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_elem.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 int hw_trans_mem_support = 0, use_local_cache = 0;
 140         unsigned int ext_table_support = 0;
 141         unsigned int readwrite_concur_support = 0;
 142         unsigned int writer_takes_lock = 0;
 143         unsigned int no_free_on_del = 0;
 144         uint32_t *ext_bkt_to_free = NULL;
 145         uint32_t *tbl_chng_cnt = NULL;
 146         unsigned int readwrite_concur_lf_support = 0;
 147         uint32_t i;
 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_elem(ring_name, sizeof(uint32_t),
 217                         rte_align32pow2(num_key_slots), 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_elem(ext_ring_name, sizeof(uint32_t),
 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_elem(r_ext, &i, sizeof(uint32_t));
 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_elem(r, &i, sizeof(uint32_t));
 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         uint32_t tot_ring_cnt, i;
 574
 575         if (h == NULL)
 576                 return;
 577
 578         __hash_rw_writer_lock(h);
 579         memset(h->buckets, 0, h->num_buckets * sizeof(struct rte_hash_bucket));
 580         memset(h->key_store, 0, h->key_entry_size * (h->entries + 1));
 581         *h->tbl_chng_cnt = 0;
 582
 583         /* reset the free ring */
 584         rte_ring_reset(h->free_slots);
 585
 586         /* flush free extendable bucket ring and memory */
 587         if (h->ext_table_support) {
 588                 memset(h->buckets_ext, 0, h->num_buckets *
 589                                                 sizeof(struct rte_hash_bucket));
 590                 rte_ring_reset(h->free_ext_bkts);
 591         }
 592
 593         /* Repopulate the free slots ring. Entry zero is reserved for key misses */
 594         if (h->use_local_cache)
 595                 tot_ring_cnt = h->entries + (RTE_MAX_LCORE - 1) *
 596                                         (LCORE_CACHE_SIZE - 1);
 597         else
 598                 tot_ring_cnt = h->entries;
 599
 600         for (i = 1; i < tot_ring_cnt + 1; i++)
 601                 rte_ring_sp_enqueue_elem(h->free_slots, &i, sizeof(uint32_t));
 602
 603         /* Repopulate the free ext bkt ring. */
 604         if (h->ext_table_support) {
 605                 for (i = 1; i <= h->num_buckets; i++)
 606                         rte_ring_sp_enqueue_elem(h->free_ext_bkts, &i,
 607                                                         sizeof(uint32_t));
 608         }
 609
 610         if (h->use_local_cache) {
 611                 /* Reset local caches per lcore */
 612                 for (i = 0; i < RTE_MAX_LCORE; i++)
 613                         h->local_free_slots[i].len = 0;
 614         }
 615         __hash_rw_writer_unlock(h);
 616 }
 617
 618 /*
 619  * Function called to enqueue back an index in the cache/ring,
 620  * as slot has not being used and it can be used in the
 621  * next addition attempt.
 622  */
 623 static inline void
 624 enqueue_slot_back(const struct rte_hash *h,
 625                 struct lcore_cache *cached_free_slots,
 626                 uint32_t slot_id)
 627 {
 628         if (h->use_local_cache) {
 629                 cached_free_slots->objs[cached_free_slots->len] = slot_id;
 630                 cached_free_slots->len++;
 631         } else
 632                 rte_ring_sp_enqueue_elem(h->free_slots, &slot_id,
 633                                                 sizeof(uint32_t));
 634 }
 635
 636 /* Search a key from bucket and update its data.
 637  * Writer holds the lock before calling this.
 638  */
 639 static inline int32_t
 640 search_and_update(const struct rte_hash *h, void *data, const void *key,
 641         struct rte_hash_bucket *bkt, uint16_t sig)
 642 {
 643         int i;
 644         struct rte_hash_key *k, *keys = h->key_store;
 645
 646         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 647                 if (bkt->sig_current[i] == sig) {
 648                         k = (struct rte_hash_key *) ((char *)keys +
 649                                         bkt->key_idx[i] * h->key_entry_size);
 650                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 651                                 /* The store to application data at *data
 652                                  * should not leak after the store to pdata
 653                                  * in the key store. i.e. pdata is the guard
 654                                  * variable. Release the application data
 655                                  * to the readers.
 656                                  */
 657                                 __atomic_store_n(&k->pdata,
 658                                         data,
 659                                         __ATOMIC_RELEASE);
 660                                 /*
 661                                  * Return index where key is stored,
 662                                  * subtracting the first dummy index
 663                                  */
 664                                 return bkt->key_idx[i] - 1;
 665                         }
 666                 }
 667         }
 668         return -1;
 669 }
 670
 671 /* Only tries to insert at one bucket (@prim_bkt) without trying to push
 672  * buckets around.
 673  * return 1 if matching existing key, return 0 if succeeds, return -1 for no
 674  * empty entry.
 675  */
 676 static inline int32_t
 677 rte_hash_cuckoo_insert_mw(const struct rte_hash *h,
 678                 struct rte_hash_bucket *prim_bkt,
 679                 struct rte_hash_bucket *sec_bkt,
 680                 const struct rte_hash_key *key, void *data,
 681                 uint16_t sig, uint32_t new_idx,
 682                 int32_t *ret_val)
 683 {
 684         unsigned int i;
 685         struct rte_hash_bucket *cur_bkt;
 686         int32_t ret;
 687
 688         __hash_rw_writer_lock(h);
 689         /* Check if key was inserted after last check but before this
 690          * protected region in case of inserting duplicated keys.
 691          */
 692         ret = search_and_update(h, data, key, prim_bkt, sig);
 693         if (ret != -1) {
 694                 __hash_rw_writer_unlock(h);
 695                 *ret_val = ret;
 696                 return 1;
 697         }
 698
 699         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
 700                 ret = search_and_update(h, data, key, cur_bkt, sig);
 701                 if (ret != -1) {
 702                         __hash_rw_writer_unlock(h);
 703                         *ret_val = ret;
 704                         return 1;
 705                 }
 706         }
 707
 708         /* Insert new entry if there is room in the primary
 709          * bucket.
 710          */
 711         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 712                 /* Check if slot is available */
 713                 if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
 714                         prim_bkt->sig_current[i] = sig;
 715                         /* Store to signature and key should not
 716                          * leak after the store to key_idx. i.e.
 717                          * key_idx is the guard variable for signature
 718                          * and key.
 719                          */
 720                         __atomic_store_n(&prim_bkt->key_idx[i],
 721                                          new_idx,
 722                                          __ATOMIC_RELEASE);
 723                         break;
 724                 }
 725         }
 726         __hash_rw_writer_unlock(h);
 727
 728         if (i != RTE_HASH_BUCKET_ENTRIES)
 729                 return 0;
 730
 731         /* no empty entry */
 732         return -1;
 733 }
 734
 735 /* Shift buckets along provided cuckoo_path (@leaf and @leaf_slot) and fill
 736  * the path head with new entry (sig, alt_hash, new_idx)
 737  * return 1 if matched key found, return -1 if cuckoo path invalided and fail,
 738  * return 0 if succeeds.
 739  */
 740 static inline int
 741 rte_hash_cuckoo_move_insert_mw(const struct rte_hash *h,
 742                         struct rte_hash_bucket *bkt,
 743                         struct rte_hash_bucket *alt_bkt,
 744                         const struct rte_hash_key *key, void *data,
 745                         struct queue_node *leaf, uint32_t leaf_slot,
 746                         uint16_t sig, uint32_t new_idx,
 747                         int32_t *ret_val)
 748 {
 749         uint32_t prev_alt_bkt_idx;
 750         struct rte_hash_bucket *cur_bkt;
 751         struct queue_node *prev_node, *curr_node = leaf;
 752         struct rte_hash_bucket *prev_bkt, *curr_bkt = leaf->bkt;
 753         uint32_t prev_slot, curr_slot = leaf_slot;
 754         int32_t ret;
 755
 756         __hash_rw_writer_lock(h);
 757
 758         /* In case empty slot was gone before entering protected region */
 759         if (curr_bkt->key_idx[curr_slot] != EMPTY_SLOT) {
 760                 __hash_rw_writer_unlock(h);
 761                 return -1;
 762         }
 763
 764         /* Check if key was inserted after last check but before this
 765          * protected region.
