
1024
1025         /* Also search secondary bucket to get better occupancy */
1026         ret = rte_hash_cuckoo_make_space_mw(h, sec_bkt, prim_bkt, key, data,
1027                                 short_sig, sec_bucket_idx, new_idx, &ret_val);
1028
1029         if (ret == 0)
1030                 return new_idx - 1;
1031         else if (ret == 1) {
1032                 enqueue_slot_back(h, cached_free_slots, slot_id);
1033                 return ret_val;
1034         }
1035
1036         /* if ext table not enabled, we failed the insertion */
1037         if (!h->ext_table_support) {
1038                 enqueue_slot_back(h, cached_free_slots, slot_id);
1039                 return ret;
1040         }
1041
1042         /* Now we need to go through the extendable bucket. Protection is needed
1043          * to protect all extendable bucket processes.
1044          */
1045         __hash_rw_writer_lock(h);
1046         /* We check for duplicates again since could be inserted before the lock */
1047         ret = search_and_update(h, data, key, prim_bkt, short_sig);
1048         if (ret != -1) {
1049                 enqueue_slot_back(h, cached_free_slots, slot_id);
1050                 goto failure;
1051         }
1052
1053         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1054                 ret = search_and_update(h, data, key, cur_bkt, short_sig);
1055                 if (ret != -1) {
1056                         enqueue_slot_back(h, cached_free_slots, slot_id);
1057                         goto failure;
1058                 }
1059         }
1060
1061         /* Search sec and ext buckets to find an empty entry to insert. */
1062         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1063                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1064                         /* Check if slot is available */
1065                         if (likely(cur_bkt->key_idx[i] == EMPTY_SLOT)) {
1066                                 cur_bkt->sig_current[i] = short_sig;
1067                                 /* Store to signature and key should not
1068                                  * leak after the store to key_idx. i.e.
1069                                  * key_idx is the guard variable for signature
1070                                  * and key.
1071                                  */
1072                                 __atomic_store_n(&cur_bkt->key_idx[i],
1073                                                  new_idx,
1074                                                  __ATOMIC_RELEASE);
1075                                 __hash_rw_writer_unlock(h);
1076                                 return new_idx - 1;
1077                         }
1078                 }
1079         }
1080
1081         /* Failed to get an empty entry from extendable buckets. Link a new
1082          * extendable bucket. We first get a free bucket from ring.
1083          */
1084         if (rte_ring_sc_dequeue(h->free_ext_bkts, &ext_bkt_id) != 0) {
1085                 ret = -ENOSPC;
1086                 goto failure;
1087         }
1088
1089         bkt_id = (uint32_t)((uintptr_t)ext_bkt_id) - 1;
1090         /* Use the first location of the new bucket */
1091         (h->buckets_ext[bkt_id]).sig_current[0] = short_sig;
1092         /* Store to signature and key should not leak after
1093          * the store to key_idx. i.e. key_idx is the guard variable
1094          * for signature and key.
1095          */
1096         __atomic_store_n(&(h->buckets_ext[bkt_id]).key_idx[0],
1097                          new_idx,
1098                          __ATOMIC_RELEASE);
1099         /* Link the new bucket to sec bucket linked list */
1100         last = rte_hash_get_last_bkt(sec_bkt);
1101         last->next = &h->buckets_ext[bkt_id];
1102         __hash_rw_writer_unlock(h);
1103         return new_idx - 1;
1104
1105 failure:
1106         __hash_rw_writer_unlock(h);
1107         return ret;
1108
1109 }
1110
1111 int32_t
1112 rte_hash_add_key_with_hash(const struct rte_hash *h,
1113                         const void *key, hash_sig_t sig)
1114 {
1115         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1116         return __rte_hash_add_key_with_hash(h, key, sig, 0);
1117 }
1118
1119 int32_t
1120 rte_hash_add_key(const struct rte_hash *h, const void *key)
1121 {
1122         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1123         return __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), 0);
1124 }
1125
1126 int
1127 rte_hash_add_key_with_hash_data(const struct rte_hash *h,
1128                         const void *key, hash_sig_t sig, void *data)
1129 {
1130         int ret;
1131
1132         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1133         ret = __rte_hash_add_key_with_hash(h, key, sig, data);
1134         if (ret >= 0)
1135                 return 0;
1136         else
1137                 return ret;
1138 }
1139
1140 int
1141 rte_hash_add_key_data(const struct rte_hash *h, const void *key, void *data)
1142 {
1143         int ret;
1144
1145         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1146
1147         ret = __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), data);
1148         if (ret >= 0)
1149                 return 0;
1150         else
1151                 return ret;
1152 }
1153
1154 /* Search one bucket to find the match key - uses rw lock */
1155 static inline int32_t
1156 search_one_bucket_l(const struct rte_hash *h, const void *key,
1157                 uint16_t sig, void **data,
1158                 const struct rte_hash_bucket *bkt)
1159 {
1160         int i;
1161         struct rte_hash_key *k, *keys = h->key_store;
1162
1163         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1164                 if (bkt->sig_current[i] == sig &&
1165                                 bkt->key_idx[i] != EMPTY_SLOT) {
1166                         k = (struct rte_hash_key *) ((char *)keys +
1167                                         bkt->key_idx[i] * h->key_entry_size);
1168
1169                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1170                                 if (data != NULL)
1171                                         *data = k->pdata;
1172                                 /*
1173                                  * Return index where key is stored,
1174                                  * subtracting the first dummy index
1175                                  */
1176                                 return bkt->key_idx[i] - 1;
1177                         }
1178                 }
1179         }
1180         return -1;
1181 }
1182
1183 /* Search one bucket to find the match key */
1184 static inline int32_t
1185 search_one_bucket_lf(const struct rte_hash *h, const void *key, uint16_t sig,
1186                         void **data, const struct rte_hash_bucket *bkt)
1187 {
1188         int i;
1189         uint32_t key_idx;
1190         struct rte_hash_key *k, *keys = h->key_store;
1191
1192         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1193                 /* Signature comparison is done before the acquire-load
1194                  * of the key index to achieve better performance.
