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         /* Turn on multi-writer only with explicit flat from user and TM
 288          * support.
 289          */
 290         if (params->extra_flag & RTE_HASH_EXTRA_FLAGS_MULTI_WRITER_ADD) {
 291                 if (h->hw_trans_mem_support) {
 292                         h->add_key = ADD_KEY_MULTIWRITER_TM;
 293                 } else {
 294                         h->add_key = ADD_KEY_MULTIWRITER;
 295                         h->multiwriter_lock = rte_malloc(NULL,
 296                                                         sizeof(rte_spinlock_t),
 297                                                         LCORE_CACHE_SIZE);
 298                         rte_spinlock_init(h->multiwriter_lock);
 299                 }
 300         } else
 301                 h->add_key = ADD_KEY_SINGLEWRITER;
 302
 303         /* Populate free slots ring. Entry zero is reserved for key misses. */
 304         for (i = 1; i < params->entries + 1; i++)
 305                 rte_ring_sp_enqueue(r, (void *)((uintptr_t) i));
 306
 307         te->data = (void *) h;
 308         TAILQ_INSERT_TAIL(hash_list, te, next);
 309         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 310
 311         return h;
 312 err_unlock:
 313         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 314 err:
 315         rte_ring_free(r);
 316         rte_free(te);
 317         rte_free(h);
 318         rte_free(buckets);
 319         rte_free(k);
 320         return NULL;
 321 }
 322
 323 void
 324 rte_hash_free(struct rte_hash *h)
 325 {
 326         struct rte_tailq_entry *te;
 327         struct rte_hash_list *hash_list;
 328
 329         if (h == NULL)
 330                 return;
 331
 332         hash_list = RTE_TAILQ_CAST(rte_hash_tailq.head, rte_hash_list);
 333
 334         rte_rwlock_write_lock(RTE_EAL_TAILQ_RWLOCK);
 335
 336         /* find out tailq entry */
 337         TAILQ_FOREACH(te, hash_list, next) {
 338                 if (te->data == (void *) h)
 339                         break;
 340         }
 341
 342         if (te == NULL) {
 343                 rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 344                 return;
 345         }
 346
 347         TAILQ_REMOVE(hash_list, te, next);
 348
 349         rte_rwlock_write_unlock(RTE_EAL_TAILQ_RWLOCK);
 350
 351         if (h->hw_trans_mem_support)
 352                 rte_free(h->local_free_slots);
 353
 354         if (h->add_key == ADD_KEY_MULTIWRITER)
 355                 rte_free(h->multiwriter_lock);
 356         rte_ring_free(h->free_slots);
 357         rte_free(h->key_store);
 358         rte_free(h->buckets);
 359         rte_free(h);
 360         rte_free(te);
 361 }
 362
 363 hash_sig_t
 364 rte_hash_hash(const struct rte_hash *h, const void *key)
 365 {
 366         /* calc hash result by key */
 367         return h->hash_func(key, h->key_len, h->hash_func_init_val);
 368 }
 369
 370 /* Calc the secondary hash value from the primary hash value of a given key */
 371 static inline hash_sig_t
 372 rte_hash_secondary_hash(const hash_sig_t primary_hash)
 373 {
 374         static const unsigned all_bits_shift = 12;
 375         static const unsigned alt_bits_xor = 0x5bd1e995;
 376
 377         uint32_t tag = primary_hash >> all_bits_shift;
 378
 379         return primary_hash ^ ((tag + 1) * alt_bits_xor);
 380 }
 381
 382 void
 383 rte_hash_reset(struct rte_hash *h)
 384 {
 385         void *ptr;
 386         unsigned i;
 387
 388         if (h == NULL)
 389                 return;
 390
 391         memset(h->buckets, 0, h->num_buckets * sizeof(struct rte_hash_bucket));
 392         memset(h->key_store, 0, h->key_entry_size * (h->entries + 1));
 393
 394         /* clear the free ring */
 395         while (rte_ring_dequeue(h->free_slots, &ptr) == 0)
 396                 rte_pause();
 397
 398         /* Repopulate the free slots ring. Entry zero is reserved for key misses */
 399         for (i = 1; i < h->entries + 1; i++)
 400                 rte_ring_sp_enqueue(h->free_slots, (void *)((uintptr_t) i));
 401
 402         if (h->hw_trans_mem_support) {
 403                 /* Reset local caches per lcore */
 404                 for (i = 0; i < RTE_MAX_LCORE; i++)
 405                         h->local_free_slots[i].len = 0;
 406         }
 407 }
 408
 409 /* Search for an entry that can be pushed to its alternative location */
 410 static inline int
 411 make_space_bucket(const struct rte_hash *h, struct rte_hash_bucket *bkt)
 412 {
 413         unsigned i, j;
 414         int ret;
 415         uint32_t next_bucket_idx;
 416         struct rte_hash_bucket *next_bkt[RTE_HASH_BUCKET_ENTRIES];
 417
 418         /*
 419          * Push existing item (search for bucket with space in
 420          * alternative locations) to its alternative location
 421          */
 422         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 423                 /* Search for space in alternative locations */
 424                 next_bucket_idx = bkt->sig_alt[i] & h->bucket_bitmask;
 425                 next_bkt[i] = &h->buckets[next_bucket_idx];
 426                 for (j = 0; j < RTE_HASH_BUCKET_ENTRIES; j++) {
 427                         if (next_bkt[i]->key_idx[j] == EMPTY_SLOT)
 428                                 break;
 429                 }
 430
 431                 if (j != RTE_HASH_BUCKET_ENTRIES)
 432                         break;
 433         }
 434
 435         /* Alternative location has spare room (end of recursive function) */
 436         if (i != RTE_HASH_BUCKET_ENTRIES) {
 437                 next_bkt[i]->sig_alt[j] = bkt->sig_current[i];
 438                 next_bkt[i]->sig_current[j] = bkt->sig_alt[i];
 439                 next_bkt[i]->key_idx[j] = bkt->key_idx[i];
 440                 return i;
 441         }
 442
 443         /* Pick entry that has not been pushed yet */
 444         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++)
 445                 if (bkt->flag[i] == 0)
 446                         break;
 447
 448         /* All entries have been pushed, so entry cannot be added */
 449         if (i == RTE_HASH_BUCKET_ENTRIES)
 450                 return -ENOSPC;
 451
 452         /* Set flag to indicate that this entry is going to be pushed */
 453         bkt->flag[i] = 1;
 454         /* Need room in alternative bucket to insert the pushed entry */
 455         ret = make_space_bucket(h, next_bkt[i]);
 456         /*
 457          * After recursive function.