 766          */
 767         ret = search_and_update(h, data, key, bkt, sig);
 768         if (ret != -1) {
 769                 __hash_rw_writer_unlock(h);
 770                 *ret_val = ret;
 771                 return 1;
 772         }
 773
 774         FOR_EACH_BUCKET(cur_bkt, alt_bkt) {
 775                 ret = search_and_update(h, data, key, cur_bkt, sig);
 776                 if (ret != -1) {
 777                         __hash_rw_writer_unlock(h);
 778                         *ret_val = ret;
 779                         return 1;
 780                 }
 781         }
 782
 783         while (likely(curr_node->prev != NULL)) {
 784                 prev_node = curr_node->prev;
 785                 prev_bkt = prev_node->bkt;
 786                 prev_slot = curr_node->prev_slot;
 787
 788                 prev_alt_bkt_idx = get_alt_bucket_index(h,
 789                                         prev_node->cur_bkt_idx,
 790                                         prev_bkt->sig_current[prev_slot]);
 791
 792                 if (unlikely(&h->buckets[prev_alt_bkt_idx]
 793                                 != curr_bkt)) {
 794                         /* revert it to empty, otherwise duplicated keys */
 795                         __atomic_store_n(&curr_bkt->key_idx[curr_slot],
 796                                 EMPTY_SLOT,
 797                                 __ATOMIC_RELEASE);
 798                         __hash_rw_writer_unlock(h);
 799                         return -1;
 800                 }
 801
 802                 if (h->readwrite_concur_lf_support) {
 803                         /* Inform the previous move. The current move need
 804                          * not be informed now as the current bucket entry
 805                          * is present in both primary and secondary.
 806                          * Since there is one writer, load acquires on
 807                          * tbl_chng_cnt are not required.
 808                          */
 809                         __atomic_store_n(h->tbl_chng_cnt,
 810                                          *h->tbl_chng_cnt + 1,
 811                                          __ATOMIC_RELEASE);
 812                         /* The store to sig_current should not
 813                          * move above the store to tbl_chng_cnt.
 814                          */
 815                         __atomic_thread_fence(__ATOMIC_RELEASE);
 816                 }
 817
 818                 /* Need to swap current/alt sig to allow later
 819                  * Cuckoo insert to move elements back to its
 820                  * primary bucket if available
 821                  */
 822                 curr_bkt->sig_current[curr_slot] =
 823                         prev_bkt->sig_current[prev_slot];
 824                 /* Release the updated bucket entry */
 825                 __atomic_store_n(&curr_bkt->key_idx[curr_slot],
 826                         prev_bkt->key_idx[prev_slot],
 827                         __ATOMIC_RELEASE);
 828
 829                 curr_slot = prev_slot;
 830                 curr_node = prev_node;
 831                 curr_bkt = curr_node->bkt;
 832         }
 833
 834         if (h->readwrite_concur_lf_support) {
 835                 /* Inform the previous move. The current move need
 836                  * not be informed now as the current bucket entry
 837                  * is present in both primary and secondary.
 838                  * Since there is one writer, load acquires on
 839                  * tbl_chng_cnt are not required.
 840                  */
 841                 __atomic_store_n(h->tbl_chng_cnt,
 842                                  *h->tbl_chng_cnt + 1,
 843                                  __ATOMIC_RELEASE);
 844                 /* The store to sig_current should not
 845                  * move above the store to tbl_chng_cnt.
 846                  */
 847                 __atomic_thread_fence(__ATOMIC_RELEASE);
 848         }
 849
 850         curr_bkt->sig_current[curr_slot] = sig;
 851         /* Release the new bucket entry */
 852         __atomic_store_n(&curr_bkt->key_idx[curr_slot],
 853                          new_idx,
 854                          __ATOMIC_RELEASE);
 855
 856         __hash_rw_writer_unlock(h);
 857
 858         return 0;
 859
 860 }
 861
 862 /*
 863  * Make space for new key, using bfs Cuckoo Search and Multi-Writer safe
 864  * Cuckoo
 865  */
 866 static inline int
 867 rte_hash_cuckoo_make_space_mw(const struct rte_hash *h,
 868                         struct rte_hash_bucket *bkt,
 869                         struct rte_hash_bucket *sec_bkt,
 870                         const struct rte_hash_key *key, void *data,
 871                         uint16_t sig, uint32_t bucket_idx,
 872                         uint32_t new_idx, int32_t *ret_val)
 873 {
 874         unsigned int i;
 875         struct queue_node queue[RTE_HASH_BFS_QUEUE_MAX_LEN];
 876         struct queue_node *tail, *head;
 877         struct rte_hash_bucket *curr_bkt, *alt_bkt;
 878         uint32_t cur_idx, alt_idx;
 879
 880         tail = queue;
 881         head = queue + 1;
 882         tail->bkt = bkt;
 883         tail->prev = NULL;
 884         tail->prev_slot = -1;
 885         tail->cur_bkt_idx = bucket_idx;
 886
 887         /* Cuckoo bfs Search */
 888         while (likely(tail != head && head <
 889                                         queue + RTE_HASH_BFS_QUEUE_MAX_LEN -
 890                                         RTE_HASH_BUCKET_ENTRIES)) {
 891                 curr_bkt = tail->bkt;
 892                 cur_idx = tail->cur_bkt_idx;
 893                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 894                         if (curr_bkt->key_idx[i] == EMPTY_SLOT) {
 895                                 int32_t ret = rte_hash_cuckoo_move_insert_mw(h,
 896                                                 bkt, sec_bkt, key, data,
 897                                                 tail, i, sig,
 898                                                 new_idx, ret_val);
 899                                 if (likely(ret != -1))
 900                                         return ret;
 901                         }
 902
 903                         /* Enqueue new node and keep prev node info */
 904                         alt_idx = get_alt_bucket_index(h, cur_idx,
 905                                                 curr_bkt->sig_current[i]);
 906                         alt_bkt = &(h->buckets[alt_idx]);
 907                         head->bkt = alt_bkt;
 908                         head->cur_bkt_idx = alt_idx;
 909                         head->prev = tail;
 910                         head->prev_slot = i;
 911                         head++;
 912                 }
 913                 tail++;
 914         }
 915
 916         return -ENOSPC;
 917 }
 918
 919 static inline int32_t
 920 __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
 921                                                 hash_sig_t sig, void *data)
 922 {
 923         uint16_t short_sig;
 924         uint32_t prim_bucket_idx, sec_bucket_idx;
 925         struct rte_hash_bucket *prim_bkt, *sec_bkt, *cur_bkt;
 926         struct rte_hash_key *new_k, *keys = h->key_store;
 927         uint32_t slot_id;
 928         uint32_t ext_bkt_id;
 929         int ret;
 930         unsigned n_slots;
 931         unsigned lcore_id;
 932         unsigned int i;
 933         struct lcore_cache *cached_free_slots = NULL;
 934         int32_t ret_val;
 935         struct rte_hash_bucket *last;
 936
 937         short_sig = get_short_sig(sig);
 938         prim_bucket_idx = get_prim_bucket_index(h, sig);
 939         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
 940         prim_bkt = &h->buckets[prim_bucket_idx];
 941         sec_bkt = &h->buckets[sec_bucket_idx];
 942         rte_prefetch0(prim_bkt);
 943         rte_prefetch0(sec_bkt);
 944
 945         /* Check if key is already inserted in primary location */
 946         __hash_rw_writer_lock(h);
 947         ret = search_and_update(h, data, key, prim_bkt, short_sig);
 948         if (ret != -1) {
 949                 __hash_rw_writer_unlock(h);
 950                 return ret;
 951         }
 952
 953         /* Check if key is already inserted in secondary location */
 954         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
 955                 ret = search_and_update(h, data, key, cur_bkt, short_sig);
 956                 if (ret != -1) {
 957                         __hash_rw_writer_unlock(h);
 958                         return ret;
 959                 }
 960         }
 961
 962         __hash_rw_writer_unlock(h);
 963
 964         /* Did not find a match, so get a new slot for storing the new key */
 965         if (h->use_local_cache) {
 966                 lcore_id = rte_lcore_id();
 967                 cached_free_slots = &h->local_free_slots[lcore_id];
 968                 /* Try to get a free slot from the local cache */
 969                 if (cached_free_slots->len == 0) {
 970                         /* Need to get another burst of free slots from global ring */
 971                         n_slots = rte_ring_mc_dequeue_burst_elem(h->free_slots,
 972                                         cached_free_slots->objs,
 973                                         sizeof(uint32_t),
 974                                         LCORE_CACHE_SIZE, NULL);
 975                         if (n_slots == 0) {
 976                                 return -ENOSPC;
 977                         }
 978
 979                         cached_free_slots->len += n_slots;
 980                 }
 981
 982                 /* Get a free slot from the local cache */
 983                 cached_free_slots->len--;
 984                 slot_id = cached_free_slots->objs[cached_free_slots->len];
 985         } else {
 986                 if (rte_ring_sc_dequeue_elem(h->free_slots, &slot_id,
 987                                                 sizeof(uint32_t)) != 0) {
 988                         return -ENOSPC;
 989                 }
 990         }
 991
 992         new_k = RTE_PTR_ADD(keys, slot_id * h->key_entry_size);
 993         /* The store to application data (by the application) at *data should
 994          * not leak after the store of pdata in the key store. i.e. pdata is
 995          * the guard variable. Release the application data to the readers.
