lib/librte_hash/rte_cuckoo_hash.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <string.h>
  35 #include <stdint.h>
  36 #include <errno.h>
  37 #include <stdio.h>
  38 #include <stdarg.h>
  39 #include <sys/queue.h>
  40
  41 #include <rte_common.h>
  42 #include <rte_memory.h>         /* for definition of RTE_CACHE_LINE_SIZE */
  43 #include <rte_log.h>
  44 #include <rte_memcpy.h>
  45 #include <rte_prefetch.h>
  46 #include <rte_branch_prediction.h>
  47 #include <rte_memzone.h>
  48 #include <rte_malloc.h>
  49 #include <rte_eal.h>
  50 #include <rte_eal_memconfig.h>
  51 #include <rte_per_lcore.h>
  52 #include <rte_errno.h>
  53 #include <rte_string_fns.h>
  54 #include <rte_cpuflags.h>
  55 #include <rte_log.h>
  56 #include <rte_rwlock.h>
  57 #include <rte_spinlock.h>
  58 #include <rte_ring.h>
  59 #include <rte_compat.h>
  60
  61 #include "rte_hash.h"
  62 #include "rte_cuckoo_hash.h"
  63
  64 #if defined(RTE_ARCH_X86)
  65 #include "rte_cuckoo_hash_x86.h"
  66 #endif
  67
  68 TAILQ_HEAD(rte_hash_list, rte_tailq_entry);
  69
  70 static struct rte_tailq_elem rte_hash_tailq = {
  71         .name = "RTE_HASH",
  72 };
  73 EAL_REGISTER_TAILQ(rte_hash_tailq)
  74
  75 struct rte_hash *
  76 rte_hash_find_existing(const char *name)
  77 {
  78         struct rte_hash *h = NULL;
  79         struct rte_tailq_entry *te;
  80         struct rte_hash_list *hash_list;
  81
  82         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
  83
  84         rte_rwlock_read_lock(RTE_EAL_TAILQ_RWLOCK);
  85         TAILQ_FOREACH(te, hash_list, next) {
  86                 h = (struct rte_hash *) te->data;
  87                 if (strncmp(name, h->name, RTE_HASH_NAMESIZE) == 0)
  88                         break;
  89         }
  90         rte_rwlock_read_unlock(RTE_EAL_TAILQ_RWLOCK);
  91
  92         if (te == NULL) {
  93                 rte_errno = ENOENT;
  94                 return NULL;
  95         }
  96         return h;
  97 }
  98
  99 void rte_hash_set_cmp_func(struct rte_hash *h, rte_hash_cmp_eq_t func)
 100 {
 101         h->cmp_jump_table_idx = KEY_CUSTOM;
 102         h->rte_hash_custom_cmp_eq = func;
 103 }
 104
 105 static inline int
 106 rte_hash_cmp_eq(const void *key1, const void *key2, const struct rte_hash *h)
 107 {
 108         if (h->cmp_jump_table_idx == KEY_CUSTOM)
 109                 return h->rte_hash_custom_cmp_eq(key1, key2, h->key_len);
 110         else
 111                 return cmp_jump_table[h->cmp_jump_table_idx](key1, key2, h->key_len);
 112 }
 113
 114 struct rte_hash *
 115 rte_hash_create(const struct rte_hash_parameters *params)
 116 {
 117         struct rte_hash *h = NULL;
 118         struct rte_tailq_entry *te = NULL;
 119         struct rte_hash_list *hash_list;
 120         struct rte_ring *r = NULL;
 121         char hash_name[RTE_HASH_NAMESIZE];
 122         void *k = NULL;
 123         void *buckets = NULL;
 124         char ring_name[RTE_RING_NAMESIZE];
 125         unsigned num_key_slots;
 126         unsigned hw_trans_mem_support = 0;
 127         unsigned i;
 128
 129         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
 130
 131         if (params == NULL) {
 132                 RTE_LOG(ERR, HASH, "rte_hash_create has no parameters\n");
 133                 return NULL;
 134         }
 135
 136         /* Check for valid parameters */
 137         if ((params->entries > RTE_HASH_ENTRIES_MAX) ||
 138                         (params->entries < RTE_HASH_BUCKET_ENTRIES) ||
 139                         !rte_is_power_of_2(RTE_HASH_BUCKET_ENTRIES) ||
 140                         (params->key_len == 0)) {
 141                 rte_errno = EINVAL;
 142                 RTE_LOG(ERR, HASH, "rte_hash_create has invalid parameters\n");
 143                 return NULL;
 144         }
 145
 146         /* Check extra flags field to check extra options. */
 147         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_TRANS_MEM_SUPPORT)
 148                 hw_trans_mem_support = 1;
 149
 150         /* Store all keys and leave the first entry as a dummy entry for lookup_bulk */
 151         if (hw_trans_mem_support)
 152                 /*
 153                  * Increase number of slots by total number of indices
 154                  * that can be stored in the lcore caches
 155                  * except for the first cache
 156                  */
 157                 num_key_slots = params->entries + (RTE_MAX_LCORE - 1) *
 158                                         LCORE_CACHE_SIZE + 1;
 159         else
 160                 num_key_slots = params->entries + 1;
 161
 162         snprintf(ring_name, sizeof(ring_name), "HT_%s", params->name);
 163         /* Create ring (Dummy slot index is not enqueued) */
 164         r = rte_ring_create(ring_name, rte_align32pow2(num_key_slots - 1),
 165                         params->socket_id, 0);
 166         if (r == NULL) {
 167                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
 168                 goto err;
 169         }
 170
 171         snprintf(hash_name, sizeof(hash_name), "HT_%s", params->name);
 172
 173         rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
 174
 175         /* guarantee there's no existing: this is normally already checked
 176          * by ring creation above */
 177         TAILQ_FOREACH(te, hash_list, next) {
 178                 h = (struct rte_hash *) te->data;
 179                 if (strncmp(params->name, h->name, RTE_HASH_NAMESIZE) == 0)
 180                         break;
 181         }
 182         h = NULL;
 183         if (te != NULL) {
 184                 rte_errno = EEXIST;
 185                 te = NULL;
 186                 goto err_unlock;
 187         }
 188
 189         te = rte_zmalloc("HASH_TAILQ_ENTRY", sizeof(*te), 0);
 190         if (te == NULL) {
 191                 RTE_LOG(ERR, HASH, "tailq entry allocation failed\n");
 192                 goto err_unlock;
 193         }
 194
 195         h = (struct rte_hash *)rte_zmalloc_socket(hash_name, sizeof(struct rte_hash),
 196                                         RTE_CACHE_LINE_SIZE, params->socket_id);
 197
 198         if (h == NULL) {
 199                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
 200                 goto err_unlock;
 201         }
 202
 203         const uint32_t num_buckets = rte_align32pow2(params->entries)
 204                                         / RTE_HASH_BUCKET_ENTRIES;
 205
 206         buckets = rte_zmalloc_socket(NULL,
 207                                 num_buckets * sizeof(struct rte_hash_bucket),
 208                                 RTE_CACHE_LINE_SIZE, params->socket_id);
 209
 210         if (buckets == NULL) {
 211                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
 212                 goto err_unlock;
 213         }
 214
 215         const uint32_t key_entry_size = sizeof(struct rte_hash_key) + params->key_len;