 458          * Clear flags and insert the pushed entry
 459          * in its alternative location if successful,
 460          * or return error
 461          */
 462         bkt->flag[i] = 0;
 463         if (ret >= 0) {
 464                 next_bkt[i]->sig_alt[ret] = bkt->sig_current[i];
 465                 next_bkt[i]->sig_current[ret] = bkt->sig_alt[i];
 466                 next_bkt[i]->key_idx[ret] = bkt->key_idx[i];
 467                 return i;
 468         } else
 469                 return ret;
 470
 471 }
 472
 473 /*
 474  * Function called to enqueue back an index in the cache/ring,
 475  * as slot has not being used and it can be used in the
 476  * next addition attempt.
 477  */
 478 static inline void
 479 enqueue_slot_back(const struct rte_hash *h,
 480                 struct lcore_cache *cached_free_slots,
 481                 void *slot_id)
 482 {
 483         if (h->hw_trans_mem_support) {
 484                 cached_free_slots->objs[cached_free_slots->len] = slot_id;
 485                 cached_free_slots->len++;
 486         } else
 487                 rte_ring_sp_enqueue(h->free_slots, slot_id);
 488 }
 489
 490 static inline int32_t
 491 __rte_hash_add_key_with_hash(const struct rte_hash *h, const void *key,
 492                                                 hash_sig_t sig, void *data)
 493 {
 494         hash_sig_t alt_hash;
 495         uint32_t prim_bucket_idx, sec_bucket_idx;
 496         unsigned i;
 497         struct rte_hash_bucket *prim_bkt, *sec_bkt;
 498         struct rte_hash_key *new_k, *k, *keys = h->key_store;
 499         void *slot_id = NULL;
 500         uint32_t new_idx;
 501         int ret;
 502         unsigned n_slots;
 503         unsigned lcore_id;
 504         struct lcore_cache *cached_free_slots = NULL;
 505
 506         if (h->add_key == ADD_KEY_MULTIWRITER)
 507                 rte_spinlock_lock(h->multiwriter_lock);
 508
 509         prim_bucket_idx = sig & h->bucket_bitmask;
 510         prim_bkt = &h->buckets[prim_bucket_idx];
 511         rte_prefetch0(prim_bkt);
 512
 513         alt_hash = rte_hash_secondary_hash(sig);
 514         sec_bucket_idx = alt_hash & h->bucket_bitmask;
 515         sec_bkt = &h->buckets[sec_bucket_idx];
 516         rte_prefetch0(sec_bkt);
 517
 518         /* Get a new slot for storing the new key */
 519         if (h->hw_trans_mem_support) {
 520                 lcore_id = rte_lcore_id();
 521                 cached_free_slots = &h->local_free_slots[lcore_id];
 522                 /* Try to get a free slot from the local cache */
 523                 if (cached_free_slots->len == 0) {
 524                         /* Need to get another burst of free slots from global ring */
 525                         n_slots = rte_ring_mc_dequeue_burst(h->free_slots,
 526                                         cached_free_slots->objs, LCORE_CACHE_SIZE);
 527                         if (n_slots == 0)
 528                                 return -ENOSPC;
 529
 530                         cached_free_slots->len += n_slots;
 531                 }
 532
 533                 /* Get a free slot from the local cache */
 534                 cached_free_slots->len--;
 535                 slot_id = cached_free_slots->objs[cached_free_slots->len];
 536         } else {
 537                 if (rte_ring_sc_dequeue(h->free_slots, &slot_id) != 0)
 538                         return -ENOSPC;
 539         }
 540
 541         new_k = RTE_PTR_ADD(keys, (uintptr_t)slot_id * h->key_entry_size);
 542         rte_prefetch0(new_k);
 543         new_idx = (uint32_t)((uintptr_t) slot_id);
 544
 545         /* Check if key is already inserted in primary location */
 546         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 547                 if (prim_bkt->sig_current[i] == sig &&
 548                                 prim_bkt->sig_alt[i] == alt_hash) {
 549                         k = (struct rte_hash_key *) ((char *)keys +
 550                                         prim_bkt->key_idx[i] * h->key_entry_size);
 551                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 552                                 /* Enqueue index of free slot back in the ring. */
 553                                 enqueue_slot_back(h, cached_free_slots, slot_id);
 554                                 /* Update data */
 555                                 k->pdata = data;
 556                                 /*
 557                                  * Return index where key is stored,
 558                                  * substracting the first dummy index
 559                                  */
 560                                 return prim_bkt->key_idx[i] - 1;
 561                         }
 562                 }
 563         }
 564
 565         /* Check if key is already inserted in secondary location */
 566         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 567                 if (sec_bkt->sig_alt[i] == sig &&
 568                                 sec_bkt->sig_current[i] == alt_hash) {
 569                         k = (struct rte_hash_key *) ((char *)keys +