 996          */
 997         __atomic_store_n(&new_k->pdata,
 998                 data,
 999                 __ATOMIC_RELEASE);
1000         /* Copy key */
1001         memcpy(new_k->key, key, h->key_len);
1002
1003         /* Find an empty slot and insert */
1004         ret = rte_hash_cuckoo_insert_mw(h, prim_bkt, sec_bkt, key, data,
1005                                         short_sig, slot_id, &ret_val);
1006         if (ret == 0)
1007                 return slot_id - 1;
1008         else if (ret == 1) {
1009                 enqueue_slot_back(h, cached_free_slots, slot_id);
1010                 return ret_val;
1011         }
1012
1013         /* Primary bucket full, need to make space for new entry */
1014         ret = rte_hash_cuckoo_make_space_mw(h, prim_bkt, sec_bkt, key, data,
1015                                 short_sig, prim_bucket_idx, slot_id, &ret_val);
1016         if (ret == 0)
1017                 return slot_id - 1;
1018         else if (ret == 1) {
1019                 enqueue_slot_back(h, cached_free_slots, slot_id);
1020                 return ret_val;
1021         }
1022
1023         /* Also search secondary bucket to get better occupancy */
1024         ret = rte_hash_cuckoo_make_space_mw(h, sec_bkt, prim_bkt, key, data,
1025                                 short_sig, sec_bucket_idx, slot_id, &ret_val);
1026
1027         if (ret == 0)
1028                 return slot_id - 1;
1029         else if (ret == 1) {
1030                 enqueue_slot_back(h, cached_free_slots, slot_id);
1031                 return ret_val;
1032         }
1033
1034         /* if ext table not enabled, we failed the insertion */
1035         if (!h->ext_table_support) {
1036                 enqueue_slot_back(h, cached_free_slots, slot_id);
1037                 return ret;
1038         }
1039
1040         /* Now we need to go through the extendable bucket. Protection is needed
1041          * to protect all extendable bucket processes.
1042          */
1043         __hash_rw_writer_lock(h);
1044         /* We check for duplicates again since could be inserted before the lock */
1045         ret = search_and_update(h, data, key, prim_bkt, short_sig);
1046         if (ret != -1) {
1047                 enqueue_slot_back(h, cached_free_slots, slot_id);
1048                 goto failure;
1049         }
1050
1051         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1052                 ret = search_and_update(h, data, key, cur_bkt, short_sig);
1053                 if (ret != -1) {
1054                         enqueue_slot_back(h, cached_free_slots, slot_id);
1055                         goto failure;
1056                 }
1057         }
1058
1059         /* Search sec and ext buckets to find an empty entry to insert. */
1060         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1061                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1062                         /* Check if slot is available */
1063                         if (likely(cur_bkt->key_idx[i] == EMPTY_SLOT)) {
1064                                 cur_bkt->sig_current[i] = short_sig;
1065                                 /* Store to signature and key should not
1066                                  * leak after the store to key_idx. i.e.
1067                                  * key_idx is the guard variable for signature
1068                                  * and key.
1069                                  */
1070                                 __atomic_store_n(&cur_bkt->key_idx[i],
1071                                                  slot_id,
1072                                                  __ATOMIC_RELEASE);
1073                                 __hash_rw_writer_unlock(h);
1074                                 return slot_id - 1;
1075                         }
1076                 }
1077         }
1078
1079         /* Failed to get an empty entry from extendable buckets. Link a new
1080          * extendable bucket. We first get a free bucket from ring.
1081          */
1082         if (rte_ring_sc_dequeue_elem(h->free_ext_bkts, &ext_bkt_id,
1083                                                 sizeof(uint32_t)) != 0) {
1084                 ret = -ENOSPC;
1085                 goto failure;
1086         }
1087
1088         /* Use the first location of the new bucket */
1089         (h->buckets_ext[ext_bkt_id - 1]).sig_current[0] = short_sig;
1090         /* Store to signature and key should not leak after
1091          * the store to key_idx. i.e. key_idx is the guard variable
1092          * for signature and key.
1093          */
1094         __atomic_store_n(&(h->buckets_ext[ext_bkt_id - 1]).key_idx[0],
1095                          slot_id,
1096                          __ATOMIC_RELEASE);
1097         /* Link the new bucket to sec bucket linked list */
1098         last = rte_hash_get_last_bkt(sec_bkt);
1099         last->next = &h->buckets_ext[ext_bkt_id - 1];
1100         __hash_rw_writer_unlock(h);
1101         return slot_id - 1;
1102
1103 failure:
1104         __hash_rw_writer_unlock(h);
1105         return ret;
1106
1107 }
1108
1109 int32_t
1110 rte_hash_add_key_with_hash(const struct rte_hash *h,
1111                         const void *key, hash_sig_t sig)
1112 {
1113         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1114         return __rte_hash_add_key_with_hash(h, key, sig, 0);
1115 }
1116
1117 int32_t
1118 rte_hash_add_key(const struct rte_hash *h, const void *key)
1119 {
1120         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1121         return __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), 0);
1122 }
1123
1124 int
1125 rte_hash_add_key_with_hash_data(const struct rte_hash *h,
1126                         const void *key, hash_sig_t sig, void *data)
1127 {
1128         int ret;
1129
1130         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1131         ret = __rte_hash_add_key_with_hash(h, key, sig, data);
1132         if (ret >= 0)
1133                 return 0;
1134         else
1135                 return ret;
1136 }
1137
1138 int
1139 rte_hash_add_key_data(const struct rte_hash *h, const void *key, void *data)
1140 {
1141         int ret;
1142
1143         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1144
1145         ret = __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), data);
1146         if (ret >= 0)
1147                 return 0;
1148         else
1149                 return ret;
1150 }
1151
1152 /* Search one bucket to find the match key - uses rw lock */
1153 static inline int32_t
1154 search_one_bucket_l(const struct rte_hash *h, const void *key,
1155                 uint16_t sig, void **data,
1156                 const struct rte_hash_bucket *bkt)
1157 {
1158         int i;
1159         struct rte_hash_key *k, *keys = h->key_store;
1160
1161         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1162                 if (bkt->sig_current[i] == sig &&
1163                                 bkt->key_idx[i] != EMPTY_SLOT) {
1164                         k = (struct rte_hash_key *) ((char *)keys +
1165                                         bkt->key_idx[i] * h->key_entry_size);
1166
1167                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1168                                 if (data != NULL)
1169                                         *data = k->pdata;
1170                                 /*
1171                                  * Return index where key is stored,
1172                                  * subtracting the first dummy index
1173                                  */
1174                                 return bkt->key_idx[i] - 1;
1175                         }
1176                 }
1177         }
1178         return -1;
1179 }
1180
1181 /* Search one bucket to find the match key */
1182 static inline int32_t
1183 search_one_bucket_lf(const struct rte_hash *h, const void *key, uint16_t sig,
1184                         void **data, const struct rte_hash_bucket *bkt)
1185 {
1186         int i;
1187         uint32_t key_idx;
1188         struct rte_hash_key *k, *keys = h->key_store;
1189
1190         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1191                 /* Signature comparison is done before the acquire-load
1192                  * of the key index to achieve better performance.