 216         const uint64_t key_tbl_size = (uint64_t) key_entry_size * num_key_slots;
 217
 218         k = rte_zmalloc_socket(NULL, key_tbl_size,
 219                         RTE_CACHE_LINE_SIZE, params->socket_id);
 220
 221         if (k == NULL) {
 222                 RTE_LOG(ERR, HASH, "memory allocation failed\n");
 223                 goto err_unlock;
 224         }
 225
 226 /*
 227  * If x86 architecture is used, select appropriate compare function,
 228  * which may use x86 intrinsics, otherwise use memcmp
 229  */
 230 #if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64)
 231         /* Select function to compare keys */
 232         switch (params->key_len) {
 233         case 16:
 234                 h->cmp_jump_table_idx = KEY_16_BYTES;
 235                 break;
 236         case 32:
 237                 h->cmp_jump_table_idx = KEY_32_BYTES;
 238                 break;
 239         case 48:
 240                 h->cmp_jump_table_idx = KEY_48_BYTES;
 241                 break;
 242         case 64:
 243                 h->cmp_jump_table_idx = KEY_64_BYTES;
 244                 break;
 245         case 80:
 246                 h->cmp_jump_table_idx = KEY_80_BYTES;
 247                 break;
 248         case 96:
 249                 h->cmp_jump_table_idx = KEY_96_BYTES;
 250                 break;
 251         case 112:
 252                 h->cmp_jump_table_idx = KEY_112_BYTES;
 253                 break;
 254         case 128:
 255                 h->cmp_jump_table_idx = KEY_128_BYTES;
 256                 break;
 257         default:
 258                 /* If key is not multiple of 16, use generic memcmp */
 259                 h->cmp_jump_table_idx = KEY_OTHER_BYTES;
 260         }
 261 #else
 262         h->cmp_jump_table_idx = KEY_OTHER_BYTES;
 263 #endif
 264
 265         if (hw_trans_mem_support) {
 266                 h->local_free_slots = rte_zmalloc_socket(NULL,
 267                                 sizeof(struct lcore_cache) * RTE_MAX_LCORE,
 268                                 RTE_CACHE_LINE_SIZE, params->socket_id);
 269         }
 270
 271         /* Setup hash context */
 272         snprintf(h->name, sizeof(h->name), "%s", params->name);
 273         h->entries = params->entries;
 274         h->key_len = params->key_len;
 275         h->key_entry_size = key_entry_size;
 276         h->hash_func_init_val = params->hash_func_init_val;
 277
 278         h->num_buckets = num_buckets;
 279         h->bucket_bitmask = h->num_buckets - 1;
 280         h->buckets = buckets;
 281         h->hash_func = (params->hash_func == NULL) ?
 282                 DEFAULT_HASH_FUNC : params->hash_func;
 283         h->key_store = k;
 284         h->free_slots = r;
 285         h->hw_trans_mem_support = hw_trans_mem_support;
 286
 287 #if defined(RTE_ARCH_X86)
 288         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
 289                 h->sig_cmp_fn = RTE_HASH_COMPARE_AVX2;
 290         else if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_SSE2))
 291                 h->sig_cmp_fn = RTE_HASH_COMPARE_SSE;
 292         else
 293 #endif
 294                 h->sig_cmp_fn = RTE_HASH_COMPARE_SCALAR;
 295
 296         /* Turn on multi-writer only with explicit flat from user and TM
 297          * support.
 298          */
 299         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) {
 300                 if (h->hw_trans_mem_support) {
 301                         h->add_key = ADD_KEY_MULTIWRITER_TM;
 302                 } else {
 303                         h->add_key = ADD_KEY_MULTIWRITER;
 304                         h->multiwriter_lock = rte_malloc(NULL,
 305                                                         sizeof(rte_spinlock_t),
 306                                                         LCORE_CACHE_SIZE);
 307                         rte_spinlock_init(h->multiwriter_lock);
 308                 }
 309         } else
 310                 h->add_key = ADD_KEY_SINGLEWRITER;
 311
 312         /* Populate free slots ring. Entry zero is reserved for key misses. */
 313         for (i = 1; i < params->entries + 1; i++)
 314                 rte_ring_sp_enqueue(r, (void *)((uintptr_t) i));
 315
 316         te->data = (void *) h;
 317         TAILQ_INSERT_TAIL(hash_list, te, next);
 318         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 319
 320         return h;
 321 err_unlock:
 322         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 323 err:
 324         rte_ring_free(r);
 325         rte_free(te);
 326         rte_free(h);
 327         rte_free(buckets);
 328         rte_free(k);
 329         return NULL;
 330 }
 331
 332 void
 333 rte_hash_free(struct rte_hash *h)
 334 {
 335         struct rte_tailq_entry *te;
 336         struct rte_hash_list *hash_list;
 337
 338         if (h == NULL)
 339                 return;
 340
 341         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
 342
 343         rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
 344
 345         /* find out tailq entry */
 346         TAILQ_FOREACH(te, hash_list, next) {
 347                 if (te->data == (void *) h)
 348                         break;
 349         }
 350
 351         if (te == NULL) {
 352                 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 353                 return;
 354         }
 355
 356         TAILQ_REMOVE(hash_list, te, next);
 357
 358         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 359
 360         if (h->hw_trans_mem_support)
 361                 rte_free(h->local_free_slots);
 362
 363         if (h->add_key == ADD_KEY_MULTIWRITER)
 364                 rte_free(h->multiwriter_lock);
 365         rte_ring_free(h->free_slots);
 366         rte_free(h->key_store);
 367         rte_free(h->buckets);
 368         rte_free(h);
 369         rte_free(te);
 370 }
 371
 372 hash_sig_t
 373 rte_hash_hash(const struct rte_hash *h, const void *key)
 374 {
 375         /* calc hash result by key */
 376         return h->hash_func(key, h->key_len, h->hash_func_init_val);
 377 }
 378
 379 /* Calc the secondary hash value from the primary hash value of a given key */
 380 static inline hash_sig_t
 381 rte_hash_secondary_hash(const hash_sig_t primary_hash)
 382 {
 383         static const unsigned all_bits_shift = 12;
 384         static const unsigned alt_bits_xor = 0x5bd1e995;
 385
 386         uint32_t tag = primary_hash >> all_bits_shift;
 387
 388         return primary_hash ^ ((tag + 1) * alt_bits_xor);
 389 }
 390
 391 void
 392 rte_hash_reset(struct rte_hash *h)
 393 {
 394         void *ptr;
 395         unsigned i;
 396
 397         if (h == NULL)
 398                 return;
 399
 400         memset(h->buckets, 0, h->num_buckets * sizeof(struct rte_hash_bucket));
 401         memset(h->key_store, 0, h->key_entry_size * (h->entries + 1));
 402
 403         /* clear the free ring */
