lib/librte_table/rte_table_hash_key16.c

   1 /*-
   2  *       BSD LICENSE
   3  *
   4  *       Copyright(c) 2010-2014 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 #include <string.h>
  34 #include <stdio.h>
  35
  36 #include <rte_common.h>
  37 #include <rte_mbuf.h>
  38 #include <rte_memory.h>
  39 #include <rte_malloc.h>
  40 #include <rte_log.h>
  41
  42 #include "rte_table_hash.h"
  43 #include "rte_lru.h"
  44
  45 #define RTE_TABLE_HASH_KEY_SIZE                                         16
  46
  47 #define RTE_BUCKET_ENTRY_VALID                                          0x1LLU
  48
  49 #ifdef RTE_TABLE_STATS_COLLECT
  50
  51 #define RTE_TABLE_HASH_KEY16_STATS_PKTS_IN_ADD(table, val) \
  52         table->stats.n_pkts_in += val
  53 #define RTE_TABLE_HASH_KEY16_STATS_PKTS_LOOKUP_MISS(table, val) \
  54         table->stats.n_pkts_lookup_miss += val
  55
  56 #else
  57
  58 #define RTE_TABLE_HASH_KEY16_STATS_PKTS_IN_ADD(table, val)
  59 #define RTE_TABLE_HASH_KEY16_STATS_PKTS_LOOKUP_MISS(table, val)
  60
  61 #endif
  62
  63 struct rte_bucket_4_16 {
  64         /* Cache line 0 */
  65         uint64_t signature[4 + 1];
  66         uint64_t lru_list;
  67         struct rte_bucket_4_16 *next;
  68         uint64_t next_valid;
  69
  70         /* Cache line 1 */
  71         uint64_t key[4][2];
  72
  73         /* Cache line 2 */
  74         uint8_t data[0];
  75 };
  76
  77 struct rte_table_hash {
  78         struct rte_table_stats stats;
  79
  80         /* Input parameters */
  81         uint32_t n_buckets;
  82         uint32_t n_entries_per_bucket;
  83         uint32_t key_size;
  84         uint32_t entry_size;
  85         uint32_t bucket_size;
  86         uint32_t signature_offset;
  87         uint32_t key_offset;
  88         uint64_t key_mask[2];
  89         rte_table_hash_op_hash f_hash;
  90         uint64_t seed;
  91
  92         /* Extendible buckets */
  93         uint32_t n_buckets_ext;
  94         uint32_t stack_pos;
  95         uint32_t *stack;
  96
  97         /* Lookup table */
  98         uint8_t memory[0] __rte_cache_aligned;
  99 };
 100
 101 static int
 102 check_params_create_lru(struct rte_table_hash_key16_lru_params *params) {
 103         /* n_entries */
 104         if (params->n_entries == 0) {
 105                 RTE_LOG(ERR, TABLE, "%s: n_entries is zero\n", __func__);
 106                 return -EINVAL;
 107         }
 108
 109         /* f_hash */
 110         if (params->f_hash == NULL) {
 111                 RTE_LOG(ERR, TABLE,
 112                         "%s: f_hash function pointer is NULL\n", __func__);
 113                 return -EINVAL;
 114         }
 115
 116         return 0;
 117 }
 118
 119 static void *
 120 rte_table_hash_create_key16_lru(void *params,
 121                 int socket_id,
 122                 uint32_t entry_size)
 123 {
 124         struct rte_table_hash_key16_lru_params *p =
 125                         (struct rte_table_hash_key16_lru_params *) params;
 126         struct rte_table_hash *f;
 127         uint32_t n_buckets, n_entries_per_bucket,
 128                         key_size, bucket_size_cl, total_size, i;
 129
 130         /* Check input parameters */
 131         if ((check_params_create_lru(p) != 0) ||
 132                 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
 133                 ((sizeof(struct rte_bucket_4_16) % RTE_CACHE_LINE_SIZE) != 0))
 134                 return NULL;
 135         n_entries_per_bucket = 4;
 136         key_size = 16;
 137
 138         /* Memory allocation */
 139         n_buckets = rte_align32pow2((p->n_entries + n_entries_per_bucket - 1) /
 140                 n_entries_per_bucket);
 141         bucket_size_cl = (sizeof(struct rte_bucket_4_16) + n_entries_per_bucket
 142                 * entry_size + RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
 143         total_size = sizeof(struct rte_table_hash) + n_buckets *
 144                 bucket_size_cl * RTE_CACHE_LINE_SIZE;
 145
 146         f = rte_zmalloc_socket("TABLE", total_size, RTE_CACHE_LINE_SIZE, socket_id);