1193                  * This can result in the reader loading old signature
1194                  * (which matches), while the key_idx is updated to a
1195                  * value that belongs to a new key. However, the full
1196                  * key comparison will ensure that the lookup fails.
1197                  */
1198                 if (bkt->sig_current[i] == sig) {
1199                         key_idx = __atomic_load_n(&bkt->key_idx[i],
1200                                           __ATOMIC_ACQUIRE);
1201                         if (key_idx != EMPTY_SLOT) {
1202                                 k = (struct rte_hash_key *) ((char *)keys +
1203                                                 key_idx * h->key_entry_size);
1204
1205                                 if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1206                                         if (data != NULL) {
1207                                                 *data = __atomic_load_n(
1208                                                         &k->pdata,
1209                                                         __ATOMIC_ACQUIRE);
1210                                         }
1211                                         /*
1212                                          * Return index where key is stored,
1213                                          * subtracting the first dummy index
1214                                          */
1215                                         return key_idx - 1;
1216                                 }
1217                         }
1218                 }
1219         }
1220         return -1;
1221 }
1222
1223 static inline int32_t
1224 __rte_hash_lookup_with_hash_l(const struct rte_hash *h, const void *key,
1225                                 hash_sig_t sig, void **data)
1226 {
1227         uint32_t prim_bucket_idx, sec_bucket_idx;
1228         struct rte_hash_bucket *bkt, *cur_bkt;
1229         int ret;
1230         uint16_t short_sig;
1231
1232         short_sig = get_short_sig(sig);
1233         prim_bucket_idx = get_prim_bucket_index(h, sig);
1234         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1235
1236         bkt = &h->buckets[prim_bucket_idx];
1237
1238         __hash_rw_reader_lock(h);
1239
1240         /* Check if key is in primary location */
1241         ret = search_one_bucket_l(h, key, short_sig, data, bkt);
1242         if (ret != -1) {
1243                 __hash_rw_reader_unlock(h);
1244                 return ret;
1245         }
1246         /* Calculate secondary hash */
1247         bkt = &h->buckets[sec_bucket_idx];
1248
1249         /* Check if key is in secondary location */
1250         FOR_EACH_BUCKET(cur_bkt, bkt) {
1251                 ret = search_one_bucket_l(h, key, short_sig,
1252                                         data, cur_bkt);
1253                 if (ret != -1) {
1254                         __hash_rw_reader_unlock(h);
1255                         return ret;
1256                 }
1257         }
1258
1259         __hash_rw_reader_unlock(h);
1260
1261         return -ENOENT;
1262 }
1263
1264 static inline int32_t
1265 __rte_hash_lookup_with_hash_lf(const struct rte_hash *h, const void *key,
1266                                         hash_sig_t sig, void **data)
1267 {
1268         uint32_t prim_bucket_idx, sec_bucket_idx;
1269         struct rte_hash_bucket *bkt, *cur_bkt;
1270         uint32_t cnt_b, cnt_a;
1271         int ret;
1272         uint16_t short_sig;
1273
1274         short_sig = get_short_sig(sig);
1275         prim_bucket_idx = get_prim_bucket_index(h, sig);
1276         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1277
1278         do {
1279                 /* Load the table change counter before the lookup
1280                  * starts. Acquire semantics will make sure that
1281                  * loads in search_one_bucket are not hoisted.
1282                  */
1283                 cnt_b = __atomic_load_n(h->tbl_chng_cnt,
1284                                 __ATOMIC_ACQUIRE);
1285
1286                 /* Check if key is in primary location */
1287                 bkt = &h->buckets[prim_bucket_idx];
1288                 ret = search_one_bucket_lf(h, key, short_sig, data, bkt);
1289                 if (ret != -1) {
1290                         __hash_rw_reader_unlock(h);
1291                         return ret;
1292                 }
1293                 /* Calculate secondary hash */
1294                 bkt = &h->buckets[sec_bucket_idx];
1295
1296                 /* Check if key is in secondary location */
1297                 FOR_EACH_BUCKET(cur_bkt, bkt) {
1298                         ret = search_one_bucket_lf(h, key, short_sig,
1299                                                 data, cur_bkt);
1300                         if (ret != -1) {
1301                                 __hash_rw_reader_unlock(h);
1302                                 return ret;
1303                         }
1304                 }
1305
1306                 /* The loads of sig_current in search_one_bucket
1307                  * should not move below the load from tbl_chng_cnt.
1308                  */
1309                 __atomic_thread_fence(__ATOMIC_ACQUIRE);
1310                 /* Re-read the table change counter to check if the
1311                  * table has changed during search. If yes, re-do
1312                  * the search.
1313                  * This load should not get hoisted. The load
1314                  * acquires on cnt_b, key index in primary bucket
1315                  * and key index in secondary bucket will make sure
1316                  * that it does not get hoisted.
1317                  */
1318                 cnt_a = __atomic_load_n(h->tbl_chng_cnt,
1319                                         __ATOMIC_ACQUIRE);
1320         } while (cnt_b != cnt_a);
1321
1322         return -ENOENT;
1323 }
1324
1325 static inline int32_t
1326 __rte_hash_lookup_with_hash(const struct rte_hash *h, const void *key,
1327                                         hash_sig_t sig, void **data)
1328 {
1329         if (h->readwrite_concur_lf_support)
1330                 return __rte_hash_lookup_with_hash_lf(h, key, sig, data);
1331         else
1332                 return __rte_hash_lookup_with_hash_l(h, key, sig, data);
1333 }
1334
1335 int32_t
1336 rte_hash_lookup_with_hash(const struct rte_hash *h,
1337                         const void *key, hash_sig_t sig)
1338 {
1339         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1340         return __rte_hash_lookup_with_hash(h, key, sig, NULL);
1341 }
1342
1343 int32_t
1344 rte_hash_lookup(const struct rte_hash *h, const void *key)
1345 {
1346         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1347         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), NULL);
1348 }
1349
1350 int
1351 rte_hash_lookup_with_hash_data(const struct rte_hash *h,
1352                         const void *key, hash_sig_t sig, void **data)
1353 {
1354         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1355         return __rte_hash_lookup_with_hash(h, key, sig, data);
1356 }
1357
1358 int
1359 rte_hash_lookup_data(const struct rte_hash *h, const void *key, void **data)
1360 {
1361         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
1362         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), data);
1363 }
1364
1365 static inline void
1366 remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i)
1367 {
1368         unsigned lcore_id, n_slots;
1369         struct lcore_cache *cached_free_slots;
1370
1371         if (h->use_local_cache) {
1372                 lcore_id = rte_lcore_id();
1373                 cached_free_slots = &h->local_free_slots[lcore_id];
1374                 /* Cache full, need to free it. */
1375                 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
1376                         /* Need to enqueue the free slots in global ring. */
1377                         n_slots = rte_ring_mp_enqueue_burst_elem(h->free_slots,
1378                                                 cached_free_slots->objs,
1379                                                 sizeof(uint32_t),
1380                                                 LCORE_CACHE_SIZE, NULL);
1381                         ERR_IF_TRUE((n_slots == 0),
1382                                 "%s: could not enqueue free slots in global ring\n",
1383                                 __func__);
1384                         cached_free_slots->len -= n_slots;
1385                 }
1386                 /* Put index of new free slot in cache. */
1387                 cached_free_slots->objs[cached_free_slots->len] =
1388                                                         bkt->key_idx[i];
1389                 cached_free_slots->len++;
1390         } else {
1391                 rte_ring_sp_enqueue_elem(h->free_slots,
1392                                 &bkt->key_idx[i], sizeof(uint32_t));
1393         }
1394 }
1395
1396 /* Compact the linked list by moving key from last entry in linked list to the
1397  * empty slot.