1195                  * This can result in the reader loading old signature
1196                  * (which matches), while the key_idx is updated to a
1197                  * value that belongs to a new key. However, the full
1198                  * key comparison will ensure that the lookup fails.
1199                  */
1200                 if (bkt->sig_current[i] == sig) {
1201                         key_idx = __atomic_load_n(&bkt->key_idx[i],
1202                                           __ATOMIC_ACQUIRE);
1203                         if (key_idx != EMPTY_SLOT) {
1204                                 k = (struct rte_hash_key *) ((char *)keys +
1205                                                 key_idx * h->key_entry_size);
1206
1207                                 if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1208                                         if (data != NULL) {
1209                                                 *data = __atomic_load_n(
1210                                                         &k->pdata,
1211                                                         __ATOMIC_ACQUIRE);
1212                                         }
1213                                         /*
1214                                          * Return index where key is stored,
1215                                          * subtracting the first dummy index
1216                                          */
1217                                         return key_idx - 1;
1218                                 }
1219                         }
1220                 }
1221         }
1222         return -1;
1223 }
1224
1225 static inline int32_t
1226 __rte_hash_lookup_with_hash_l(const struct rte_hash *h, const void *key,
1227                                 hash_sig_t sig, void **data)
1228 {
1229         uint32_t prim_bucket_idx, sec_bucket_idx;
1230         struct rte_hash_bucket *bkt, *cur_bkt;
1231         int ret;
1232         uint16_t short_sig;
1233
1234         short_sig = get_short_sig(sig);
1235         prim_bucket_idx = get_prim_bucket_index(h, sig);
1236         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1237
1238         bkt = &h->buckets[prim_bucket_idx];
1239
1240         __hash_rw_reader_lock(h);
1241
1242         /* Check if key is in primary location */
1243         ret = search_one_bucket_l(h, key, short_sig, data, bkt);
1244         if (ret != -1) {
1245                 __hash_rw_reader_unlock(h);
1246                 return ret;
1247         }
1248         /* Calculate secondary hash */
1249         bkt = &h->buckets[sec_bucket_idx];
1250
1251         /* Check if key is in secondary location */
1252         FOR_EACH_BUCKET(cur_bkt, bkt) {
1253                 ret = search_one_bucket_l(h, key, short_sig,
1254                                         data, cur_bkt);
1255                 if (ret != -1) {
1256                         __hash_rw_reader_unlock(h);
1257                         return ret;
1258                 }
1259         }
1260
1261         __hash_rw_reader_unlock(h);
1262
1263         return -ENOENT;
1264 }
1265
1266 static inline int32_t
1267 __rte_hash_lookup_with_hash_lf(const struct rte_hash *h, const void *key,
1268                                         hash_sig_t sig, void **data)
1269 {
1270         uint32_t prim_bucket_idx, sec_bucket_idx;
1271         struct rte_hash_bucket *bkt, *cur_bkt;
1272         uint32_t cnt_b, cnt_a;
1273         int ret;
1274         uint16_t short_sig;
1275
1276         short_sig = get_short_sig(sig);
1277         prim_bucket_idx = get_prim_bucket_index(h, sig);
1278         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1279
1280         do {
1281                 /* Load the table change counter before the lookup
1282                  * starts. Acquire semantics will make sure that
1283                  * loads in search_one_bucket are not hoisted.
1284                  */
1285                 cnt_b = __atomic_load_n(h->tbl_chng_cnt,
1286                                 __ATOMIC_ACQUIRE);
1287
1288                 /* Check if key is in primary location */
1289                 bkt = &h->buckets[prim_bucket_idx];
1290                 ret = search_one_bucket_lf(h, key, short_sig, data, bkt);
1291                 if (ret != -1) {
1292                         __hash_rw_reader_unlock(h);
1293                         return ret;
1294                 }
1295                 /* Calculate secondary hash */
1296                 bkt = &h->buckets[sec_bucket_idx];
1297
1298                 /* Check if key is in secondary location */
1299                 FOR_EACH_BUCKET(cur_bkt, bkt) {
1300                         ret = search_one_bucket_lf(h, key, short_sig,
1301                                                 data, cur_bkt);
1302                         if (ret != -1) {
1303                                 __hash_rw_reader_unlock(h);
1304                                 return ret;
1305                         }
1306                 }
1307
1308                 /* The loads of sig_current in search_one_bucket
1309                  * should not move below the load from tbl_chng_cnt.
1310                  */
1311                 __atomic_thread_fence(__ATOMIC_ACQUIRE);
1312                 /* Re-read the table change counter to check if the
1313                  * table has changed during search. If yes, re-do
1314                  * the search.
1315                  * This load should not get hoisted. The load
1316                  * acquires on cnt_b, key index in primary bucket
1317                  * and key index in secondary bucket will make sure
1318                  * that it does not get hoisted.