 570                                         sec_bkt->key_idx[i] * h->key_entry_size);
 571                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 572                                 /* Enqueue index of free slot back in the ring. */
 573                                 enqueue_slot_back(h, cached_free_slots, slot_id);
 574                                 /* Update data */
 575                                 k->pdata = data;
 576                                 /*
 577                                  * Return index where key is stored,
 578                                  * substracting the first dummy index
 579                                  */
 580                                 return sec_bkt->key_idx[i] - 1;
 581                         }
 582                 }
 583         }
 584
 585         /* Copy key */
 586         rte_memcpy(new_k->key, key, h->key_len);
 587         new_k->pdata = data;
 588
 589 #if defined(RTE_ARCH_X86) /* currently only x86 support HTM */
 590         if (h->add_key == ADD_KEY_MULTIWRITER_TM) {
 591                 ret = rte_hash_cuckoo_insert_mw_tm(prim_bkt,
 592                                 sig, alt_hash, new_idx);
 593                 if (ret >= 0)
 594                         return new_idx - 1;
 595
 596                 /* Primary bucket full, need to make space for new entry */
 597                 ret = rte_hash_cuckoo_make_space_mw_tm(h, prim_bkt, sig,
 598                                                         alt_hash, new_idx);
 599
 600                 if (ret >= 0)
 601                         return new_idx - 1;
 602
 603                 /* Also search secondary bucket to get better occupancy */
 604                 ret = rte_hash_cuckoo_make_space_mw_tm(h, sec_bkt, sig,
 605                                                         alt_hash, new_idx);
 606
 607                 if (ret >= 0)
 608                         return new_idx - 1;
 609         } else {
 610 #endif
 611                 for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 612                         /* Check if slot is available */
 613                         if (likely(prim_bkt->key_idx[i] == EMPTY_SLOT)) {
 614                                 prim_bkt->sig_current[i] = sig;
 615                                 prim_bkt->sig_alt[i] = alt_hash;
 616                                 prim_bkt->key_idx[i] = new_idx;
 617                                 break;
 618                         }
 619                 }
 620
 621                 if (i != RTE_HASH_BUCKET_ENTRIES) {
 622                         if (h->add_key == ADD_KEY_MULTIWRITER)
 623                                 rte_spinlock_unlock(h->multiwriter_lock);
 624                         return new_idx - 1;
 625                 }
 626
 627                 /* Primary bucket full, need to make space for new entry
 628                  * After recursive function.
 629                  * Insert the new entry in the position of the pushed entry
 630                  * if successful or return error and
 631                  * store the new slot back in the ring
 632                  */
 633                 ret = make_space_bucket(h, prim_bkt);
 634                 if (ret >= 0) {
 635                         prim_bkt->sig_current[ret] = sig;
 636                         prim_bkt->sig_alt[ret] = alt_hash;
 637                         prim_bkt->key_idx[ret] = new_idx;
 638                         if (h->add_key == ADD_KEY_MULTIWRITER)
 639                                 rte_spinlock_unlock(h->multiwriter_lock);
 640                         return new_idx - 1;
 641                 }
 642 #if defined(RTE_ARCH_X86)
 643         }
 644 #endif
 645         /* Error in addition, store new slot back in the ring and return error */
 646         enqueue_slot_back(h, cached_free_slots, (void *)((uintptr_t) new_idx));
 647
 648         if (h->add_key == ADD_KEY_MULTIWRITER)
 649                 rte_spinlock_unlock(h->multiwriter_lock);
 650         return ret;
 651 }
 652
 653 int32_t
 654 rte_hash_add_key_with_hash(const struct rte_hash *h,
 655                         const void *key, hash_sig_t sig)
 656 {
 657         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 658         return __rte_hash_add_key_with_hash(h, key, sig, 0);
 659 }
 660
 661 int32_t
 662 rte_hash_add_key(const struct rte_hash *h, const void *key)
 663 {
 664         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 665         return __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), 0);
 666 }
 667
 668 int
 669 rte_hash_add_key_with_hash_data(const struct rte_hash *h,
 670                         const void *key, hash_sig_t sig, void *data)
 671 {
 672         int ret;
 673
 674         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 675         ret = __rte_hash_add_key_with_hash(h, key, sig, data);
 676         if (ret >= 0)
 677                 return 0;
 678         else
 679                 return ret;
 680 }
 681
 682 int
 683 rte_hash_add_key_data(const struct rte_hash *h, const void *key, void *data)
 684 {
 685         int ret;
 686