 404         while (rte_ring_dequeue(h->free_slots, &ptr) == 0)
 405                 rte_pause();
 406
 407         /* Repopulate the free slots ring. Entry zero is reserved for key misses */
 408         for (i = 1; i < h->entries + 1; i++)
 409                 rte_ring_sp_enqueue(h->free_slots, (void *)((uintptr_t) i));
 410
 411         if (h->hw_trans_mem_support) {
 412                 /* Reset local caches per lcore */
 413                 for (i = 0; i < RTE_MAX_LCORE; i++)
 414                         h->local_free_slots[i].len = 0;
 415         }
 416 }
 417
 418 /* Search for an entry that can be pushed to its alternative location */
 419 static inline int
 420 make_space_bucket(const struct rte_hash *h, struct rte_hash_bucket *bkt)
 421 {
 422         unsigned i, j;
 423         int ret;
 424         uint32_t next_bucket_idx;
 425         struct rte_hash_bucket *next_bkt[RTE_HASH_BUCKET_ENTRIES];
 426
 427         /*
 428          * Push existing item (search for bucket with space in
 429          * alternative locations) to its alternative location
 430          */
 431         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 432                 /* Search for space in alternative locations */
 433                 next_bucket_idx = bkt->sig_alt[i] & h->bucket_bitmask;
 434                 next_bkt[i] = &h->buckets[next_bucket_idx];
 435                 for (j = 0; j < RTE_HASH_BUCKET_ENTRIES; j++) {
 436                         if (next_bkt[i]->key_idx[j] == EMPTY_SLOT)
 437                                 break;
 438                 }
 439
 440                 if (j != RTE_HASH_BUCKET_ENTRIES)
 441                         break;
 442         }
 443
 444         /* Alternative location has spare room (end of recursive function) */
 445         if (i != RTE_HASH_BUCKET_ENTRIES) {
 446                 next_bkt[i]->sig_alt[j] = bkt->sig_current[i];
 447                 next_bkt[i]->sig_current[j] = bkt->sig_alt[i];
 448                 next_bkt[i]->key_idx[j] = bkt->key_idx[i];
 449                 return i;
 450         }
 451
 452         /* Pick entry that has not been pushed yet */
 453         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++)
 454                 if (bkt->flag[i] == 0)
 455                         break;
 456
 457         /* All entries have been pushed, so entry cannot be added */
 458         if (i == RTE_HASH_BUCKET_ENTRIES)
 459                 return -ENOSPC;
 460
 461         /* Set flag to indicate that this entry is going to be pushed */
 462         bkt->flag[i] = 1;
 463         /* Need room in alternative bucket to insert the pushed entry */
 464         ret = make_space_bucket(h, next_bkt[i]);
 465         /*
 466          * After recursive function.
 467          * Clear flags and insert the pushed entry
 468          * in its alternative location if successful,
 469          * or return error
 470          */
 471         bkt->flag[i] = 0;
 472         if (ret >= 0) {
 473                 next_bkt[i]->sig_alt[ret] = bkt->sig_current[i];
 474                 next_bkt[i]->sig_current[ret] = bkt->sig_alt[i];
 475                 next_bkt[i]->key_idx[ret] = bkt->key_idx[i];
 476                 return i;
 477         } else
 478                 return ret;
 479
 480 }
 481
 482 /*
 483  * Function called to enqueue back an index in the cache/ring,
 484  * as slot has not being used and it can be used in the
 485  * next addition attempt.
 486  */
 487 static inline void
 488 enqueue_slot_back(const struct rte_hash *h,
 489                 struct lcore_cache *cached_free_slots,
 490                 void *slot_id)
 491 {
 492         if (h->hw_trans_mem_support) {
 493                 cached_free_slots->objs[cached_free_slots->len] = slot_id;
 494                 cached_free_slots->len++;
 495         } else
 496                 rte_ring_sp_enqueue(h->free_slots, slot_id);
 497 }
 498
 499 static inline int32_t
 500 __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
 501                                                 hash_sig_t sig, void *data)
 502 {
 503         hash_sig_t alt_hash;
 504         uint32_t prim_bucket_idx, sec_bucket_idx;
 505         unsigned i;
 506         struct rte_hash_bucket *prim_bkt, *sec_bkt;
 507         struct rte_hash_key *new_k, *k, *keys = h->key_store;
 508         void *slot_id = NULL;
 509         uint32_t new_idx;
 510         int ret;
 511         unsigned n_slots;
 512         unsigned lcore_id;
 513         struct lcore_cache *cached_free_slots = NULL;
 514
 515         if (h->add_key == ADD_KEY_MULTIWRITER)
 516                 rte_spinlock_lock(h->multiwriter_lock);
 517
 518         prim_bucket_idx = sig & h->bucket_bitmask;
 519         prim_bkt = &h->buckets[prim_bucket_idx];
 520         rte_prefetch0(prim_bkt);
 521
 522         alt_hash = rte_hash_secondary_hash(sig);
 523         sec_bucket_idx = alt_hash & h->bucket_bitmask;
 524         sec_bkt = &h->buckets[sec_bucket_idx];
 525         rte_prefetch0(sec_bkt);
 526
 527         /* Get a new slot for storing the new key */
 528         if (h->hw_trans_mem_support) {
 529                 lcore_id = rte_lcore_id();
 530                 cached_free_slots = &h->local_free_slots[lcore_id];
 531                 /* Try to get a free slot from the local cache */
 532                 if (cached_free_slots->len == 0) {
 533                         /* Need to get another burst of free slots from global ring */
 534                         n_slots = rte_ring_mc_dequeue_burst(h->free_slots,
 535                                         cached_free_slots->objs, LCORE_CACHE_SIZE);
 536                         if (n_slots == 0)
 537                                 return -ENOSPC;
 538
 539                         cached_free_slots->len += n_slots;
 540                 }
 541
 542                 /* Get a free slot from the local cache */
 543                 cached_free_slots->len--;
 544                 slot_id = cached_free_slots->objs[cached_free_slots->len];
 545         } else {
 546                 if (rte_ring_sc_dequeue(h->free_slots, &slot_id) != 0)
 547                         return -ENOSPC;
 548         }
 549
 550         new_k = RTE_PTR_ADD(keys, (uintptr_t)slot_id * h->key_entry_size);
 551         rte_prefetch0(new_k);
 552         new_idx = (uint32_t)((uintptr_t) slot_id);
 553
 554         /* Check if key is already inserted in primary location */
 555         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 556                 if (prim_bkt->sig_current[i] == sig &&
 557                                 prim_bkt->sig_alt[i] == alt_hash) {
 558                         k = (struct rte_hash_key *) ((char *)keys +
 559                                         prim_bkt->key_idx[i] * h->key_entry_size);
 560                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 561                                 /* Enqueue index of free slot back in the ring. */