1319                  */
1320                 cnt_a = __atomic_load_n(h->tbl_chng_cnt,
1321                                         __ATOMIC_ACQUIRE);
1322         } while (cnt_b != cnt_a);
1323
1324         return -ENOENT;
1325 }
1326
1327 static inline int32_t
1328 __rte_hash_lookup_with_hash(const struct rte_hash *h, const void *key,
1329                                         hash_sig_t sig, void **data)
1330 {
1331         if (h->readwrite_concur_lf_support)
1332                 return __rte_hash_lookup_with_hash_lf(h, key, sig, data);
1333         else
1334                 return __rte_hash_lookup_with_hash_l(h, key, sig, data);
1335 }
1336
1337 int32_t
1338 rte_hash_lookup_with_hash(const struct rte_hash *h,
1339                         const void *key, hash_sig_t sig)
1340 {
1341         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1342         return __rte_hash_lookup_with_hash(h, key, sig, NULL);
1343 }
1344
1345 int32_t
1346 rte_hash_lookup(const struct rte_hash *h, const void *key)
1347 {
1348         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1349         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), NULL);
1350 }
1351
1352 int
1353 rte_hash_lookup_with_hash_data(const struct rte_hash *h,
1354                         const void *key, hash_sig_t sig, void **data)
1355 {
1356         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1357         return __rte_hash_lookup_with_hash(h, key, sig, data);
1358 }
1359
1360 int
1361 rte_hash_lookup_data(const struct rte_hash *h, const void *key, void **data)
1362 {
1363         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1364         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), data);
1365 }
1366
1367 static inline void
1368 remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i)
1369 {
1370         unsigned lcore_id, n_slots;
1371         struct lcore_cache *cached_free_slots;
1372
1373         if (h->use_local_cache) {
1374                 lcore_id = rte_lcore_id();
1375                 cached_free_slots = &h->local_free_slots[lcore_id];
1376                 /* Cache full, need to free it. */
1377                 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
1378                         /* Need to enqueue the free slots in global ring. */
1379                         n_slots = rte_ring_mp_enqueue_burst(h->free_slots,
1380                                                 cached_free_slots->objs,
1381                                                 LCORE_CACHE_SIZE, NULL);
1382                         ERR_IF_TRUE((n_slots == 0),
1383                                 "%s: could not enqueue free slots in global ring\n",
1384                                 __func__);
1385                         cached_free_slots->len -= n_slots;
1386                 }
1387                 /* Put index of new free slot in cache. */
1388                 cached_free_slots->objs[cached_free_slots->len] =
1389                                 (void *)((uintptr_t)bkt->key_idx[i]);
1390                 cached_free_slots->len++;
1391         } else {
1392                 rte_ring_sp_enqueue(h->free_slots,
1393                                 (void *)((uintptr_t)bkt->key_idx[i]));
1394         }
1395 }
1396
1397 /* Compact the linked list by moving key from last entry in linked list to the
1398  * empty slot.
1399  */
1400 static inline void
1401 __rte_hash_compact_ll(const struct rte_hash *h,
1402                         struct rte_hash_bucket *cur_bkt, int pos) {
1403         int i;
1404         struct rte_hash_bucket *last_bkt;
1405
1406         if (!cur_bkt->next)
1407                 return;
1408
1409         last_bkt = rte_hash_get_last_bkt(cur_bkt);
1410
1411         for (i = RTE_HASH_BUCKET_ENTRIES - 1; i >= 0; i--) {
1412                 if (last_bkt->key_idx[i] != EMPTY_SLOT) {
1413                         cur_bkt->sig_current[pos] = last_bkt->sig_current[i];
1414                         __atomic_store_n(&cur_bkt->key_idx[pos],
1415                                          last_bkt->key_idx[i],
1416                                          __ATOMIC_RELEASE);
1417                         if (h->readwrite_concur_lf_support) {
1418                                 /* Inform the readers that the table has changed
1419                                  * Since there is one writer, load acquire on
1420                                  * tbl_chng_cnt is not required.
1421                                  */
1422                                 __atomic_store_n(h->tbl_chng_cnt,
1423                                          *h->tbl_chng_cnt + 1,
1424                                          __ATOMIC_RELEASE);
1425                                 /* The store to sig_current should
1426                                  * not move above the store to tbl_chng_cnt.
1427                                  */
1428                                 __atomic_thread_fence(__ATOMIC_RELEASE);
1429                         }
1430                         last_bkt->sig_current[i] = NULL_SIGNATURE;
1431                         __atomic_store_n(&last_bkt->key_idx[i],
1432                                          EMPTY_SLOT,
1433                                          __ATOMIC_RELEASE);
1434                         return;
1435                 }
1436         }
1437 }
1438
1439 /* Search one bucket and remove the matched key.
1440  * Writer is expected to hold the lock while calling this
1441  * function.
1442  */
1443 static inline int32_t
1444 search_and_remove(const struct rte_hash *h, const void *key,
1445                         struct rte_hash_bucket *bkt, uint16_t sig, int *pos)
1446 {
1447         struct rte_hash_key *k, *keys = h->key_store;
1448         unsigned int i;
1449         uint32_t key_idx;
1450
1451         /* Check if key is in bucket */
1452         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1453                 key_idx = __atomic_load_n(&bkt->key_idx[i],
1454                                           __ATOMIC_ACQUIRE);
1455                 if (bkt->sig_current[i] == sig && key_idx != EMPTY_SLOT) {
1456                         k = (struct rte_hash_key *) ((char *)keys +
1457                                         key_idx * h->key_entry_size);
1458                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1459                                 bkt->sig_current[i] = NULL_SIGNATURE;
1460                                 /* Free the key store index if
1461                                  * no_free_on_del is disabled.