 687         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 688
 689         ret = __rte_hash_add_key_with_hash(h, key, rte_hash_hash(h, key), data);
 690         if (ret >= 0)
 691                 return 0;
 692         else
 693                 return ret;
 694 }
 695 static inline int32_t
 696 __rte_hash_lookup_with_hash(const struct rte_hash *h, const void *key,
 697                                         hash_sig_t sig, void **data)
 698 {
 699         uint32_t bucket_idx;
 700         hash_sig_t alt_hash;
 701         unsigned i;
 702         struct rte_hash_bucket *bkt;
 703         struct rte_hash_key *k, *keys = h->key_store;
 704
 705         bucket_idx = sig & h->bucket_bitmask;
 706         bkt = &h->buckets[bucket_idx];
 707
 708         /* Check if key is in primary location */
 709         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 710                 if (bkt->sig_current[i] == sig &&
 711                                 bkt->key_idx[i] != EMPTY_SLOT) {
 712                         k = (struct rte_hash_key *) ((char *)keys +
 713                                         bkt->key_idx[i] * h->key_entry_size);
 714                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 715                                 if (data != NULL)
 716                                         *data = k->pdata;
 717                                 /*
 718                                  * Return index where key is stored,
 719                                  * substracting the first dummy index
 720                                  */
 721                                 return bkt->key_idx[i] - 1;
 722                         }
 723                 }
 724         }
 725
 726         /* Calculate secondary hash */
 727         alt_hash = rte_hash_secondary_hash(sig);
 728         bucket_idx = alt_hash & h->bucket_bitmask;
 729         bkt = &h->buckets[bucket_idx];
 730
 731         /* Check if key is in secondary location */
 732         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 733                 if (bkt->sig_current[i] == alt_hash &&
 734                                 bkt->sig_alt[i] == sig) {
 735                         k = (struct rte_hash_key *) ((char *)keys +
 736                                         bkt->key_idx[i] * h->key_entry_size);
 737                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 738                                 if (data != NULL)
 739                                         *data = k->pdata;
 740                                 /*
 741                                  * Return index where key is stored,
 742                                  * substracting the first dummy index
 743                                  */
 744                                 return bkt->key_idx[i] - 1;
 745                         }
 746                 }
 747         }
 748
 749         return -ENOENT;
 750 }
 751
 752 int32_t
 753 rte_hash_lookup_with_hash(const struct rte_hash *h,
 754                         const void *key, hash_sig_t sig)
 755 {
 756         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 757         return __rte_hash_lookup_with_hash(h, key, sig, NULL);
 758 }
 759
 760 int32_t
 761 rte_hash_lookup(const struct rte_hash *h, const void *key)
 762 {
 763         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 764         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), NULL);
 765 }
 766
 767 int
 768 rte_hash_lookup_with_hash_data(const struct rte_hash *h,
 769                         const void *key, hash_sig_t sig, void **data)
 770 {
 771         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 772         return __rte_hash_lookup_with_hash(h, key, sig, data);
 773 }
 774
 775 int
 776 rte_hash_lookup_data(const struct rte_hash *h, const void *key, void **data)
 777 {
 778         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 779         return __rte_hash_lookup_with_hash(h, key, rte_hash_hash(h, key), data);
 780 }
 781
 782 static inline void
 783 remove_entry(const struct rte_hash *h, struct rte_hash_bucket *bkt, unsigned i)
 784 {
 785         unsigned lcore_id, n_slots;
 786         struct lcore_cache *cached_free_slots;
 787
 788         bkt->sig_current[i] = NULL_SIGNATURE;
 789         bkt->sig_alt[i] = NULL_SIGNATURE;
 790         if (h->hw_trans_mem_support) {
 791                 lcore_id = rte_lcore_id();
 792                 cached_free_slots = &h->local_free_slots[lcore_id];
 793                 /* Cache full, need to free it. */
 794                 if (cached_free_slots->len == LCORE_CACHE_SIZE) {
 795                         /* Need to enqueue the free slots in global ring. */
 796                         n_slots = rte_ring_mp_enqueue_burst(h->free_slots,
 797                                                 cached_free_slots->objs,
 798                                                 LCORE_CACHE_SIZE);
 799                         cached_free_slots->len -= n_slots;
 800                 }
 801                 /* Put index of new free slot in cache. */
 802                 cached_free_slots->objs[cached_free_slots->len] =