 562                                 enqueue_slot_back(h, cached_free_slots, slot_id);
 563                                 /* Update data */
 564                                 k->pdata = data;
 565                                 /*
 566                                  * Return index where key is stored,
 567                                  * substracting the first dummy index
 568                                  */
 569                                 return prim_bkt->key_idx[i] - 1;
 570                         }
 571                 }
 572         }
 573
 574         /* Check if key is already inserted in secondary location */
 575         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 576                 if (sec_bkt->sig_alt[i] == sig &&
 577                                 sec_bkt->sig_current[i] == alt_hash) {
 578                         k = (struct rte_hash_key *) ((char *)keys +
 579                                         sec_bkt->key_idx[i] * h->key_entry_size);
 580                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 581                                 /* Enqueue index of free slot back in the ring. */
 582                                 enqueue_slot_back(h, cached_free_slots, slot_id);
 583                                 /* Update data */
 584                                 k->pdata = data;
 585                                 /*
 586                                  * Return index where key is stored,
 587                                  * substracting the first dummy index
 588                                  */
 589                                 return sec_bkt->key_idx[i] - 1;
 590                         }
 591                 }
 592         }
 593
 594         /* Copy key */
 595         rte_memcpy(new_k->key, key, h->key_len);
 596         new_k->pdata = data;
 597
 598 #if defined(RTE_ARCH_X86) /* currently only x86 support HTM */
 599         if (h->add_key == ADD_KEY_MULTIWRITER_TM) {
 600                 ret = rte_hash_cuckoo_insert_mw_tm(prim_bkt,
 601                                 sig, alt_hash, new_idx);
 602                 if (ret >= 0)
 603                         return new_idx - 1;
 604
 605                 /* Primary bucket full, need to make space for new entry */
 606                 ret = rte_hash_cuckoo_make_space_mw_tm(h, prim_bkt, sig,
 607                                                         alt_hash, new_idx);
 608
 609                 if (ret >= 0)
 610                         return new_idx - 1;
 611
 612                 /* Also search secondary bucket to get better occupancy */
 613                 ret = rte_hash_cuckoo_make_space_mw_tm(h, sec_bkt, sig,
 614                                                         alt_hash, new_idx);
 615
 616                 if (ret >= 0)
 617                         return new_idx - 1;
 618         } else {
 619 #endif
 620                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 621                         /* Check if slot is available */
 622                         if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
 623                                 prim_bkt->sig_current[i] = sig;
 624                                 prim_bkt->sig_alt[i] = alt_hash;
 625                                 prim_bkt->key_idx[i] = new_idx;
 626                                 break;
 627                         }
 628                 }
 629
 630                 if (i != RTE_HASH_BUCKET_ENTRIES) {
 631                         if (h->add_key == ADD_KEY_MULTIWRITER)
 632                                 rte_spinlock_unlock(h->multiwriter_lock);
 633                         return new_idx - 1;
 634                 }
 635
 636                 /* Primary bucket full, need to make space for new entry
 637                  * After recursive function.
 638                  * Insert the new entry in the position of the pushed entry
 639                  * if successful or return error and
 640                  * store the new slot back in the ring
 641                  */
 642                 ret = make_space_bucket(h, prim_bkt);
 643                 if (ret >= 0) {
 644                         prim_bkt->sig_current[ret] = sig;
 645                         prim_bkt->sig_alt[ret] = alt_hash;
 646                         prim_bkt->key_idx[ret] = new_idx;
 647                         if (h->add_key == ADD_KEY_MULTIWRITER)
 648                                 rte_spinlock_unlock(h->multiwriter_lock);
 649                         return new_idx - 1;
 650                 }
 651 #if defined(RTE_ARCH_X86)
 652         }
 653 #endif
 654         /* Error in addition, store new slot back in the ring and return error */
 655         enqueue_slot_back(h, cached_free_slots, (void *)((uintptr_t) new_idx));
 656
 657         if (h->add_key == ADD_KEY_MULTIWRITER)
 658                 rte_spinlock_unlock(h->multiwriter_lock);
 659         return ret;
 660 }
 661
 662 int32_t
 663 rte_hash_add_key_with_hash(const struct rte_hash *h,
 664                         const void *key, hash_sig_t sig)
 665 {
 666         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 667         return __rte_hash_add_key_with_hash(h, key, sig, 0);
 668 }
 669
 670 int32_t
 671 rte_hash_add_key(const struct rte_hash *h, const void *key)
 672 {
 673         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 674         return __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), 0);
 675 }
 676
 677 int
 678 rte_hash_add_key_with_hash_data(const struct rte_hash *h,
 679                         const void *key, hash_sig_t sig, void *data)
 680 {
 681         int ret;
 682
 683         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 684         ret = __rte_hash_add_key_with_hash(h, key, sig, data);
 685         if (ret >= 0)
 686                 return 0;
 687         else
 688                 return ret;
 689 }
 690
 691 int
 692 rte_hash_add_key_data(const struct rte_hash *h, const void *key, void *data)
 693 {
 694         int ret;
 695
 696         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 697
 698         ret = __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), data);
 699         if (ret >= 0)
 700                 return 0;
 701         else
 702                 return ret;
 703 }
 704 static inline int32_t
 705 __rte_hash_lookup_with_hash(const struct rte_hash *h, const void *key,
 706                                         hash_sig_t sig, void **data)
 707 {
 708         uint32_t bucket_idx;
 709         hash_sig_t alt_hash;
 710         unsigned i;
 711         struct rte_hash_bucket *bkt;
 712         struct rte_hash_key *k, *keys = h->key_store;
 713
 714         bucket_idx = sig & h->bucket_bitmask;
 715         bkt = &h->buckets[bucket_idx];
 716
 717         /* Check if key is in primary location */