1462                                  */
1463                                 if (!h->no_free_on_del)
1464                                         remove_entry(h, bkt, i);
1465
1466                                 __atomic_store_n(&bkt->key_idx[i],
1467                                                  EMPTY_SLOT,
1468                                                  __ATOMIC_RELEASE);
1469
1470                                 *pos = i;
1471                                 /*
1472                                  * Return index where key is stored,
1473                                  * subtracting the first dummy index
1474                                  */
1475                                 return key_idx - 1;
1476                         }
1477                 }
1478         }
1479         return -1;
1480 }
1481
1482 static inline int32_t
1483 __rte_hash_del_key_with_hash(const struct rte_hash *h, const void *key,
1484                                                 hash_sig_t sig)
1485 {
1486         uint32_t prim_bucket_idx, sec_bucket_idx;
1487         struct rte_hash_bucket *prim_bkt, *sec_bkt, *prev_bkt, *last_bkt;
1488         struct rte_hash_bucket *cur_bkt;
1489         int pos;
1490         int32_t ret, i;
1491         uint16_t short_sig;
1492
1493         short_sig = get_short_sig(sig);
1494         prim_bucket_idx = get_prim_bucket_index(h, sig);
1495         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1496         prim_bkt = &h->buckets[prim_bucket_idx];
1497
1498         __hash_rw_writer_lock(h);
1499         /* look for key in primary bucket */
1500         ret = search_and_remove(h, key, prim_bkt, short_sig, &pos);
1501         if (ret != -1) {
1502                 __rte_hash_compact_ll(h, prim_bkt, pos);
1503                 last_bkt = prim_bkt->next;
1504                 prev_bkt = prim_bkt;
1505                 goto return_bkt;
1506         }
1507
1508         /* Calculate secondary hash */
1509         sec_bkt = &h->buckets[sec_bucket_idx];
1510
1511         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1512                 ret = search_and_remove(h, key, cur_bkt, short_sig, &pos);
1513                 if (ret != -1) {
1514                         __rte_hash_compact_ll(h, cur_bkt, pos);
1515                         last_bkt = sec_bkt->next;
1516                         prev_bkt = sec_bkt;
1517                         goto return_bkt;
1518                 }
1519         }
1520
1521         __hash_rw_writer_unlock(h);
1522         return -ENOENT;
1523
1524 /* Search last bucket to see if empty to be recycled */
1525 return_bkt:
1526         if (!last_bkt) {
1527                 __hash_rw_writer_unlock(h);
1528                 return ret;
1529         }
1530         while (last_bkt->next) {
1531                 prev_bkt = last_bkt;
1532                 last_bkt = last_bkt->next;
1533         }
1534
1535         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1536                 if (last_bkt->key_idx[i] != EMPTY_SLOT)
1537                         break;
1538         }
1539         /* found empty bucket and recycle */
1540         if (i == RTE_HASH_BUCKET_ENTRIES) {
1541                 prev_bkt->next = NULL;
1542                 uint32_t index = last_bkt - h->buckets_ext + 1;
1543                 /* Recycle the empty bkt if
1544                  * no_free_on_del is disabled.
1545                  */
1546                 if (h->no_free_on_del)
1547                         /* Store index of an empty ext bkt to be recycled
1548                          * on calling rte_hash_del_xxx APIs.
1549                          * When lock free read-write concurrency is enabled,
1550                          * an empty ext bkt cannot be put into free list
1551                          * immediately (as readers might be using it still).
1552                          * Hence freeing of the ext bkt is piggy-backed to
1553                          * freeing of the key index.
1554                          */
1555                         h->ext_bkt_to_free[ret] = index;
1556                 else
1557                         rte_ring_sp_enqueue(h->free_ext_bkts, (void *)(uintptr_t)index);
1558         }
1559         __hash_rw_writer_unlock(h);
1560         return ret;
1561 }
1562
1563 int32_t
1564 rte_hash_del_key_with_hash(const struct rte_hash *h,
1565                         const void *key, hash_sig_t sig)
1566 {
1567         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1568         return __rte_hash_del_key_with_hash(h, key, sig);
1569 }
1570
1571 int32_t
1572 rte_hash_del_key(const struct rte_hash *h, const void *key)
1573 {
1574         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1575         return __rte_hash_del_key_with_hash(h, key, rte_hash_hash(h, key));
1576 }
1577
1578 int
1579 rte_hash_get_key_with_position(const struct rte_hash *h, const int32_t position,
1580                                void **key)
1581 {
1582         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1583
1584         struct rte_hash_key *k, *keys = h->key_store;
1585         k = (struct rte_hash_key *) ((char *) keys + (position + 1) *
1586                                      h->key_entry_size);
1587         *key = k->key;
1588
1589         if (position !=
1590             __rte_hash_lookup_with_hash(h, *key, rte_hash_hash(h, *key),
1591                                         NULL)) {
1592                 return -ENOENT;
1593         }
1594
1595         return 0;
1596 }
1597
1598 int
1599 rte_hash_free_key_with_position(const struct rte_hash *h,
1600                                 const int32_t position)
1601 {
1602         /* Key index where key is stored, adding the first dummy index */
1603         uint32_t key_idx = position + 1;
1604
1605         RETURN_IF_TRUE(((h == NULL) || (key_idx == EMPTY_SLOT)), -EINVAL);
1606
1607         unsigned int lcore_id, n_slots;
1608         struct lcore_cache *cached_free_slots;
1609         const uint32_t total_entries = h->use_local_cache ?