 803                                 (void *)((uintptr_t)bkt->key_idx[i]);
 804                 cached_free_slots->len++;
 805         } else {
 806                 rte_ring_sp_enqueue(h->free_slots,
 807                                 (void *)((uintptr_t)bkt->key_idx[i]));
 808         }
 809 }
 810
 811 static inline int32_t
 812 __rte_hash_del_key_with_hash(const struct rte_hash *h, const void *key,
 813                                                 hash_sig_t sig)
 814 {
 815         uint32_t bucket_idx;
 816         hash_sig_t alt_hash;
 817         unsigned i;
 818         struct rte_hash_bucket *bkt;
 819         struct rte_hash_key *k, *keys = h->key_store;
 820         int32_t ret;
 821
 822         bucket_idx = sig & h->bucket_bitmask;
 823         bkt = &h->buckets[bucket_idx];
 824
 825         /* Check if key is in primary location */
 826         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 827                 if (bkt->sig_current[i] == sig &&
 828                                 bkt->key_idx[i] != EMPTY_SLOT) {
 829                         k = (struct rte_hash_key *) ((char *)keys +
 830                                         bkt->key_idx[i] * h->key_entry_size);
 831                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 832                                 remove_entry(h, bkt, i);
 833
 834                                 /*
 835                                  * Return index where key is stored,
 836                                  * substracting the first dummy index
 837                                  */
 838                                 ret = bkt->key_idx[i] - 1;
 839                                 bkt->key_idx[i] = EMPTY_SLOT;
 840                                 return ret;
 841                         }
 842                 }
 843         }
 844
 845         /* Calculate secondary hash */
 846         alt_hash = rte_hash_secondary_hash(sig);
 847         bucket_idx = alt_hash & h->bucket_bitmask;
 848         bkt = &h->buckets[bucket_idx];
 849
 850         /* Check if key is in secondary location */
 851         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 852                 if (bkt->sig_current[i] == alt_hash &&
 853                                 bkt->key_idx[i] != EMPTY_SLOT) {
 854                         k = (struct rte_hash_key *) ((char *)keys +
 855                                         bkt->key_idx[i] * h->key_entry_size);
 856                         if (rte_hash_cmp_eq(key, k->key, h) == 0) {
 857                                 remove_entry(h, bkt, i);
 858
 859                                 /*
 860                                  * Return index where key is stored,
 861                                  * substracting the first dummy index
 862                                  */
 863                                 ret = bkt->key_idx[i] - 1;
 864                                 bkt->key_idx[i] = EMPTY_SLOT;
 865                                 return ret;
 866                         }
 867                 }
 868         }
 869
 870         return -ENOENT;
 871 }
 872
 873 int32_t
 874 rte_hash_del_key_with_hash(const struct rte_hash *h,
 875                         const void *key, hash_sig_t sig)
 876 {
 877         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 878         return __rte_hash_del_key_with_hash(h, key, sig);
 879 }
 880
 881 int32_t
 882 rte_hash_del_key(const struct rte_hash *h, const void *key)
 883 {
 884         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 885         return __rte_hash_del_key_with_hash(h, key, rte_hash_hash(h, key));
 886 }
 887
 888 int
 889 rte_hash_get_key_with_position(const struct rte_hash *h, const int32_t position,
 890                                void **key)
 891 {
 892         RETURN_IF_TRUE(((h == NULL) || (key == NULL)), -EINVAL);
 893
 894         struct rte_hash_key *k, *keys = h->key_store;
 895         k = (struct rte_hash_key *) ((char *) keys + (position + 1) *
 896                                      h->key_entry_size);
 897         *key = k->key;
 898
 899         if (position !=
 900             __rte_hash_lookup_with_hash(h, *key, rte_hash_hash(h, *key),
 901                                         NULL)) {
 902                 return -ENOENT;
 903         }
 904
 905         return 0;
 906 }
 907
 908 /* Lookup bulk stage 0: Prefetch input key */
 909 static inline void
 910 lookup_stage0(unsigned *idx, uint64_t *lookup_mask,
 911                 const void * const *keys)
 912 {
 913         *idx = __builtin_ctzl(*lookup_mask);
 914         if (*lookup_mask == 0)
 915                 *idx = 0;
 916
 917         rte_prefetch0(keys[*idx]);
 918         *lookup_mask &= ~(1llu << *idx);
 919 }
 920
 921 /*
 922  * Lookup bulk stage 1: Calculate primary/secondary hashes
 923  * and prefetch primary/secondary buckets
 924  */
 925 static inline void
 926 lookup_stage1(unsigned idx, hash_sig_t *prim_hash, hash_sig_t *sec_hash,
 927                 const struct rte_hash_bucket **primary_bkt,
 928                 const struct rte_hash_bucket **secondary_bkt,