1398  */
1399 static inline void
1400 __rte_hash_compact_ll(const struct rte_hash *h,
1401                         struct rte_hash_bucket *cur_bkt, int pos) {
1402         int i;
1403         struct rte_hash_bucket *last_bkt;
1404
1405         if (!cur_bkt->next)
1406                 return;
1407
1408         last_bkt = rte_hash_get_last_bkt(cur_bkt);
1409
1410         for (i = RTE_HASH_BUCKET_ENTRIES - 1; i >= 0; i--) {
1411                 if (last_bkt->key_idx[i] != EMPTY_SLOT) {
1412                         cur_bkt->sig_current[pos] = last_bkt->sig_current[i];
1413                         __atomic_store_n(&cur_bkt->key_idx[pos],
1414                                          last_bkt->key_idx[i],
1415                                          __ATOMIC_RELEASE);
1416                         if (h->readwrite_concur_lf_support) {
1417                                 /* Inform the readers that the table has changed
1418                                  * Since there is one writer, load acquire on
1419                                  * tbl_chng_cnt is not required.
1420                                  */
1421                                 __atomic_store_n(h->tbl_chng_cnt,
1422                                          *h->tbl_chng_cnt + 1,
1423                                          __ATOMIC_RELEASE);
1424                                 /* The store to sig_current should
1425                                  * not move above the store to tbl_chng_cnt.
1426                                  */
1427                                 __atomic_thread_fence(__ATOMIC_RELEASE);
1428                         }
1429                         last_bkt->sig_current[i] = NULL_SIGNATURE;
1430                         __atomic_store_n(&last_bkt->key_idx[i],
1431                                          EMPTY_SLOT,
1432                                          __ATOMIC_RELEASE);
1433                         return;
1434                 }
1435         }
1436 }
1437
1438 /* Search one bucket and remove the matched key.
1439  * Writer is expected to hold the lock while calling this
1440  * function.
1441  */
1442 static inline int32_t
1443 search_and_remove(const struct rte_hash *h, const void *key,
1444                         struct rte_hash_bucket *bkt, uint16_t sig, int *pos)
1445 {
1446         struct rte_hash_key *k, *keys = h->key_store;
1447         unsigned int i;
1448         uint32_t key_idx;
1449
1450         /* Check if key is in bucket */
1451         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1452                 key_idx = __atomic_load_n(&bkt->key_idx[i],
1453                                           __ATOMIC_ACQUIRE);
1454                 if (bkt->sig_current[i] == sig && key_idx != EMPTY_SLOT) {
1455                         k = (struct rte_hash_key *) ((char *)keys +
1456                                         key_idx * h->key_entry_size);
1457                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
1458                                 bkt->sig_current[i] = NULL_SIGNATURE;
1459                                 /* Free the key store index if
1460                                  * no_free_on_del is disabled.
1461                                  */
1462                                 if (!h->no_free_on_del)
1463                                         remove_entry(h, bkt, i);
1464
1465                                 __atomic_store_n(&bkt->key_idx[i],
1466                                                  EMPTY_SLOT,
1467                                                  __ATOMIC_RELEASE);
1468
1469                                 *pos = i;
1470                                 /*
1471                                  * Return index where key is stored,
1472                                  * subtracting the first dummy index
1473                                  */
1474                                 return key_idx - 1;
1475                         }
1476                 }
1477         }
1478         return -1;
1479 }
1480
1481 static inline int32_t
1482 __rte_hash_del_key_with_hash(const struct rte_hash *h, const void *key,
1483                                                 hash_sig_t sig)
1484 {
1485         uint32_t prim_bucket_idx, sec_bucket_idx;
1486         struct rte_hash_bucket *prim_bkt, *sec_bkt, *prev_bkt, *last_bkt;
1487         struct rte_hash_bucket *cur_bkt;
1488         int pos;
1489         int32_t ret, i;
1490         uint16_t short_sig;
1491
1492         short_sig = get_short_sig(sig);
1493         prim_bucket_idx = get_prim_bucket_index(h, sig);
1494         sec_bucket_idx = get_alt_bucket_index(h, prim_bucket_idx, short_sig);
1495         prim_bkt = &h->buckets[prim_bucket_idx];
1496
1497         __hash_rw_writer_lock(h);
1498         /* look for key in primary bucket */
1499         ret = search_and_remove(h, key, prim_bkt, short_sig, &pos);
1500         if (ret != -1) {
1501                 __rte_hash_compact_ll(h, prim_bkt, pos);
1502                 last_bkt = prim_bkt->next;
1503                 prev_bkt = prim_bkt;
1504                 goto return_bkt;
1505         }
1506
1507         /* Calculate secondary hash */
1508         sec_bkt = &h->buckets[sec_bucket_idx];
1509
1510         FOR_EACH_BUCKET(cur_bkt, sec_bkt) {
1511                 ret = search_and_remove(h, key, cur_bkt, short_sig, &pos);
1512                 if (ret != -1) {
1513                         __rte_hash_compact_ll(h, cur_bkt, pos);
1514                         last_bkt = sec_bkt->next;
1515                         prev_bkt = sec_bkt;
1516                         goto return_bkt;
1517                 }
1518         }
1519
1520         __hash_rw_writer_unlock(h);
1521         return -ENOENT;
1522
1523 /* Search last bucket to see if empty to be recycled */
1524 return_bkt:
1525         if (!last_bkt) {
1526                 __hash_rw_writer_unlock(h);
1527                 return ret;
1528         }
1529         while (last_bkt->next) {
1530                 prev_bkt = last_bkt;
1531                 last_bkt = last_bkt->next;
1532         }
1533
1534         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1535                 if (last_bkt->key_idx[i] != EMPTY_SLOT)
1536                         break;
1537         }
1538         /* found empty bucket and recycle */
1539         if (i == RTE_HASH_BUCKET_ENTRIES) {
1540                 prev_bkt->next = NULL;
1541                 uint32_t index = last_bkt - h->buckets_ext + 1;
1542                 /* Recycle the empty bkt if
1543                  * no_free_on_del is disabled.
1544                  */
1545                 if (h->no_free_on_del)
1546                         /* Store index of an empty ext bkt to be recycled
1547                          * on calling rte_hash_del_xxx APIs.
1548                          * When lock free read-write concurrency is enabled,
1549                          * an empty ext bkt cannot be put into free list
1550                          * immediately (as readers might be using it still).
1551                          * Hence freeing of the ext bkt is piggy-backed to
1552                          * freeing of the key index.