 718         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 719                 if (bkt->sig_current[i] == sig &&
 720                                 bkt->key_idx[i] != EMPTY_SLOT) {
 721                         k = (struct rte_hash_key *) ((char *)keys +
 722                                         bkt->key_idx[i] * h->key_entry_size);
 723                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 724                                 if (data != NULL)
 725                                         *data = k->pdata;
 726                                 /*
 727                                  * Return index where key is stored,
 728                                  * substracting the first dummy index
 729                                  */
 730                                 return bkt->key_idx[i] - 1;
 731                         }
 732                 }
 733         }
 734
 735         /* Calculate secondary hash */
 736         alt_hash = rte_hash_secondary_hash(sig);
 737         bucket_idx = alt_hash & h->bucket_bitmask;
 738         bkt = &h->buckets[bucket_idx];
 739
 740         /* Check if key is in secondary location */
 741         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 742                 if (bkt->sig_current[i] == alt_hash &&
 743                                 bkt->sig_alt[i] == sig) {
 744                         k = (struct rte_hash_key *) ((char *)keys +
 745                                         bkt->key_idx[i] * h->key_entry_size);
 746                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 747                                 if (data != NULL)
 748                                         *data = k->pdata;
 749                                 /*
 750                                  * Return index where key is stored,
 751                                  * substracting the first dummy index
 752                                  */
 753                                 return bkt->key_idx[i] - 1;
 754                         }
 755                 }
 756         }
 757
 758         return -ENOENT;
 759 }
 760
 761 int32_t
 762 rte_hash_lookup_with_hash(const struct rte_hash *h,
 763                         const void *key, hash_sig_t sig)
 764 {
 765         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 766         return __rte_hash_lookup_with_hash(h, key, sig, NULL);
 767 }
 768
 769 int32_t
 770 rte_hash_lookup(const struct rte_hash *h, const void *key)
 771 {
 772         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 773         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), NULL);
 774 }
 775
 776 int
 777 rte_hash_lookup_with_hash_data(const struct rte_hash *h,
 778                         const void *key, hash_sig_t sig, void **data)
 779 {
 780         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 781         return __rte_hash_lookup_with_hash(h, key, sig, data);
 782 }
 783
 784 int
 785 rte_hash_lookup_data(const struct rte_hash *h, const void *key, void **data)
 786 {
 787         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 788         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), data);
 789 }
 790
 791 static inline void
 792 remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i)
 793 {
 794         unsigned lcore_id, n_slots;
 795         struct lcore_cache *cached_free_slots;
 796
 797         bkt->sig_current[i] = NULL_SIGNATURE;
 798         bkt->sig_alt[i] = NULL_SIGNATURE;
 799         if (h->hw_trans_mem_support) {
 800                 lcore_id = rte_lcore_id();
 801                 cached_free_slots = &h->local_free_slots[lcore_id];
 802                 /* Cache full, need to free it. */
 803                 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
 804                         /* Need to enqueue the free slots in global ring. */
 805                         n_slots = rte_ring_mp_enqueue_burst(h->free_slots,
 806                                                 cached_free_slots->objs,
 807                                                 LCORE_CACHE_SIZE);
 808                         cached_free_slots->len -= n_slots;
 809                 }
 810                 /* Put index of new free slot in cache. */
 811                 cached_free_slots->objs[cached_free_slots->len] =
 812                                 (void *)((uintptr_t)bkt->key_idx[i]);
 813                 cached_free_slots->len++;
 814         } else {
 815                 rte_ring_sp_enqueue(h->free_slots,
 816                                 (void *)((uintptr_t)bkt->key_idx[i]));
 817         }
 818 }
 819
 820 static inline int32_t
 821 __rte_hash_del_key_with_hash(const struct rte_hash *h, const void *key,
 822                                                 hash_sig_t sig)
 823 {
 824         uint32_t bucket_idx;
 825         hash_sig_t alt_hash;
 826         unsigned i;
 827         struct rte_hash_bucket *bkt;
 828         struct rte_hash_key *k, *keys = h->key_store;
 829         int32_t ret;
 830
 831         bucket_idx = sig & h->bucket_bitmask;
 832         bkt = &h->buckets[bucket_idx];
 833
 834         /* Check if key is in primary location */
 835         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 836                 if (bkt->sig_current[i] == sig &&
 837                                 bkt->key_idx[i] != EMPTY_SLOT) {
 838                         k = (struct rte_hash_key *) ((char *)keys +
 839                                         bkt->key_idx[i] * h->key_entry_size);
 840                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 841                                 remove_entry(h, bkt, i);
 842
 843                                 /*
 844                                  * Return index where key is stored,
 845                                  * substracting the first dummy index
 846                                  */
 847                                 ret = bkt->key_idx[i] - 1;
 848                                 bkt->key_idx[i] = EMPTY_SLOT;
 849                                 return ret;
 850                         }
 851                 }
 852         }
 853
 854         /* Calculate secondary hash */
 855         alt_hash = rte_hash_secondary_hash(sig);
 856         bucket_idx = alt_hash & h->bucket_bitmask;
 857         bkt = &h->buckets[bucket_idx];
 858
 859         /* Check if key is in secondary location */
 860         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 861                 if (bkt->sig_current[i] == alt_hash &&
 862                                 bkt->key_idx[i] != EMPTY_SLOT) {
 863                         k = (struct rte_hash_key *) ((char *)keys +
 864                                         bkt->key_idx[i] * h->key_entry_size);
 865                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 866                                 remove_entry(h, bkt, i);
 867
 868                                 /*
 869                                  * Return index where key is stored,