1610                 h->entries + (RTE_MAX_LCORE - 1) * (LCORE_CACHE_SIZE - 1) + 1
1611                                                         : h->entries + 1;
1612
1613         /* Out of bounds */
1614         if (key_idx >= total_entries)
1615                 return -EINVAL;
1616         if (h->ext_table_support && h->readwrite_concur_lf_support) {
1617                 uint32_t index = h->ext_bkt_to_free[position];
1618                 if (index) {
1619                         /* Recycle empty ext bkt to free list. */
1620                         rte_ring_sp_enqueue(h->free_ext_bkts, (void *)(uintptr_t)index);
1621                         h->ext_bkt_to_free[position] = 0;
1622                 }
1623         }
1624
1625         if (h->use_local_cache) {
1626                 lcore_id = rte_lcore_id();
1627                 cached_free_slots = &h->local_free_slots[lcore_id];
1628                 /* Cache full, need to free it. */
1629                 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
1630                         /* Need to enqueue the free slots in global ring. */
1631                         n_slots = rte_ring_mp_enqueue_burst(h->free_slots,
1632                                                 cached_free_slots->objs,
1633                                                 LCORE_CACHE_SIZE, NULL);
1634                         RETURN_IF_TRUE((n_slots == 0), -EFAULT);
1635                         cached_free_slots->len -= n_slots;
1636                 }
1637                 /* Put index of new free slot in cache. */
1638                 cached_free_slots->objs[cached_free_slots->len] =
1639                                         (void *)((uintptr_t)key_idx);
1640                 cached_free_slots->len++;
1641         } else {
1642                 rte_ring_sp_enqueue(h->free_slots,
1643                                 (void *)((uintptr_t)key_idx));
1644         }
1645
1646         return 0;
1647 }
1648
1649 static inline void
1650 compare_signatures(uint32_t *prim_hash_matches, uint32_t *sec_hash_matches,
1651                         const struct rte_hash_bucket *prim_bkt,
1652                         const struct rte_hash_bucket *sec_bkt,
1653                         uint16_t sig,
1654                         enum rte_hash_sig_compare_function sig_cmp_fn)
1655 {
1656         unsigned int i;
1657
1658         /* For match mask the first bit of every two bits indicates the match */
1659         switch (sig_cmp_fn) {
1660 #if defined(RTE_MACHINE_CPUFLAG_SSE2)
1661         case RTE_HASH_COMPARE_SSE:
1662                 /* Compare all signatures in the bucket */
1663                 *prim_hash_matches = _mm_movemask_epi8(_mm_cmpeq_epi16(
1664                                 _mm_load_si128(
1665                                         (__m128i const *)prim_bkt->sig_current),
1666                                 _mm_set1_epi16(sig)));
1667                 /* Compare all signatures in the bucket */
1668                 *sec_hash_matches = _mm_movemask_epi8(_mm_cmpeq_epi16(
1669                                 _mm_load_si128(
1670                                         (__m128i const *)sec_bkt->sig_current),
1671                                 _mm_set1_epi16(sig)));
1672                 break;
1673 #elif defined(RTE_MACHINE_CPUFLAG_NEON)
1674         case RTE_HASH_COMPARE_NEON: {
1675                 uint16x8_t vmat, vsig, x;
1676                 int16x8_t shift = {-15, -13, -11, -9, -7, -5, -3, -1};
1677
1678                 vsig = vld1q_dup_u16((uint16_t const *)&sig);
1679                 /* Compare all signatures in the primary bucket */
1680                 vmat = vceqq_u16(vsig,
1681                         vld1q_u16((uint16_t const *)prim_bkt->sig_current));
1682                 x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
1683                 *prim_hash_matches = (uint32_t)(vaddvq_u16(x));
1684                 /* Compare all signatures in the secondary bucket */
1685                 vmat = vceqq_u16(vsig,
1686                         vld1q_u16((uint16_t const *)sec_bkt->sig_current));
1687                 x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
1688                 *sec_hash_matches = (uint32_t)(vaddvq_u16(x));
1689                 }
1690                 break;
1691 #endif
1692         default:
1693                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1694                         *prim_hash_matches |=
1695                                 ((sig == prim_bkt->sig_current[i]) << (i << 1));
1696                         *sec_hash_matches |=
1697                                 ((sig == sec_bkt->sig_current[i]) << (i << 1));
1698                 }
1699         }
1700 }
1701
1702 #define PREFETCH_OFFSET 4
1703 static inline void
1704 __rte_hash_lookup_bulk_l(const struct rte_hash *h, const void **keys,
1705                         int32_t num_keys, int32_t *positions,
1706                         uint64_t *hit_mask, void *data[])
1707 {
1708         uint64_t hits = 0;
1709         int32_t i;
1710         int32_t ret;
1711         uint32_t prim_hash[RTE_HASH_LOOKUP_BULK_MAX];
1712         uint32_t prim_index[RTE_HASH_LOOKUP_BULK_MAX];
1713         uint32_t sec_index[RTE_HASH_LOOKUP_BULK_MAX];
1714         uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
1715         const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1716         const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1717         uint32_t prim_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1718         uint32_t sec_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1719         struct rte_hash_bucket *cur_bkt, *next_bkt;
1720
1721         /* Prefetch first keys */
1722         for (i = 0; i < PREFETCH_OFFSET && i < num_keys; i++)
1723                 rte_prefetch0(keys[i]);
1724
1725         /*
1726          * Prefetch rest of the keys, calculate primary and
1727          * secondary bucket and prefetch them
1728          */
1729         for (i = 0; i < (num_keys - PREFETCH_OFFSET); i++) {
1730                 rte_prefetch0(keys[i + PREFETCH_OFFSET]);
1731
1732                 prim_hash[i] = rte_hash_hash(h, keys[i]);
1733
1734                 sig[i] = get_short_sig(prim_hash[i]);
1735                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1736                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1737
1738                 primary_bkt[i] = &h->buckets[prim_index[i]];
1739                 secondary_bkt[i] = &h->buckets[sec_index[i]];
1740
1741                 rte_prefetch0(primary_bkt[i]);
1742                 rte_prefetch0(secondary_bkt[i]);
1743         }
1744
1745         /* Calculate and prefetch rest of the buckets */
1746         for (; i < num_keys; i++) {
1747                 prim_hash[i] = rte_hash_hash(h, keys[i]);
1748
1749                 sig[i] = get_short_sig(prim_hash[i]);
1750                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1751                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1752
1753                 primary_bkt[i] = &h->buckets[prim_index[i]];
1754                 secondary_bkt[i] = &h->buckets[sec_index[i]];
1755
1756                 rte_prefetch0(primary_bkt[i]);
1757                 rte_prefetch0(secondary_bkt[i]);
1758         }
1759
1760         __hash_rw_reader_lock(h);
1761
1762         /* Compare signatures and prefetch key slot of first hit */
1763         for (i = 0; i < num_keys; i++) {
1764                 compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
1765                         primary_bkt[i], secondary_bkt[i],
1766                         sig[i], h->sig_cmp_fn);
1767
1768                 if (prim_hitmask[i]) {
1769                         uint32_t first_hit =
1770                                         __builtin_ctzl(prim_hitmask[i])
1771                                         >> 1;
1772                         uint32_t key_idx =
1773                                 primary_bkt[i]->key_idx[first_hit];
1774                         const struct rte_hash_key *key_slot =
1775                                 (const struct rte_hash_key *)(
1776                                 (const char *)h->key_store +
1777                                 key_idx * h->key_entry_size);