1553                          */
1554                         h->ext_bkt_to_free[ret] = index;
1555                 else
1556                         rte_ring_sp_enqueue_elem(h->free_ext_bkts, &index,
1557                                                         sizeof(uint32_t));
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_elem(h->free_ext_bkts, &index,
1621                                                         sizeof(uint32_t));
1622                         h->ext_bkt_to_free[position] = 0;
1623                 }
1624         }
1625
1626         if (h->use_local_cache) {
1627                 lcore_id = rte_lcore_id();
1628                 cached_free_slots = &h->local_free_slots[lcore_id];
1629                 /* Cache full, need to free it. */
1630                 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
1631                         /* Need to enqueue the free slots in global ring. */
1632                         n_slots = rte_ring_mp_enqueue_burst_elem(h->free_slots,
1633                                                 cached_free_slots->objs,
1634                                                 sizeof(uint32_t),
1635                                                 LCORE_CACHE_SIZE, NULL);
1636                         RETURN_IF_TRUE((n_slots == 0), -EFAULT);
1637                         cached_free_slots->len -= n_slots;
1638                 }
1639                 /* Put index of new free slot in cache. */
1640                 cached_free_slots->objs[cached_free_slots->len] = key_idx;
1641                 cached_free_slots->len++;
1642         } else {
1643                 rte_ring_sp_enqueue_elem(h->free_slots, &key_idx,
1644                                                 sizeof(uint32_t));
1645         }
1646
1647         return 0;
1648 }
1649
1650 static inline void
1651 compare_signatures(uint32_t *prim_hash_matches, uint32_t *sec_hash_matches,
1652                         const struct rte_hash_bucket *prim_bkt,
1653                         const struct rte_hash_bucket *sec_bkt,
1654                         uint16_t sig,
1655                         enum rte_hash_sig_compare_function sig_cmp_fn)
1656 {
1657         unsigned int i;
1658
1659         /* For match mask the first bit of every two bits indicates the match */
1660         switch (sig_cmp_fn) {
1661 #if defined(RTE_MACHINE_CPUFLAG_SSE2)
1662         case RTE_HASH_COMPARE_SSE:
1663                 /* Compare all signatures in the bucket */
1664                 *prim_hash_matches = _mm_movemask_epi8(_mm_cmpeq_epi16(
1665                                 _mm_load_si128(
1666                                         (__m128i const *)prim_bkt->sig_current),
1667                                 _mm_set1_epi16(sig)));
1668                 /* Compare all signatures in the bucket */
1669                 *sec_hash_matches = _mm_movemask_epi8(_mm_cmpeq_epi16(
1670                                 _mm_load_si128(
1671                                         (__m128i const *)sec_bkt->sig_current),
1672                                 _mm_set1_epi16(sig)));
1673                 break;
1674 #elif defined(RTE_MACHINE_CPUFLAG_NEON)
1675         case RTE_HASH_COMPARE_NEON: {
1676                 uint16x8_t vmat, vsig, x;
1677                 int16x8_t shift = {-15, -13, -11, -9, -7, -5, -3, -1};
1678
1679                 vsig = vld1q_dup_u16((uint16_t const *)&sig);
1680                 /* Compare all signatures in the primary bucket */
1681                 vmat = vceqq_u16(vsig,
1682                         vld1q_u16((uint16_t const *)prim_bkt->sig_current));
1683                 x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
1684                 *prim_hash_matches = (uint32_t)(vaddvq_u16(x));
1685                 /* Compare all signatures in the secondary bucket */
1686                 vmat = vceqq_u16(vsig,
1687                         vld1q_u16((uint16_t const *)sec_bkt->sig_current));
1688                 x = vshlq_u16(vandq_u16(vmat, vdupq_n_u16(0x8000)), shift);
1689                 *sec_hash_matches = (uint32_t)(vaddvq_u16(x));
1690                 }
1691                 break;
1692 #endif
1693         default:
1694                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
1695                         *prim_hash_matches |=
1696                                 ((sig == prim_bkt->sig_current[i]) << (i << 1));
1697                         *sec_hash_matches |=
1698                                 ((sig == sec_bkt->sig_current[i]) << (i << 1));
1699                 }
1700         }
1701 }
1702
1703 #define PREFETCH_OFFSET 4
1704 static inline void
1705 __rte_hash_lookup_bulk_l(const struct rte_hash *h, const void **keys,
1706                         int32_t num_keys, int32_t *positions,
1707                         uint64_t *hit_mask, void *data[])
1708 {
1709         uint64_t hits = 0;
1710         int32_t i;
1711         int32_t ret;
1712         uint32_t prim_hash[RTE_HASH_LOOKUP_BULK_MAX];
1713         uint32_t prim_index[RTE_HASH_LOOKUP_BULK_MAX];
1714         uint32_t sec_index[RTE_HASH_LOOKUP_BULK_MAX];
1715         uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
1716         const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1717         const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1718         uint32_t prim_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1719         uint32_t sec_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1720         struct rte_hash_bucket *cur_bkt, *next_bkt;
1721
1722         /* Prefetch first keys */
1723         for (i = 0; i < PREFETCH_OFFSET && i < num_keys; i++)
1724                 rte_prefetch0(keys[i]);
1725
1726         /*
1727          * Prefetch rest of the keys, calculate primary and
1728          * secondary bucket and prefetch them
1729          */
1730         for (i = 0; i < (num_keys - PREFETCH_OFFSET); i++) {
1731                 rte_prefetch0(keys[i + PREFETCH_OFFSET]);
1732
1733                 prim_hash[i] = rte_hash_hash(h, keys[i]);
1734
1735                 sig[i] = get_short_sig(prim_hash[i]);
1736                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1737                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1738
1739                 primary_bkt[i] = &h->buckets[prim_index[i]];
1740                 secondary_bkt[i] = &h->buckets[sec_index[i]];
1741
1742                 rte_prefetch0(primary_bkt[i]);
1743                 rte_prefetch0(secondary_bkt[i]);
1744         }
1745
1746         /* Calculate and prefetch rest of the buckets */
1747         for (; i < num_keys; i++) {
1748                 prim_hash[i] = rte_hash_hash(h, keys[i]);
1749
1750                 sig[i] = get_short_sig(prim_hash[i]);
1751                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1752                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1753
1754                 primary_bkt[i] = &h->buckets[prim_index[i]];
1755                 secondary_bkt[i] = &h->buckets[sec_index[i]];
1756
1757                 rte_prefetch0(primary_bkt[i]);
1758                 rte_prefetch0(secondary_bkt[i]);
1759         }
1760
1761         __hash_rw_reader_lock(h);
1762
1763         /* Compare signatures and prefetch key slot of first hit */
1764         for (i = 0; i < num_keys; i++) {
1765                 compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
1766                         primary_bkt[i], secondary_bkt[i],
1767                         sig[i], h->sig_cmp_fn);
1768
1769                 if (prim_hitmask[i]) {
1770                         uint32_t first_hit =
1771                                         __builtin_ctzl(prim_hitmask[i])
1772                                         >> 1;
1773                         uint32_t key_idx =
1774                                 primary_bkt[i]->key_idx[first_hit];
1775                         const struct rte_hash_key *key_slot =
1776                                 (const struct rte_hash_key *)(
1777                                 (const char *)h->key_store +
1778                                 key_idx * h->key_entry_size);
1779                         rte_prefetch0(key_slot);
1780                         continue;
1781                 }
1782
1783                 if (sec_hitmask[i]) {
1784                         uint32_t first_hit =
1785                                         __builtin_ctzl(sec_hitmask[i])
1786                                         >> 1;
1787                         uint32_t key_idx =
1788                                 secondary_bkt[i]->key_idx[first_hit];
1789                         const struct rte_hash_key *key_slot =
1790                                 (const struct rte_hash_key *)(
1791                                 (const char *)h->key_store +
1792                                 key_idx * h->key_entry_size);
1793                         rte_prefetch0(key_slot);
1794                 }
1795         }
1796
1797         /* Compare keys, first hits in primary first */
1798         for (i = 0; i < num_keys; i++) {
1799                 positions[i] = -ENOENT;
1800                 while (prim_hitmask[i]) {
1801                         uint32_t hit_index =
1802                                         __builtin_ctzl(prim_hitmask[i])
1803                                         >> 1;
1804                         uint32_t key_idx =
1805                                 primary_bkt[i]->key_idx[hit_index];
1806                         const struct rte_hash_key *key_slot =
1807                                 (const struct rte_hash_key *)(
1808                                 (const char *)h->key_store +
1809                                 key_idx * h->key_entry_size);
1810
1811                         /*
1812                          * If key index is 0, do not compare key,
1813                          * as it is checking the dummy slot
1814                          */
1815                         if (!!key_idx &
1816                                 !rte_hash_cmp_eq(