 929                 hash_sig_t *hash_vals, const void * const *keys,
 930                 const struct rte_hash *h)
 931 {
 932         *prim_hash = rte_hash_hash(h, keys[idx]);
 933         hash_vals[idx] = *prim_hash;
 934         *sec_hash = rte_hash_secondary_hash(*prim_hash);
 935
 936         *primary_bkt = &h->buckets[*prim_hash & h->bucket_bitmask];
 937         *secondary_bkt = &h->buckets[*sec_hash & h->bucket_bitmask];
 938
 939         rte_prefetch0(*primary_bkt);
 940         rte_prefetch0(*secondary_bkt);
 941 }
 942
 943 /*
 944  * Lookup bulk stage 2:  Search for match hashes in primary/secondary locations
 945  * and prefetch first key slot
 946  */
 947 static inline void
 948 lookup_stage2(unsigned idx, hash_sig_t prim_hash, hash_sig_t sec_hash,
 949                 const struct rte_hash_bucket *prim_bkt,
 950                 const struct rte_hash_bucket *sec_bkt,
 951                 const struct rte_hash_key **key_slot, int32_t *positions,
 952                 uint64_t *extra_hits_mask, const void *keys,
 953                 const struct rte_hash *h)
 954 {
 955         unsigned prim_hash_matches, sec_hash_matches, key_idx, i;
 956         unsigned total_hash_matches;
 957
 958         prim_hash_matches = 1 << RTE_HASH_BUCKET_ENTRIES;
 959         sec_hash_matches = 1 << RTE_HASH_BUCKET_ENTRIES;
 960         for (i = 0; i < RTE_HASH_BUCKET_ENTRIES; i++) {
 961                 prim_hash_matches |= ((prim_hash == prim_bkt->sig_current[i]) << i);
 962                 sec_hash_matches |= ((sec_hash == sec_bkt->sig_current[i]) << i);
 963         }
 964
 965         key_idx = prim_bkt->key_idx[__builtin_ctzl(prim_hash_matches)];
 966         if (key_idx == 0)
 967                 key_idx = sec_bkt->key_idx[__builtin_ctzl(sec_hash_matches)];
 968
 969         total_hash_matches = (prim_hash_matches |
 970                                 (sec_hash_matches << (RTE_HASH_BUCKET_ENTRIES + 1)));
 971         *key_slot = (const struct rte_hash_key *) ((const char *)keys +
 972                                         key_idx * h->key_entry_size);
 973
 974         rte_prefetch0(*key_slot);
 975         /*
 976          * Return index where key is stored,
 977          * substracting the first dummy index
 978          */
 979         positions[idx] = (key_idx - 1);
 980
 981         *extra_hits_mask |= (uint64_t)(__builtin_popcount(total_hash_matches) > 3) << idx;
 982
 983 }
 984
 985
 986 /* Lookup bulk stage 3: Check if key matches, update hit mask and return data */
 987 static inline void
 988 lookup_stage3(unsigned idx, const struct rte_hash_key *key_slot, const void * const *keys,
 989                 const int32_t *positions, void *data[], uint64_t *hits,
 990                 const struct rte_hash *h)
 991 {
 992         unsigned hit;
 993         unsigned key_idx;
 994
 995         hit = !rte_hash_cmp_eq(key_slot->key, keys[idx], h);
 996         if (data != NULL)
 997                 data[idx] = key_slot->pdata;
 998
 999         key_idx = positions[idx] + 1;
1000         /*
1001          * If key index is 0, force hit to be 0, in case key to be looked up
1002          * is all zero (as in the dummy slot), which would result in a wrong hit
1003          */
1004         *hits |= (uint64_t)(hit && !!key_idx)  << idx;
1005 }
1006
1007 static inline void
1008 __rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
1009                         uint32_t num_keys, int32_t *positions,
1010                         uint64_t *hit_mask, void *data[])
1011 {
1012         uint64_t hits = 0;
1013         uint64_t extra_hits_mask = 0;
1014         uint64_t lookup_mask, miss_mask;
1015         unsigned idx;
1016         const void *key_store = h->key_store;
1017         int ret;
1018         hash_sig_t hash_vals[RTE_HASH_LOOKUP_BULK_MAX];
1019
1020         unsigned idx00, idx01, idx10, idx11, idx20, idx21, idx30, idx31;
1021         const struct rte_hash_bucket *primary_bkt10, *primary_bkt11;
1022         const struct rte_hash_bucket *secondary_bkt10, *secondary_bkt11;
1023         const struct rte_hash_bucket *primary_bkt20, *primary_bkt21;
1024         const struct rte_hash_bucket *secondary_bkt20, *secondary_bkt21;
1025         const struct rte_hash_key *k_slot20, *k_slot21, *k_slot30, *k_slot31;
1026         hash_sig_t primary_hash10, primary_hash11;
1027         hash_sig_t secondary_hash10, secondary_hash11;
1028         hash_sig_t primary_hash20, primary_hash21;
1029         hash_sig_t secondary_hash20, secondary_hash21;
1030
1031         lookup_mask = (uint64_t) -1 >> (64 - num_keys);
1032         miss_mask = lookup_mask;
1033
1034         lookup_stage0(&idx00, &lookup_mask, keys);
1035         lookup_stage0(&idx01, &lookup_mask, keys);
1036
1037         idx10 = idx00, idx11 = idx01;
1038
1039         lookup_stage0(&idx00, &lookup_mask, keys);
1040         lookup_stage0(&idx01, &lookup_mask, keys);
1041         lookup_stage1(idx10, &primary_hash10, &secondary_hash10,
1042                         &primary_bkt10, &secondary_bkt10, hash_vals, keys, h);
1043         lookup_stage1(idx11, &primary_hash11, &secondary_hash11,