1778                         rte_prefetch0(key_slot);
1779                         continue;
1780                 }
1781
1782                 if (sec_hitmask[i]) {
1783                         uint32_t first_hit =
1784                                         __builtin_ctzl(sec_hitmask[i])
1785                                         >> 1;
1786                         uint32_t key_idx =
1787                                 secondary_bkt[i]->key_idx[first_hit];
1788                         const struct rte_hash_key *key_slot =
1789                                 (const struct rte_hash_key *)(
1790                                 (const char *)h->key_store +
1791                                 key_idx * h->key_entry_size);
1792                         rte_prefetch0(key_slot);
1793                 }
1794         }
1795
1796         /* Compare keys, first hits in primary first */
1797         for (i = 0; i < num_keys; i++) {
1798                 positions[i] = -ENOENT;
1799                 while (prim_hitmask[i]) {
1800                         uint32_t hit_index =
1801                                         __builtin_ctzl(prim_hitmask[i])
1802                                         >> 1;
1803                         uint32_t key_idx =
1804                                 primary_bkt[i]->key_idx[hit_index];
1805                         const struct rte_hash_key *key_slot =
1806                                 (const struct rte_hash_key *)(
1807                                 (const char *)h->key_store +
1808                                 key_idx * h->key_entry_size);
1809
1810                         /*
1811                          * If key index is 0, do not compare key,
1812                          * as it is checking the dummy slot
1813                          */
1814                         if (!!key_idx &
1815                                 !rte_hash_cmp_eq(
1816                                         key_slot->key, keys[i], h)) {
1817                                 if (data != NULL)
1818                                         data[i] = key_slot->pdata;
1819
1820                                 hits |= 1ULL << i;
1821                                 positions[i] = key_idx - 1;
1822                                 goto next_key;
1823                         }
1824                         prim_hitmask[i] &= ~(3ULL << (hit_index << 1));
1825                 }
1826
1827                 while (sec_hitmask[i]) {
1828                         uint32_t hit_index =
1829                                         __builtin_ctzl(sec_hitmask[i])
1830                                         >> 1;
1831                         uint32_t key_idx =
1832                                 secondary_bkt[i]->key_idx[hit_index];
1833                         const struct rte_hash_key *key_slot =
1834                                 (const struct rte_hash_key *)(
1835                                 (const char *)h->key_store +
1836                                 key_idx * h->key_entry_size);
1837
1838                         /*
1839                          * If key index is 0, do not compare key,
1840                          * as it is checking the dummy slot
1841                          */
1842
1843                         if (!!key_idx &
1844                                 !rte_hash_cmp_eq(
1845                                         key_slot->key, keys[i], h)) {
1846                                 if (data != NULL)
1847                                         data[i] = key_slot->pdata;
1848
1849                                 hits |= 1ULL << i;
1850                                 positions[i] = key_idx - 1;
1851                                 goto next_key;
1852                         }
1853                         sec_hitmask[i] &= ~(3ULL << (hit_index << 1));
1854                 }
1855 next_key:
1856                 continue;
1857         }
1858
1859         /* all found, do not need to go through ext bkt */
1860         if ((hits == ((1ULL << num_keys) - 1)) || !h->ext_table_support) {
1861                 if (hit_mask != NULL)
1862                         *hit_mask = hits;
1863                 __hash_rw_reader_unlock(h);
1864                 return;
1865         }
1866
1867         /* need to check ext buckets for match */
1868         for (i = 0; i < num_keys; i++) {
1869                 if ((hits & (1ULL << i)) != 0)
1870                         continue;
1871                 next_bkt = secondary_bkt[i]->next;
1872                 FOR_EACH_BUCKET(cur_bkt, next_bkt) {
1873                         if (data != NULL)
1874                                 ret = search_one_bucket_l(h, keys[i],
1875                                                 sig[i], &data[i], cur_bkt);
1876                         else
1877                                 ret = search_one_bucket_l(h, keys[i],
1878                                                 sig[i], NULL, cur_bkt);
1879                         if (ret != -1) {
1880                                 positions[i] = ret;
1881                                 hits |= 1ULL << i;
1882                                 break;
1883                         }
1884                 }
1885         }
1886
1887         __hash_rw_reader_unlock(h);
1888
1889         if (hit_mask != NULL)
1890                 *hit_mask = hits;
1891 }
1892
1893 static inline void
1894 __rte_hash_lookup_bulk_lf(const struct rte_hash *h, const void **keys,
1895                         int32_t num_keys, int32_t *positions,
1896                         uint64_t *hit_mask, void *data[])
1897 {
1898         uint64_t hits = 0;
1899         int32_t i;
1900         int32_t ret;
1901         uint32_t prim_hash[RTE_HASH_LOOKUP_BULK_MAX];
1902         uint32_t prim_index[RTE_HASH_LOOKUP_BULK_MAX];
1903         uint32_t sec_index[RTE_HASH_LOOKUP_BULK_MAX];
1904         uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
1905         const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1906         const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1907         uint32_t prim_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1908         uint32_t sec_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1909         struct rte_hash_bucket *cur_bkt, *next_bkt;
1910         uint32_t cnt_b, cnt_a;
1911
1912         /* Prefetch first keys */
1913         for (i = 0; i < PREFETCH_OFFSET && i < num_keys; i++)
1914                 rte_prefetch0(keys[i]);
1915
1916         /*
1917          * Prefetch rest of the keys, calculate primary and
1918          * secondary bucket and prefetch them
1919          */
1920         for (i = 0; i < (num_keys - PREFETCH_OFFSET); i++) {
1921                 rte_prefetch0(keys[i + PREFETCH_OFFSET]);
1922
1923                 prim_hash[i] = rte_hash_hash(h, keys[i]);
1924
1925                 sig[i] = get_short_sig(prim_hash[i]);
1926                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1927                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1928
1929                 primary_bkt[i] = &h->buckets[prim_index[i]];
1930                 secondary_bkt[i] = &h->buckets[sec_index[i]];
1931
1932                 rte_prefetch0(primary_bkt[i]);
1933                 rte_prefetch0(secondary_bkt[i]);
1934         }
1935
1936         /* Calculate and prefetch rest of the buckets */
1937         for (; i < num_keys; i++) {
1938                 prim_hash[i] = rte_hash_hash(h, keys[i]);
1939
1940                 sig[i] = get_short_sig(prim_hash[i]);
1941                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1942                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1943
1944                 primary_bkt[i] = &h->buckets[prim_index[i]];
1945                 secondary_bkt[i] = &h->buckets[sec_index[i]];
1946
1947                 rte_prefetch0(primary_bkt[i]);
1948                 rte_prefetch0(secondary_bkt[i]);
1949         }
1950
1951         for (i = 0; i < num_keys; i++)
1952                 positions[i] = -ENOENT;
1953
1954         do {
1955                 /* Load the table change counter before the lookup
1956                  * starts. Acquire semantics will make sure that
1957                  * loads in compare_signatures are not hoisted.