1817                                         key_slot->key, keys[i], h)) {
1818                                 if (data != NULL)
1819                                         data[i] = key_slot->pdata;
1820
1821                                 hits |= 1ULL << i;
1822                                 positions[i] = key_idx - 1;
1823                                 goto next_key;
1824                         }
1825                         prim_hitmask[i] &= ~(3ULL << (hit_index << 1));
1826                 }
1827
1828                 while (sec_hitmask[i]) {
1829                         uint32_t hit_index =
1830                                         __builtin_ctzl(sec_hitmask[i])
1831                                         >> 1;
1832                         uint32_t key_idx =
1833                                 secondary_bkt[i]->key_idx[hit_index];
1834                         const struct rte_hash_key *key_slot =
1835                                 (const struct rte_hash_key *)(
1836                                 (const char *)h->key_store +
1837                                 key_idx * h->key_entry_size);
1838
1839                         /*
1840                          * If key index is 0, do not compare key,
1841                          * as it is checking the dummy slot
1842                          */
1843
1844                         if (!!key_idx &
1845                                 !rte_hash_cmp_eq(
1846                                         key_slot->key, keys[i], h)) {
1847                                 if (data != NULL)
1848                                         data[i] = key_slot->pdata;
1849
1850                                 hits |= 1ULL << i;
1851                                 positions[i] = key_idx - 1;
1852                                 goto next_key;
1853                         }
1854                         sec_hitmask[i] &= ~(3ULL << (hit_index << 1));
1855                 }
1856 next_key:
1857                 continue;
1858         }
1859
1860         /* all found, do not need to go through ext bkt */
1861         if ((hits == ((1ULL << num_keys) - 1)) || !h->ext_table_support) {
1862                 if (hit_mask != NULL)
1863                         *hit_mask = hits;
1864                 __hash_rw_reader_unlock(h);
1865                 return;
1866         }
1867
1868         /* need to check ext buckets for match */
1869         for (i = 0; i < num_keys; i++) {
1870                 if ((hits & (1ULL << i)) != 0)
1871                         continue;
1872                 next_bkt = secondary_bkt[i]->next;
1873                 FOR_EACH_BUCKET(cur_bkt, next_bkt) {
1874                         if (data != NULL)
1875                                 ret = search_one_bucket_l(h, keys[i],
1876                                                 sig[i], &data[i], cur_bkt);
1877                         else
1878                                 ret = search_one_bucket_l(h, keys[i],
1879                                                 sig[i], NULL, cur_bkt);
1880                         if (ret != -1) {
1881                                 positions[i] = ret;
1882                                 hits |= 1ULL << i;
1883                                 break;
1884                         }
1885                 }
1886         }
1887
1888         __hash_rw_reader_unlock(h);
1889
1890         if (hit_mask != NULL)
1891                 *hit_mask = hits;
1892 }
1893
1894 static inline void
1895 __rte_hash_lookup_bulk_lf(const struct rte_hash *h, const void **keys,
1896                         int32_t num_keys, int32_t *positions,
1897                         uint64_t *hit_mask, void *data[])
1898 {
1899         uint64_t hits = 0;
1900         int32_t i;
1901         int32_t ret;
1902         uint32_t prim_hash[RTE_HASH_LOOKUP_BULK_MAX];
1903         uint32_t prim_index[RTE_HASH_LOOKUP_BULK_MAX];
1904         uint32_t sec_index[RTE_HASH_LOOKUP_BULK_MAX];
1905         uint16_t sig[RTE_HASH_LOOKUP_BULK_MAX];
1906         const struct rte_hash_bucket *primary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1907         const struct rte_hash_bucket *secondary_bkt[RTE_HASH_LOOKUP_BULK_MAX];
1908         uint32_t prim_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1909         uint32_t sec_hitmask[RTE_HASH_LOOKUP_BULK_MAX] = {0};
1910         struct rte_hash_bucket *cur_bkt, *next_bkt;
1911         uint32_t cnt_b, cnt_a;
1912
1913         /* Prefetch first keys */
1914         for (i = 0; i < PREFETCH_OFFSET && i < num_keys; i++)
1915                 rte_prefetch0(keys[i]);
1916
1917         /*
1918          * Prefetch rest of the keys, calculate primary and
1919          * secondary bucket and prefetch them
1920          */
1921         for (i = 0; i < (num_keys - PREFETCH_OFFSET); i++) {
1922                 rte_prefetch0(keys[i + PREFETCH_OFFSET]);
1923
1924                 prim_hash[i] = rte_hash_hash(h, keys[i]);
1925
1926                 sig[i] = get_short_sig(prim_hash[i]);
1927                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1928                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1929
1930                 primary_bkt[i] = &h->buckets[prim_index[i]];
1931                 secondary_bkt[i] = &h->buckets[sec_index[i]];
1932
1933                 rte_prefetch0(primary_bkt[i]);
1934                 rte_prefetch0(secondary_bkt[i]);
1935         }
1936
1937         /* Calculate and prefetch rest of the buckets */
1938         for (; i < num_keys; i++) {
1939                 prim_hash[i] = rte_hash_hash(h, keys[i]);
1940
1941                 sig[i] = get_short_sig(prim_hash[i]);
1942                 prim_index[i] = get_prim_bucket_index(h, prim_hash[i]);
1943                 sec_index[i] = get_alt_bucket_index(h, prim_index[i], sig[i]);
1944
1945                 primary_bkt[i] = &h->buckets[prim_index[i]];
1946                 secondary_bkt[i] = &h->buckets[sec_index[i]];
1947
1948                 rte_prefetch0(primary_bkt[i]);
1949                 rte_prefetch0(secondary_bkt[i]);
1950         }
1951
1952         for (i = 0; i < num_keys; i++)
1953                 positions[i] = -ENOENT;
1954
1955         do {
1956                 /* Load the table change counter before the lookup
1957                  * starts. Acquire semantics will make sure that
1958                  * loads in compare_signatures are not hoisted.
1959                  */
1960                 cnt_b = __atomic_load_n(h->tbl_chng_cnt,
1961                                         __ATOMIC_ACQUIRE);
1962
1963                 /* Compare signatures and prefetch key slot of first hit */
1964                 for (i = 0; i < num_keys; i++) {
1965                         compare_signatures(&prim_hitmask[i], &sec_hitmask[i],
1966                                 primary_bkt[i], secondary_bkt[i],
1967                                 sig[i], h->sig_cmp_fn);
1968
1969                         if (prim_hitmask[i]) {
1970                                 uint32_t first_hit =
1971                                                 __builtin_ctzl(prim_hitmask[i])
1972                                                 >> 1;
1973                                 uint32_t key_idx =
1974                                         primary_bkt[i]->key_idx[first_hit];
1975                                 const struct rte_hash_key *key_slot =
1976                                         (const struct rte_hash_key *)(
1977                                         (const char *)h->key_store +
1978                                         key_idx * h->key_entry_size);
1979                                 rte_prefetch0(key_slot);
1980                                 continue;
1981                         }
1982
1983                         if (sec_hitmask[i]) {
1984                                 uint32_t first_hit =
1985                                                 __builtin_ctzl(sec_hitmask[i])
1986                                                 >> 1;
1987                                 uint32_t key_idx =
1988                                         secondary_bkt[i]->key_idx[first_hit];
1989                                 const struct rte_hash_key *key_slot =
1990                                         (const struct rte_hash_key *)(
1991                                         (const char *)h->key_store +
1992                                         key_idx * h->key_entry_size);
1993                                 rte_prefetch0(key_slot);
1994                         }
1995                 }
1996
1997                 /* Compare keys, first hits in primary first */
1998                 for (i = 0; i < num_keys; i++) {
1999                         while (prim_hitmask[i]) {
2000                                 uint32_t hit_index =
2001                                                 __builtin_ctzl(prim_hitmask[i])
2002                                                 >> 1;
2003                                 uint32_t key_idx =
2004                                 __atomic_load_n(
2005                                         &primary_bkt[i]->key_idx[hit_index],
2006                                         __ATOMIC_ACQUIRE);
2007                                 const struct rte_hash_key *key_slot =
2008                                         (const struct rte_hash_key *)(
2009                                         (const char *)h->key_store +
2010                                         key_idx * h->key_entry_size);
2011
2012                                 /*
2013                                  * If key index is 0, do not compare key,
2014                                  * as it is checking the dummy slot
2015                                  */
2016                                 if (!!key_idx &
2017                                         !rte_hash_cmp_eq(
2018                                                 key_slot->key, keys[i], h)) {
2019                                         if (data != NULL)
2020                                                 data[i] = __atomic_load_n(