 870                                  * substracting the first dummy index
 871                                  */
 872                                 ret = bkt->key_idx[i] - 1;
 873                                 bkt->key_idx[i] = EMPTY_SLOT;
 874                                 return ret;
 875                         }
 876                 }
 877         }
 878
 879         return -ENOENT;
 880 }
 881
 882 int32_t
 883 rte_hash_del_key_with_hash(const struct rte_hash *h,
 884                         const void *key, hash_sig_t sig)
 885 {
 886         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 887         return __rte_hash_del_key_with_hash(h, key, sig);
 888 }
 889
 890 int32_t
 891 rte_hash_del_key(const struct rte_hash *h, const void *key)
 892 {
 893         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 894         return __rte_hash_del_key_with_hash(h, key, rte_hash_hash(h, key));
 895 }
 896
 897 int
 898 rte_hash_get_key_with_position(const struct rte_hash *h, const int32_t position,
 899                                void **key)
 900 {
 901         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 902
 903         struct rte_hash_key *k, *keys = h->key_store;
 904         k = (struct rte_hash_key *) ((char *) keys + (position + 1) *
 905                                      h->key_entry_size);
 906         *key = k->key;
 907
 908         if (position !=
 909             __rte_hash_lookup_with_hash(h, *key, rte_hash_hash(h, *key),
 910                                         NULL)) {
 911                 return -ENOENT;
 912         }
 913
 914         return 0;
 915 }
 916
 917 /* Lookup bulk stage 0: Prefetch input key */
 918 static inline void
 919 lookup_stage0(unsigned *idx, uint64_t *lookup_mask,
 920                 const void * const *keys)
 921 {
 922         *idx = __builtin_ctzl(*lookup_mask);
 923         if (*lookup_mask == 0)
 924                 *idx = 0;
 925
 926         rte_prefetch0(keys[*idx]);
 927         *lookup_mask &= ~(1llu << *idx);
 928 }
 929
 930 /*
 931  * Lookup bulk stage 1: Calculate primary/secondary hashes
 932  * and prefetch primary/secondary buckets
 933  */
 934 static inline void
 935 lookup_stage1(unsigned idx, hash_sig_t *prim_hash, hash_sig_t *sec_hash,
 936                 const struct rte_hash_bucket **primary_bkt,
 937                 const struct rte_hash_bucket **secondary_bkt,
 938                 hash_sig_t *hash_vals, const void * const *keys,
 939                 const struct rte_hash *h)
 940 {
 941         *prim_hash = rte_hash_hash(h, keys[idx]);
 942         hash_vals[idx] = *prim_hash;
 943         *sec_hash = rte_hash_secondary_hash(*prim_hash);
 944
 945         *primary_bkt = &h->buckets[*prim_hash & h->bucket_bitmask];
 946         *secondary_bkt = &h->buckets[*sec_hash & h->bucket_bitmask];
 947
 948         rte_prefetch0(*primary_bkt);
 949         rte_prefetch0(*secondary_bkt);
 950 }
 951
 952 static inline void
 953 compare_signatures(unsigned int *prim_hash_matches,
 954                         unsigned int *sec_hash_matches,
 955                         const struct rte_hash_bucket *prim_bkt,
 956                         const struct rte_hash_bucket *sec_bkt,
 957                         hash_sig_t prim_hash, hash_sig_t sec_hash,
 958                         enum rte_hash_sig_compare_function sig_cmp_fn)
 959 {
 960         unsigned int i;
 961
 962         switch (sig_cmp_fn) {
 963 #ifdef RTE_MACHINE_CPUFLAG_AVX2
 964         case RTE_HASH_COMPARE_AVX2:
 965                 *prim_hash_matches |= _mm256_movemask_ps((__m256)_mm256_cmpeq_epi32(
 966                                 _mm256_load_si256(
 967                                         (__m256i const *)prim_bkt->sig_current),
 968                                 _mm256_set1_epi32(prim_hash)));
 969                 *sec_hash_matches |= _mm256_movemask_ps((__m256)_mm256_cmpeq_epi32(
 970                                 _mm256_load_si256(
 971                                         (__m256i const *)sec_bkt->sig_current),
 972                                 _mm256_set1_epi32(sec_hash)));
 973                 break;
 974 #endif
 975 #ifdef RTE_MACHINE_CPUFLAG_SSE2
 976         case RTE_HASH_COMPARE_SSE:
 977                 /* Compare the first 4 signatures in the bucket */
 978                 *prim_hash_matches |= _mm_movemask_ps((__m128)_mm_cmpeq_epi16(
 979                                 _mm_load_si128(
 980                                         (__m128i const *)prim_bkt->sig_current),
 981                                 _mm_set1_epi32(prim_hash)));
 982                 *prim_hash_matches |= (_mm_movemask_ps((__m128)_mm_cmpeq_epi16(
 983                                 _mm_load_si128(
 984                                         (__m128i const *)&prim_bkt->sig_current[4]),
 985                                 _mm_set1_epi32(prim_hash)))) << 4;
 986                 /* Compare the first 4 signatures in the bucket */
 987                 *sec_hash_matches |= _mm_movemask_ps((__m128)_mm_cmpeq_epi16(
 988                                 _mm_load_si128(
 989                                         (__m128i const *)sec_bkt->sig_current),
 990                                 _mm_set1_epi32(sec_hash)));
 991                 *sec_hash_matches |= (_mm_movemask_ps((__m128)_mm_cmpeq_epi16(
 992                                 _mm_load_si128(
 993                                         (__m128i const *)&sec_bkt->sig_current[4]),
 994                                 _mm_set1_epi32(sec_hash)))) << 4;
 995                 break;
 996 #endif
 997         default:
 998                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 999                         *prim_hash_matches |=
1000                                 ((prim_hash == prim_bkt->sig_current[i]) << i);
1001                         *sec_hash_matches |=
1002                                 ((sec_hash == sec_bkt->sig_current[i]) << i);
1003                 }
1004         }
1005
1006 }
1007
1008 /*
1009  * Lookup bulk stage 2:  Search for match hashes in primary/secondary locations
1010  * and prefetch first key slot
1011  */
1012 static inline void
1013 lookup_stage2(unsigned idx, hash_sig_t prim_hash, hash_sig_t sec_hash,
1014                 const struct rte_hash_bucket *prim_bkt,
1015                 const struct rte_hash_bucket *sec_bkt,
1016                 const struct rte_hash_key **key_slot, int32_t *positions,
1017                 uint64_t *extra_hits_mask, const void *keys,
1018                 const struct rte_hash *h)
1019 {
1020         unsigned int prim_hash_matches, sec_hash_matches, key_idx;
1021         unsigned int total_hash_matches;
1022
1023         prim_hash_matches = 1 << RTE_HASH_BUCKET_ENTRIES;
1024         sec_hash_matches = 1 << RTE_HASH_BUCKET_ENTRIES;
1025
1026         compare_signatures(&prim_hash_matches, &sec_hash_matches, prim_bkt,