1044                         &primary_bkt11, &secondary_bkt11, hash_vals, keys, h);
1045
1046         primary_bkt20 = primary_bkt10;
1047         primary_bkt21 = primary_bkt11;
1048         secondary_bkt20 = secondary_bkt10;
1049         secondary_bkt21 = secondary_bkt11;
1050         primary_hash20 = primary_hash10;
1051         primary_hash21 = primary_hash11;
1052         secondary_hash20 = secondary_hash10;
1053         secondary_hash21 = secondary_hash11;
1054         idx20 = idx10, idx21 = idx11;
1055         idx10 = idx00, idx11 = idx01;
1056
1057         lookup_stage0(&idx00, &lookup_mask, keys);
1058         lookup_stage0(&idx01, &lookup_mask, keys);
1059         lookup_stage1(idx10, &primary_hash10, &secondary_hash10,
1060                         &primary_bkt10, &secondary_bkt10, hash_vals, keys, h);
1061         lookup_stage1(idx11, &primary_hash11, &secondary_hash11,
1062                         &primary_bkt11, &secondary_bkt11, hash_vals, keys, h);
1063         lookup_stage2(idx20, primary_hash20, secondary_hash20, primary_bkt20,
1064                         secondary_bkt20, &k_slot20, positions, &extra_hits_mask,
1065                         key_store, h);
1066         lookup_stage2(idx21, primary_hash21, secondary_hash21, primary_bkt21,
1067                         secondary_bkt21, &k_slot21, positions, &extra_hits_mask,
1068                         key_store, h);
1069
1070         while (lookup_mask) {
1071                 k_slot30 = k_slot20, k_slot31 = k_slot21;
1072                 idx30 = idx20, idx31 = idx21;
1073                 primary_bkt20 = primary_bkt10;
1074                 primary_bkt21 = primary_bkt11;
1075                 secondary_bkt20 = secondary_bkt10;
1076                 secondary_bkt21 = secondary_bkt11;
1077                 primary_hash20 = primary_hash10;
1078                 primary_hash21 = primary_hash11;
1079                 secondary_hash20 = secondary_hash10;
1080                 secondary_hash21 = secondary_hash11;
1081                 idx20 = idx10, idx21 = idx11;
1082                 idx10 = idx00, idx11 = idx01;
1083
1084                 lookup_stage0(&idx00, &lookup_mask, keys);
1085                 lookup_stage0(&idx01, &lookup_mask, keys);
1086                 lookup_stage1(idx10, &primary_hash10, &secondary_hash10,
1087                         &primary_bkt10, &secondary_bkt10, hash_vals, keys, h);
1088                 lookup_stage1(idx11, &primary_hash11, &secondary_hash11,
1089                         &primary_bkt11, &secondary_bkt11, hash_vals, keys, h);
1090                 lookup_stage2(idx20, primary_hash20, secondary_hash20,
1091                         primary_bkt20, secondary_bkt20, &k_slot20, positions,
1092                         &extra_hits_mask, key_store, h);
1093                 lookup_stage2(idx21, primary_hash21, secondary_hash21,
1094                         primary_bkt21, secondary_bkt21, &k_slot21, positions,
1095                         &extra_hits_mask, key_store, h);
1096                 lookup_stage3(idx30, k_slot30, keys, positions, data, &hits, h);
1097                 lookup_stage3(idx31, k_slot31, keys, positions, data, &hits, h);
1098         }
1099
1100         k_slot30 = k_slot20, k_slot31 = k_slot21;
1101         idx30 = idx20, idx31 = idx21;
1102         primary_bkt20 = primary_bkt10;
1103         primary_bkt21 = primary_bkt11;
1104         secondary_bkt20 = secondary_bkt10;
1105         secondary_bkt21 = secondary_bkt11;
1106         primary_hash20 = primary_hash10;
1107         primary_hash21 = primary_hash11;
1108         secondary_hash20 = secondary_hash10;
1109         secondary_hash21 = secondary_hash11;
1110         idx20 = idx10, idx21 = idx11;
1111         idx10 = idx00, idx11 = idx01;
1112
1113         lookup_stage1(idx10, &primary_hash10, &secondary_hash10,
1114                 &primary_bkt10, &secondary_bkt10, hash_vals, keys, h);
1115         lookup_stage1(idx11, &primary_hash11, &secondary_hash11,
1116                 &primary_bkt11, &secondary_bkt11, hash_vals, keys, h);
1117         lookup_stage2(idx20, primary_hash20, secondary_hash20, primary_bkt20,
1118                 secondary_bkt20, &k_slot20, positions, &extra_hits_mask,
1119                 key_store, h);
1120         lookup_stage2(idx21, primary_hash21, secondary_hash21, primary_bkt21,
1121                 secondary_bkt21, &k_slot21, positions, &extra_hits_mask,
1122                 key_store, h);
1123         lookup_stage3(idx30, k_slot30, keys, positions, data, &hits, h);
1124         lookup_stage3(idx31, k_slot31, keys, positions, data, &hits, h);
1125
1126         k_slot30 = k_slot20, k_slot31 = k_slot21;
1127         idx30 = idx20, idx31 = idx21;
1128         primary_bkt20 = primary_bkt10;
1129         primary_bkt21 = primary_bkt11;
1130         secondary_bkt20 = secondary_bkt10;
1131         secondary_bkt21 = secondary_bkt11;
1132         primary_hash20 = primary_hash10;
1133         primary_hash21 = primary_hash11;
1134         secondary_hash20 = secondary_hash10;
1135         secondary_hash21 = secondary_hash11;
1136         idx20 = idx10, idx21 = idx11;