2021                                                         &key_slot->pdata,
2022                                                         __ATOMIC_ACQUIRE);
2023
2024                                         hits |= 1ULL << i;
2025                                         positions[i] = key_idx - 1;
2026                                         goto next_key;
2027                                 }
2028                                 prim_hitmask[i] &= ~(3ULL << (hit_index << 1));
2029                         }
2030
2031                         while (sec_hitmask[i]) {
2032                                 uint32_t hit_index =
2033                                                 __builtin_ctzl(sec_hitmask[i])
2034                                                 >> 1;
2035                                 uint32_t key_idx =
2036                                 __atomic_load_n(
2037                                         &secondary_bkt[i]->key_idx[hit_index],
2038                                         __ATOMIC_ACQUIRE);
2039                                 const struct rte_hash_key *key_slot =
2040                                         (const struct rte_hash_key *)(
2041                                         (const char *)h->key_store +
2042                                         key_idx * h->key_entry_size);
2043
2044                                 /*
2045                                  * If key index is 0, do not compare key,
2046                                  * as it is checking the dummy slot
2047                                  */
2048
2049                                 if (!!key_idx &
2050                                         !rte_hash_cmp_eq(
2051                                                 key_slot->key, keys[i], h)) {
2052                                         if (data != NULL)
2053                                                 data[i] = __atomic_load_n(
2054                                                         &key_slot->pdata,
2055                                                         __ATOMIC_ACQUIRE);
2056
2057                                         hits |= 1ULL << i;
2058                                         positions[i] = key_idx - 1;
2059                                         goto next_key;
2060                                 }
2061                                 sec_hitmask[i] &= ~(3ULL << (hit_index << 1));
2062                         }
2063 next_key:
2064                         continue;
2065                 }
2066
2067                 /* all found, do not need to go through ext bkt */
2068                 if (hits == ((1ULL << num_keys) - 1)) {
2069                         if (hit_mask != NULL)
2070                                 *hit_mask = hits;
2071                         return;
2072                 }
2073                 /* need to check ext buckets for match */
2074                 if (h->ext_table_support) {
2075                         for (i = 0; i < num_keys; i++) {
2076                                 if ((hits & (1ULL << i)) != 0)
2077                                         continue;
2078                                 next_bkt = secondary_bkt[i]->next;
2079                                 FOR_EACH_BUCKET(cur_bkt, next_bkt) {
2080                                         if (data != NULL)
2081                                                 ret = search_one_bucket_lf(h,
2082                                                         keys[i], sig[i],
2083                                                         &data[i], cur_bkt);
2084                                         else
2085                                                 ret = search_one_bucket_lf(h,
2086                                                                 keys[i], sig[i],
2087                                                                 NULL, cur_bkt);
2088                                         if (ret != -1) {
2089                                                 positions[i] = ret;
2090                                                 hits |= 1ULL << i;
2091                                                 break;
2092                                         }
2093                                 }
2094                         }
2095                 }
2096                 /* The loads of sig_current in compare_signatures
2097                  * should not move below the load from tbl_chng_cnt.
2098                  */
2099                 __atomic_thread_fence(__ATOMIC_ACQUIRE);
2100                 /* Re-read the table change counter to check if the
2101                  * table has changed during search. If yes, re-do
2102                  * the search.
2103                  * This load should not get hoisted. The load
2104                  * acquires on cnt_b, primary key index and secondary
2105                  * key index will make sure that it does not get
2106                  * hoisted.
2107                  */
2108                 cnt_a = __atomic_load_n(h->tbl_chng_cnt,
2109                                         __ATOMIC_ACQUIRE);
2110         } while (cnt_b != cnt_a);
2111
2112         if (hit_mask != NULL)
2113                 *hit_mask = hits;
2114 }
2115
2116 static inline void
2117 __rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
2118                         int32_t num_keys, int32_t *positions,
2119                         uint64_t *hit_mask, void *data[])
2120 {
2121         if (h->readwrite_concur_lf_support)
2122                 __rte_hash_lookup_bulk_lf(h, keys, num_keys, positions,
2123                                           hit_mask, data);
2124         else
2125                 __rte_hash_lookup_bulk_l(h, keys, num_keys, positions,
2126                                          hit_mask, data);
2127 }
2128
2129 int
2130 rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
2131                       uint32_t num_keys, int32_t *positions)
2132 {
2133         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
2134                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2135                         (positions == NULL)), -EINVAL);
2136
2137         __rte_hash_lookup_bulk(h, keys, num_keys, positions, NULL, NULL);
2138         return 0;
2139 }
2140
2141 int
2142 rte_hash_lookup_bulk_data(const struct rte_hash *h, const void **keys,
2143                       uint32_t num_keys, uint64_t *hit_mask, void *data[])
2144 {
2145         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
2146                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
2147                         (hit_mask == NULL)), -EINVAL);
2148
2149         int32_t positions[num_keys];
2150
2151         __rte_hash_lookup_bulk(h, keys, num_keys, positions, hit_mask, data);
2152
2153         /* Return number of hits */
2154         return __builtin_popcountl(*hit_mask);
2155 }
2156
2157 int32_t
2158 rte_hash_iterate(const struct rte_hash *h, const void **key, void **data, uint32_t *next)
2159 {
2160         uint32_t bucket_idx, idx, position;
2161         struct rte_hash_key *next_key;
2162
2163         RETURN_IF_TRUE(((h == NULL) || (next == NULL)), -EINVAL);
2164
2165         const uint32_t total_entries_main = h->num_buckets *
2166                                                         RTE_HASH_BUCKET_ENTRIES;
2167         const uint32_t total_entries = total_entries_main << 1;
2168
2169         /* Out of bounds of all buckets (both main table and ext table) */
2170         if (*next >= total_entries_main)
2171                 goto extend_table;
2172
2173         /* Calculate bucket and index of current iterator */
2174         bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
2175         idx = *next % RTE_HASH_BUCKET_ENTRIES;
2176
2177         /* If current position is empty, go to the next one */
2178         while ((position = __atomic_load_n(&h->buckets[bucket_idx].key_idx[idx],
2179                                         __ATOMIC_ACQUIRE)) == EMPTY_SLOT) {
2180                 (*next)++;
2181                 /* End of table */
2182                 if (*next == total_entries_main)
2183                         goto extend_table;
2184                 bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
2185                 idx = *next % RTE_HASH_BUCKET_ENTRIES;
2186         }
2187
2188         __hash_rw_reader_lock(h);
2189         next_key = (struct rte_hash_key *) ((char *)h->key_store +
2190                                 position * h->key_entry_size);
2191         /* Return key and data */
2192         *key = next_key->key;
2193         *data = next_key->pdata;
2194
2195         __hash_rw_reader_unlock(h);
2196
2197         /* Increment iterator */
2198         (*next)++;
2199
2200         return position - 1;
2201
2202 /* Begin to iterate extendable buckets */
2203 extend_table:
2204         /* Out of total bound or if ext bucket feature is not enabled */
2205         if (*next >= total_entries || !h->ext_table_support)
2206                 return -ENOENT;
2207
2208         bucket_idx = (*next - total_entries_main) / RTE_HASH_BUCKET_ENTRIES;
2209         idx = (*next - total_entries_main) % RTE_HASH_BUCKET_ENTRIES;
2210
2211         while ((position = h->buckets_ext[bucket_idx].key_idx[idx]) == EMPTY_SLOT) {
2212                 (*next)++;
2213                 if (*next == total_entries)
2214                         return -ENOENT;
2215                 bucket_idx = (*next - total_entries_main) /
2216                                                 RTE_HASH_BUCKET_ENTRIES;
2217                 idx = (*next - total_entries_main) % RTE_HASH_BUCKET_ENTRIES;
2218         }
2219         __hash_rw_reader_lock(h);
2220         next_key = (struct rte_hash_key *) ((char *)h->key_store +
2221                                 position * h->key_entry_size);
2222         /* Return key and data */
2223         *key = next_key->key;
2224         *data = next_key->pdata;
2225
2226         __hash_rw_reader_unlock(h);
2227
2228         /* Increment iterator */
2229         (*next)++;
2230         return position - 1;
2231 }