1027                                 sec_bkt, prim_hash, sec_hash, h->sig_cmp_fn);
1028
1029         key_idx = prim_bkt->key_idx[__builtin_ctzl(prim_hash_matches)];
1030         if (key_idx == 0)
1031                 key_idx = sec_bkt->key_idx[__builtin_ctzl(sec_hash_matches)];
1032
1033         total_hash_matches = (prim_hash_matches |
1034                                 (sec_hash_matches << (RTE_HASH_BUCKET_ENTRIES + 1)));
1035         *key_slot = (const struct rte_hash_key *) ((const char *)keys +
1036                                         key_idx * h->key_entry_size);
1037
1038         rte_prefetch0(*key_slot);
1039         /*
1040          * Return index where key is stored,
1041          * substracting the first dummy index
1042          */
1043         positions[idx] = (key_idx - 1);
1044
1045         *extra_hits_mask |= (uint64_t)(__builtin_popcount(total_hash_matches) > 3) << idx;
1046
1047 }
1048
1049
1050 /* Lookup bulk stage 3: Check if key matches, update hit mask and return data */
1051 static inline void
1052 lookup_stage3(unsigned idx, const struct rte_hash_key *key_slot, const void * const *keys,
1053                 const int32_t *positions, void *data[], uint64_t *hits,
1054                 const struct rte_hash *h)
1055 {
1056         unsigned hit;
1057         unsigned key_idx;
1058
1059         hit = !rte_hash_cmp_eq(key_slot->key, keys[idx], h);
1060         if (data != NULL)
1061                 data[idx] = key_slot->pdata;
1062
1063         key_idx = positions[idx] + 1;
1064         /*
1065          * If key index is 0, force hit to be 0, in case key to be looked up
1066          * is all zero (as in the dummy slot), which would result in a wrong hit
1067          */
1068         *hits |= (uint64_t)(hit && !!key_idx)  << idx;
1069 }
1070
1071 static inline void
1072 __rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
1073                         uint32_t num_keys, int32_t *positions,
1074                         uint64_t *hit_mask, void *data[])
1075 {
1076         uint64_t hits = 0;
1077         uint64_t extra_hits_mask = 0;
1078         uint64_t lookup_mask, miss_mask;
1079         unsigned idx;
1080         const void *key_store = h->key_store;
1081         int ret;
1082         hash_sig_t hash_vals[RTE_HASH_LOOKUP_BULK_MAX];
1083
1084         unsigned idx00, idx01, idx10, idx11, idx20, idx21, idx30, idx31;
1085         const struct rte_hash_bucket *primary_bkt10, *primary_bkt11;
1086         const struct rte_hash_bucket *secondary_bkt10, *secondary_bkt11;
1087         const struct rte_hash_bucket *primary_bkt20, *primary_bkt21;
1088         const struct rte_hash_bucket *secondary_bkt20, *secondary_bkt21;
1089         const struct rte_hash_key *k_slot20, *k_slot21, *k_slot30, *k_slot31;
1090         hash_sig_t primary_hash10, primary_hash11;
1091         hash_sig_t secondary_hash10, secondary_hash11;
1092         hash_sig_t primary_hash20, primary_hash21;
1093         hash_sig_t secondary_hash20, secondary_hash21;
1094
1095         lookup_mask = (uint64_t) -1 >> (64 - num_keys);
1096         miss_mask = lookup_mask;
1097
1098         lookup_stage0(&idx00, &lookup_mask, keys);
1099         lookup_stage0(&idx01, &lookup_mask, keys);
1100
1101         idx10 = idx00, idx11 = idx01;
1102
1103         lookup_stage0(&idx00, &lookup_mask, keys);
1104         lookup_stage0(&idx01, &lookup_mask, keys);
1105         lookup_stage1(idx10, &primary_hash10, &secondary_hash10,
1106                         &primary_bkt10, &secondary_bkt10, hash_vals, keys, h);
1107         lookup_stage1(idx11, &primary_hash11, &secondary_hash11,
1108                         &primary_bkt11, &secondary_bkt11, hash_vals, keys, h);
1109
1110         primary_bkt20 = primary_bkt10;
1111         primary_bkt21 = primary_bkt11;
1112         secondary_bkt20 = secondary_bkt10;
1113         secondary_bkt21 = secondary_bkt11;
1114         primary_hash20 = primary_hash10;
1115         primary_hash21 = primary_hash11;
1116         secondary_hash20 = secondary_hash10;
1117         secondary_hash21 = secondary_hash11;
1118         idx20 = idx10, idx21 = idx11;
1119         idx10 = idx00, idx11 = idx01;
1120
1121         lookup_stage0(&idx00, &lookup_mask, keys);
1122         lookup_stage0(&idx01, &lookup_mask, keys);
1123         lookup_stage1(idx10, &primary_hash10, &secondary_hash10,
1124                         &primary_bkt10, &secondary_bkt10, hash_vals, keys, h);
1125         lookup_stage1(idx11, &primary_hash11, &secondary_hash11,
1126                         &primary_bkt11, &secondary_bkt11, hash_vals, keys, h);
1127         lookup_stage2(idx20, primary_hash20, secondary_hash20, primary_bkt20,
1128                         secondary_bkt20, &k_slot20, positions, &extra_hits_mask,
1129                         key_store, h);
1130         lookup_stage2(idx21, primary_hash21, secondary_hash21, primary_bkt21,
1131                         secondary_bkt21, &k_slot21, positions, &extra_hits_mask,
1132                         key_store, h);
1133
1134         while (lookup_mask) {
1135                 k_slot30 = k_slot20, k_slot31 = k_slot21;
1136                 idx30 = idx20, idx31 = idx21;
1137                 primary_bkt20 = primary_bkt10;
1138                 primary_bkt21 = primary_bkt11;
1139                 secondary_bkt20 = secondary_bkt10;
1140                 secondary_bkt21 = secondary_bkt11;
1141                 primary_hash20 = primary_hash10;
1142                 primary_hash21 = primary_hash11;
1143                 secondary_hash20 = secondary_hash10;
1144                 secondary_hash21 = secondary_hash11;
1145                 idx20 = idx10, idx21 = idx11;
1146                 idx10 = idx00, idx11 = idx01;
1147
1148                 lookup_stage0(&idx00, &lookup_mask, keys);
1149                 lookup_stage0(&idx01, &lookup_mask, keys);
1150                 lookup_stage1(idx10, &primary_hash10, &secondary_hash10,
1151                         &primary_bkt10, &secondary_bkt10, hash_vals, keys, h);
1152                 lookup_stage1(idx11, &primary_hash11, &secondary_hash11,
1153                         &primary_bkt11, &secondary_bkt11, hash_vals, keys, h);
1154                 lookup_stage2(idx20, primary_hash20, secondary_hash20,
1155                         primary_bkt20, secondary_bkt20, &k_slot20, positions,
1156                         &extra_hits_mask, key_store, h);
1157                 lookup_stage2(idx21, primary_hash21, secondary_hash21,
1158                         primary_bkt21, secondary_bkt21, &k_slot21, positions,
1159                         &extra_hits_mask, key_store, h);
1160                 lookup_stage3(idx30, k_slot30, keys, positions, data, &hits, h);
1161                 lookup_stage3(idx31, k_slot31, keys, positions, data, &hits, h);
1162         }
1163
1164         k_slot30 = k_slot20, k_slot31 = k_slot21;
1165         idx30 = idx20, idx31 = idx21;