1958                  */
1959                 cnt_b = __atomic_load_n(h->tbl_chng_cnt,
1960                                         __ATOMIC_ACQUIRE);
1961
1962                 /* Compare signatures and prefetch key slot of first hit */
1963                 for (i = 0; i < num_keys; i++) {
1964                         compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
1965                                 primary_bkt[i], secondary_bkt[i],
1966                                 sig[i], h->sig_cmp_fn);
1967
1968                         if (prim_hitmask[i]) {
1969                                 uint32_t first_hit =
1970                                                 __builtin_ctzl(prim_hitmask[i])
1971                                                 >> 1;
1972                                 uint32_t key_idx =
1973                                         primary_bkt[i]->key_idx[first_hit];
1974                                 const struct rte_hash_key *key_slot =
1975                                         (const struct rte_hash_key *)(
1976                                         (const char *)h->key_store +
1977                                         key_idx * h->key_entry_size);
1978                                 rte_prefetch0(key_slot);
1979                                 continue;
1980                         }
1981
1982                         if (sec_hitmask[i]) {
1983                                 uint32_t first_hit =
1984                                                 __builtin_ctzl(sec_hitmask[i])
1985                                                 >> 1;
1986                                 uint32_t key_idx =
1987                                         secondary_bkt[i]->key_idx[first_hit];
1988                                 const struct rte_hash_key *key_slot =
1989                                         (const struct rte_hash_key *)(
1990                                         (const char *)h->key_store +
1991                                         key_idx * h->key_entry_size);
1992                                 rte_prefetch0(key_slot);
1993                         }
1994                 }
1995
1996                 /* Compare keys, first hits in primary first */
1997                 for (i = 0; i < num_keys; i++) {
1998                         while (prim_hitmask[i]) {
1999                                 uint32_t hit_index =
2000                                                 __builtin_ctzl(prim_hitmask[i])
2001                                                 >> 1;
2002                                 uint32_t key_idx =
2003                                 __atomic_load_n(
2004                                         &primary_bkt[i]->key_idx[hit_index],
2005                                         __ATOMIC_ACQUIRE);
2006                                 const struct rte_hash_key *key_slot =
2007                                         (const struct rte_hash_key *)(
2008                                         (const char *)h->key_store +
2009                                         key_idx * h->key_entry_size);
2010
2011                                 /*
2012                                  * If key index is 0, do not compare key,
2013                                  * as it is checking the dummy slot
2014                                  */
2015                                 if (!!key_idx &
2016                                         !rte_hash_cmp_eq(
2017                                                 key_slot->key, keys[i], h)) {
2018                                         if (data != NULL)
2019                                                 data[i] = __atomic_load_n(
2020                                                         &key_slot->pdata,
2021                                                         __ATOMIC_ACQUIRE);
2022
2023                                         hits |= 1ULL << i;
2024                                         positions[i] = key_idx - 1;
2025                                         goto next_key;
2026                                 }
2027                                 prim_hitmask[i] &= ~(3ULL << (hit_index << 1));
2028                         }
2029
2030                         while (sec_hitmask[i]) {
2031                                 uint32_t hit_index =
2032                                                 __builtin_ctzl(sec_hitmask[i])
2033                                                 >> 1;
2034                                 uint32_t key_idx =
2035                                 __atomic_load_n(
2036                                         &secondary_bkt[i]->key_idx[hit_index],
2037                                         __ATOMIC_ACQUIRE);
2038                                 const struct rte_hash_key *key_slot =
2039                                         (const struct rte_hash_key *)(
2040                                         (const char *)h->key_store +
2041                                         key_idx * h->key_entry_size);
2042
2043                                 /*
2044                                  * If key index is 0, do not compare key,
2045                                  * as it is checking the dummy slot
2046                                  */
2047
2048                                 if (!!key_idx &
2049                                         !rte_hash_cmp_eq(
2050                                                 key_slot->key, keys[i], h)) {
2051                                         if (data != NULL)
2052                                                 data[i] = __atomic_load_n(
2053                                                         &key_slot->pdata,
2054                                                         __ATOMIC_ACQUIRE);
2055
2056                                         hits |= 1ULL << i;
2057                                         positions[i] = key_idx - 1;
2058                                         goto next_key;
2059                                 }
2060                                 sec_hitmask[i] &= ~(3ULL << (hit_index << 1));
2061                         }
2062 next_key:
2063                         continue;
2064                 }
2065
2066                 /* all found, do not need to go through ext bkt */
2067                 if (hits == ((1ULL << num_keys) - 1)) {
2068                         if (hit_mask != NULL)
2069                                 *hit_mask = hits;
2070                         return;
2071                 }
2072                 /* need to check ext buckets for match */
2073                 if (h->ext_table_support) {
2074                         for (i = 0; i < num_keys; i++) {
2075                                 if ((hits & (1ULL << i)) != 0)
2076                                         continue;
2077                                 next_bkt = secondary_bkt[i]->next;
2078                                 FOR_EACH_BUCKET(cur_bkt, next_bkt) {
2079                                         if (data != NULL)
2080                                                 ret = search_one_bucket_lf(h,
2081                                                         keys[i], sig[i],
2082                                                         &data[i], cur_bkt);
2083                                         else
2084                                                 ret = search_one_bucket_lf(h,
2085                                                                 keys[i], sig[i],
2086                                                                 NULL, cur_bkt);
2087                                         if (ret != -1) {
2088                                                 positions[i] = ret;
2089                                                 hits |= 1ULL << i;
2090                                                 break;
2091                                         }
2092                                 }
2093                         }
2094                 }
2095                 /* The loads of sig_current in compare_signatures
2096                  * should not move below the load from tbl_chng_cnt.