1137
1138         lookup_stage2(idx20, primary_hash20, secondary_hash20, primary_bkt20,
1139                 secondary_bkt20, &k_slot20, positions, &extra_hits_mask,
1140                 key_store, h);
1141         lookup_stage2(idx21, primary_hash21, secondary_hash21, primary_bkt21,
1142                 secondary_bkt21, &k_slot21, positions, &extra_hits_mask,
1143                 key_store, h);
1144         lookup_stage3(idx30, k_slot30, keys, positions, data, &hits, h);
1145         lookup_stage3(idx31, k_slot31, keys, positions, data, &hits, h);
1146
1147         k_slot30 = k_slot20, k_slot31 = k_slot21;
1148         idx30 = idx20, idx31 = idx21;
1149
1150         lookup_stage3(idx30, k_slot30, keys, positions, data, &hits, h);
1151         lookup_stage3(idx31, k_slot31, keys, positions, data, &hits, h);
1152
1153         /* ignore any items we have already found */
1154         extra_hits_mask &= ~hits;
1155
1156         if (unlikely(extra_hits_mask)) {
1157                 /* run a single search for each remaining item */
1158                 do {
1159                         idx = __builtin_ctzl(extra_hits_mask);
1160                         if (data != NULL) {
1161                                 ret = rte_hash_lookup_with_hash_data(h,
1162                                                 keys[idx], hash_vals[idx], &data[idx]);
1163                                 if (ret >= 0)
1164                                         hits |= 1ULL << idx;
1165                         } else {
1166                                 positions[idx] = rte_hash_lookup_with_hash(h,
1167                                                         keys[idx], hash_vals[idx]);
1168                                 if (positions[idx] >= 0)
1169                                         hits |= 1llu << idx;
1170                         }
1171                         extra_hits_mask &= ~(1llu << idx);
1172                 } while (extra_hits_mask);
1173         }
1174
1175         miss_mask &= ~hits;
1176         if (unlikely(miss_mask)) {
1177                 do {
1178                         idx = __builtin_ctzl(miss_mask);
1179                         positions[idx] = -ENOENT;
1180                         miss_mask &= ~(1llu << idx);
1181                 } while (miss_mask);
1182         }
1183
1184         if (hit_mask != NULL)
1185                 *hit_mask = hits;
1186 }
1187
1188 int
1189 rte_hash_lookup_bulk(const struct rte_hash *h, const void **keys,
1190                       uint32_t num_keys, int32_t *positions)
1191 {
1192         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
1193                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
1194                         (positions == NULL)), -EINVAL);
1195
1196         __rte_hash_lookup_bulk(h, keys, num_keys, positions, NULL, NULL);
1197         return 0;
1198 }
1199
1200 int
1201 rte_hash_lookup_bulk_data(const struct rte_hash *h, const void **keys,
1202                       uint32_t num_keys, uint64_t *hit_mask, void *data[])
1203 {
1204         RETURN_IF_TRUE(((h == NULL) || (keys == NULL) || (num_keys == 0) ||
1205                         (num_keys > RTE_HASH_LOOKUP_BULK_MAX) ||
1206                         (hit_mask == NULL)), -EINVAL);
1207
1208         int32_t positions[num_keys];
1209
1210         __rte_hash_lookup_bulk(h, keys, num_keys, positions, hit_mask, data);
1211
1212         /* Return number of hits */
1213         return __builtin_popcountl(*hit_mask);
1214 }
1215
1216 int32_t
1217 rte_hash_iterate(const struct rte_hash *h, const void **key, void **data, uint32_t *next)
1218 {
1219         uint32_t bucket_idx, idx, position;
1220         struct rte_hash_key *next_key;
1221
1222         RETURN_IF_TRUE(((h == NULL) || (next == NULL)), -EINVAL);
1223
1224         const uint32_t total_entries = h->num_buckets * RTE_HASH_BUCKET_ENTRIES;
1225         /* Out of bounds */
1226         if (*next >= total_entries)
1227                 return -ENOENT;
1228
1229         /* Calculate bucket and index of current iterator */
1230         bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
1231         idx = *next % RTE_HASH_BUCKET_ENTRIES;
1232
1233         /* If current position is empty, go to the next one */
1234         while (h->buckets[bucket_idx].key_idx[idx] == EMPTY_SLOT) {
1235                 (*next)++;
1236                 /* End of table */
1237                 if (*next == total_entries)
1238                         return -ENOENT;
1239                 bucket_idx = *next / RTE_HASH_BUCKET_ENTRIES;
1240                 idx = *next % RTE_HASH_BUCKET_ENTRIES;
1241         }
1242
1243         /* Get position of entry in key table */
1244         position = h->buckets[bucket_idx].key_idx[idx];
1245         next_key = (struct rte_hash_key *) ((char *)h->key_store +
1246                                 position * h->key_entry_size);
1247         /* Return key and data */
1248         *key = next_key->key;
1249         *data = next_key->pdata;
1250
1251         /* Increment iterator */
1252         (*next)++;
1253
1254         return position - 1;
1255 }