1166         primary_bkt20 = primary_bkt10;
1167         primary_bkt21 = primary_bkt11;
1168         secondary_bkt20 = secondary_bkt10;
1169         secondary_bkt21 = secondary_bkt11;
1170         primary_hash20 = primary_hash10;
1171         primary_hash21 = primary_hash11;
1172         secondary_hash20 = secondary_hash10;
1173         secondary_hash21 = secondary_hash11;
1174         idx20 = idx10, idx21 = idx11;
1175         idx10 = idx00, idx11 = idx01;
1176
1177         lookup_stage1(idx10, &primary_hash10, &secondary_hash10,
1178                 &primary_bkt10, &secondary_bkt10, hash_vals, keys, h);
1179         lookup_stage1(idx11, &primary_hash11, &secondary_hash11,
1180                 &primary_bkt11, &secondary_bkt11, hash_vals, keys, h);
1181         lookup_stage2(idx20, primary_hash20, secondary_hash20, primary_bkt20,
1182                 secondary_bkt20, &k_slot20, positions, &extra_hits_mask,
1183                 key_store, h);
1184         lookup_stage2(idx21, primary_hash21, secondary_hash21, primary_bkt21,
1185                 secondary_bkt21, &k_slot21, positions, &extra_hits_mask,
1186                 key_store, h);
1187         lookup_stage3(idx30, k_slot30, keys, positions, data, &hits, h);
1188         lookup_stage3(idx31, k_slot31, keys, positions, data, &hits, h);
1189
1190         k_slot30 = k_slot20, k_slot31 = k_slot21;
1191         idx30 = idx20, idx31 = idx21;
1192         primary_bkt20 = primary_bkt10;
1193         primary_bkt21 = primary_bkt11;
1194         secondary_bkt20 = secondary_bkt10;
1195         secondary_bkt21 = secondary_bkt11;
1196         primary_hash20 = primary_hash10;
1197         primary_hash21 = primary_hash11;
1198         secondary_hash20 = secondary_hash10;
1199         secondary_hash21 = secondary_hash11;
1200         idx20 = idx10, idx21 = idx11;
1201
1202         lookup_stage2(idx20, primary_hash20, secondary_hash20, primary_bkt20,
1203                 secondary_bkt20, &k_slot20, positions, &extra_hits_mask,
1204                 key_store, h);
1205         lookup_stage2(idx21, primary_hash21, secondary_hash21, primary_bkt21,
1206                 secondary_bkt21, &k_slot21, positions, &extra_hits_mask,
1207                 key_store, h);
1208         lookup_stage3(idx30, k_slot30, keys, positions, data, &hits, h);
1209         lookup_stage3(idx31, k_slot31, keys, positions, data, &hits, h);
1210
1211         k_slot30 = k_slot20, k_slot31 = k_slot21;
1212         idx30 = idx20, idx31 = idx21;
1213
1214         lookup_stage3(idx30, k_slot30, keys, positions, data, &hits, h);
1215         lookup_stage3(idx31, k_slot31, keys, positions, data, &hits, h);
1216
1217         /* ignore any items we have already found */
1218         extra_hits_mask &= ~hits;
1219
1220         if (unlikely(extra_hits_mask)) {
1221                 /* run a single search for each remaining item */
1222                 do {
1223                         idx = __builtin_ctzl(extra_hits_mask);
1224                         if (data != NULL) {
1225                                 ret = rte_hash_lookup_with_hash_data(h,
1226                                                 keys[idx], hash_vals[idx], &data[idx]);
1227                                 if (ret >= 0)
1228                                         hits |= 1ULL << idx;
1229                         } else {
1230                                 positions[idx] = rte_hash_lookup_with_hash(h,
1231                                                         keys[idx], hash_vals[idx]);
1232                                 if (positions[idx] >= 0)
1233                                         hits |= 1llu << idx;
1234                         }
1235                         extra_hits_mask &= ~(1llu << idx);
1236                 } while (extra_hits_mask);
1237         }
1238
1239         miss_mask &= ~hits;
1240         if (unlikely(miss_mask)) {
1241                 do {
1242                         idx = __builtin_ctzl(miss_mask);
1243                         positions[idx] = -ENOENT;
1244                         miss_mask &= ~(1llu << idx);
1245                 } while (miss_mask);
1246         }
1247
1248         if (hit_mask != NULL)
1249                 *hit_mask = hits;
1250 }
1251
1252 int
1253 rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
1254                       uint32_t num_keys, int32_t *positions)
1255 {
1256         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
1257                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
1258                         (positions == NULL)), -EINVAL);
1259
1260         __rte_hash_lookup_bulk(h, keys, num_keys, positions, NULL, NULL);
1261         return 0;
1262 }
1263
1264 int
1265 rte_hash_lookup_bulk_data(const struct rte_hash *h, const void **keys,
1266                       uint32_t num_keys, uint64_t *hit_mask, void *data[])
1267 {
1268         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
1269                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
1270                         (hit_mask == NULL)), -EINVAL);
1271
1272         int32_t positions[num_keys];
1273
1274         __rte_hash_lookup_bulk(h, keys, num_keys, positions, hit_mask, data);
1275
1276         /* Return number of hits */
1277         return __builtin_popcountl(*hit_mask);
1278 }
1279
1280 int32_t
1281 rte_hash_iterate(const struct rte_hash *h, const void **key, void **data, uint32_t *next)
1282 {
1283         uint32_t bucket_idx, idx, position;
1284         struct rte_hash_key *next_key;
1285
1286         RETURN_IF_TRUE(((h == NULL) || (next == NULL)), -EINVAL);
1287
1288         const uint32_t total_entries = h->num_buckets * RTE_HASH_BUCKET_ENTRIES;
1289         /* Out of bounds */
1290         if (*next >= total_entries)
1291                 return -ENOENT;
1292
1293         /* Calculate bucket and index of current iterator */
1294         bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
1295         idx = *next % RTE_HASH_BUCKET_ENTRIES;
1296
1297         /* If current position is empty, go to the next one */
1298         while (h->buckets[bucket_idx].key_idx[idx] == EMPTY_SLOT) {
1299                 (*next)++;
1300                 /* End of table */
1301                 if (*next == total_entries)
1302                         return -ENOENT;
1303                 bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
1304                 idx = *next % RTE_HASH_BUCKET_ENTRIES;
1305         }
1306
1307         /* Get position of entry in key table */
1308         position = h->buckets[bucket_idx].key_idx[idx];
1309         next_key = (struct rte_hash_key *) ((char *)h->key_store +
1310                                 position * h->key_entry_size);
1311         /* Return key and data */
1312         *key = next_key->key;
1313         *data = next_key->pdata;
1314
1315         /* Increment iterator */
1316         (*next)++;
1317
1318         return position - 1;
1319 }