lib/librte_hash/rte_cuckoo_hash.c

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