2097                  */
2098                 __atomic_thread_fence(__ATOMIC_ACQUIRE);
2099                 /* Re-read the table change counter to check if the
2100                  * table has changed during search. If yes, re-do
2101                  * the search.
2102                  * This load should not get hoisted. The load
2103                  * acquires on cnt_b, primary key index and secondary
2104                  * key index will make sure that it does not get
2105                  * hoisted.
2106                  */
2107                 cnt_a = __atomic_load_n(h->tbl_chng_cnt,
2108                                         __ATOMIC_ACQUIRE);
2109         } while (cnt_b != cnt_a);
2110
2111         if (hit_mask != NULL)
2112                 *hit_mask = hits;
2113 }
2114
2115 static inline void
2116 __rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
2117                         int32_t num_keys, int32_t *positions,
2118                         uint64_t *hit_mask, void *data[])
2119 {
2120         if (h->readwrite_concur_lf_support)
2121                 __rte_hash_lookup_bulk_lf(h, keys, num_keys, positions,
2122                                           hit_mask, data);
2123         else
2124                 __rte_hash_lookup_bulk_l(h, keys, num_keys, positions,
2125                                          hit_mask, data);
2126 }
2127
2128 int
2129 rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
2130                       uint32_t num_keys, int32_t *positions)
2131 {
2132         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
2133                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2134                         (positions == NULL)), -EINVAL);
2135
2136         __rte_hash_lookup_bulk(h, keys, num_keys, positions, NULL, NULL);
2137         return 0;
2138 }
2139
2140 int
2141 rte_hash_lookup_bulk_data(const struct rte_hash *h, const void **keys,
2142                       uint32_t num_keys, uint64_t *hit_mask, void *data[])
2143 {
2144         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
2145                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2146                         (hit_mask == NULL)), -EINVAL);
2147
2148         int32_t positions[num_keys];
2149
2150         __rte_hash_lookup_bulk(h, keys, num_keys, positions, hit_mask, data);
2151
2152         /* Return number of hits */
2153         return __builtin_popcountl(*hit_mask);
2154 }
2155
2156 int32_t
2157 rte_hash_iterate(const struct rte_hash *h, const void **key, void **data, uint32_t *next)
2158 {
2159         uint32_t bucket_idx, idx, position;
2160         struct rte_hash_key *next_key;
2161
2162         RETURN_IF_TRUE(((h == NULL) || (next == NULL)), -EINVAL);
2163
2164         const uint32_t total_entries_main = h->num_buckets *
2165                                                         RTE_HASH_BUCKET_ENTRIES;
2166         const uint32_t total_entries = total_entries_main << 1;
2167
2168         /* Out of bounds of all buckets (both main table and ext table) */
2169         if (*next >= total_entries_main)
2170                 goto extend_table;
2171
2172         /* Calculate bucket and index of current iterator */
2173         bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
2174         idx = *next % RTE_HASH_BUCKET_ENTRIES;
2175
2176         /* If current position is empty, go to the next one */
2177         while ((position = __atomic_load_n(&h->buckets[bucket_idx].key_idx[idx],
2178                                         __ATOMIC_ACQUIRE)) == EMPTY_SLOT) {
2179                 (*next)++;
2180                 /* End of table */
2181                 if (*next == total_entries_main)
2182                         goto extend_table;
2183                 bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
2184                 idx = *next % RTE_HASH_BUCKET_ENTRIES;
2185         }
2186
2187         __hash_rw_reader_lock(h);
2188         next_key = (struct rte_hash_key *) ((char *)h->key_store +
2189                                 position * h->key_entry_size);
2190         /* Return key and data */
2191         *key = next_key->key;
2192         *data = next_key->pdata;
2193
2194         __hash_rw_reader_unlock(h);
2195
2196         /* Increment iterator */
2197         (*next)++;
2198
2199         return position - 1;
2200
2201 /* Begin to iterate extendable buckets */
2202 extend_table:
2203         /* Out of total bound or if ext bucket feature is not enabled */
2204         if (*next >= total_entries || !h->ext_table_support)
2205                 return -ENOENT;
2206
2207         bucket_idx = (*next - total_entries_main) / RTE_HASH_BUCKET_ENTRIES;
2208         idx = (*next - total_entries_main) % RTE_HASH_BUCKET_ENTRIES;
2209
2210         while ((position = h->buckets_ext[bucket_idx].key_idx[idx]) == EMPTY_SLOT) {
2211                 (*next)++;
2212                 if (*next == total_entries)
2213                         return -ENOENT;
2214                 bucket_idx = (*next - total_entries_main) /
2215                                                 RTE_HASH_BUCKET_ENTRIES;
2216                 idx = (*next - total_entries_main) % RTE_HASH_BUCKET_ENTRIES;
2217         }
2218         __hash_rw_reader_lock(h);
2219         next_key = (struct rte_hash_key *) ((char *)h->key_store +
2220                                 position * h->key_entry_size);
2221         /* Return key and data */
2222         *key = next_key->key;
2223         *data = next_key->pdata;
2224
2225         __hash_rw_reader_unlock(h);
2226
2227         /* Increment iterator */
2228         (*next)++;
2229         return position - 1;
2230 }