 147         if (f == NULL) {
 148                 RTE_LOG(ERR, TABLE,
 149                 "%s: Cannot allocate %u bytes for hash table\n",
 150                 __func__, total_size);
 151                 return NULL;
 152         }
 153         RTE_LOG(INFO, TABLE,
 154                 "%s: Hash table memory footprint is %u bytes\n",
 155                 __func__, total_size);
 156
 157         /* Memory initialization */
 158         f->n_buckets = n_buckets;
 159         f->n_entries_per_bucket = n_entries_per_bucket;
 160         f->key_size = key_size;
 161         f->entry_size = entry_size;
 162         f->bucket_size = bucket_size_cl * RTE_CACHE_LINE_SIZE;
 163         f->signature_offset = p->signature_offset;
 164         f->key_offset = p->key_offset;
 165         f->f_hash = p->f_hash;
 166         f->seed = p->seed;
 167
 168         if (p->key_mask != NULL) {
 169                 f->key_mask[0] = ((uint64_t *)p->key_mask)[0];
 170                 f->key_mask[1] = ((uint64_t *)p->key_mask)[1];
 171         } else {
 172                 f->key_mask[0] = 0xFFFFFFFFFFFFFFFFLLU;
 173                 f->key_mask[1] = 0xFFFFFFFFFFFFFFFFLLU;
 174         }
 175
 176         for (i = 0; i < n_buckets; i++) {
 177                 struct rte_bucket_4_16 *bucket;
 178
 179                 bucket = (struct rte_bucket_4_16 *) &f->memory[i *
 180                         f->bucket_size];
 181                 lru_init(bucket);
 182         }
 183
 184         return f;
 185 }
 186
 187 static int
 188 rte_table_hash_free_key16_lru(void *table)
 189 {
 190         struct rte_table_hash *f = (struct rte_table_hash *) table;
 191
 192         /* Check input parameters */
 193         if (f == NULL) {
 194                 RTE_LOG(ERR, TABLE, "%s: table parameter is NULL\n", __func__);
 195                 return -EINVAL;
 196         }
 197
 198         rte_free(f);
 199         return 0;
 200 }
 201
 202 static int
 203 rte_table_hash_entry_add_key16_lru(
 204         void *table,
 205         void *key,
 206         void *entry,
 207         int *key_found,
 208         void **entry_ptr)
 209 {
 210         struct rte_table_hash *f = (struct rte_table_hash *) table;
 211         struct rte_bucket_4_16 *bucket;
 212         uint64_t signature, pos;
 213         uint32_t bucket_index, i;
 214
 215         signature = f->f_hash(key, f->key_size, f->seed);
 216         bucket_index = signature & (f->n_buckets - 1);
 217         bucket = (struct rte_bucket_4_16 *)
 218                         &f->memory[bucket_index * f->bucket_size];
 219         signature |= RTE_BUCKET_ENTRY_VALID;
 220
 221         /* Key is present in the bucket */
 222         for (i = 0; i < 4; i++) {
 223                 uint64_t bucket_signature = bucket->signature[i];
 224                 uint8_t *bucket_key = (uint8_t *) bucket->key[i];
 225
 226                 if ((bucket_signature == signature) &&
 227                                 (memcmp(key, bucket_key, f->key_size) == 0)) {
 228                         uint8_t *bucket_data = &bucket->data[i * f->entry_size];
 229
 230                         memcpy(bucket_data, entry, f->entry_size);
 231                         lru_update(bucket, i);
 232                         *key_found = 1;
 233                         *entry_ptr = (void *) bucket_data;
 234                         return 0;
 235                 }
 236         }
 237
 238         /* Key is not present in the bucket */
 239         for (i = 0; i < 4; i++) {
 240                 uint64_t bucket_signature = bucket->signature[i];
 241                 uint8_t *bucket_key = (uint8_t *) bucket->key[i];
 242
 243                 if (bucket_signature == 0) {
 244                         uint8_t *bucket_data = &bucket->data[i * f->entry_size];
 245
 246                         bucket->signature[i] = signature;
 247                         memcpy(bucket_key, key, f->key_size);
 248                         memcpy(bucket_data, entry, f->entry_size);
 249                         lru_update(bucket, i);
 250                         *key_found = 0;
 251                         *entry_ptr = (void *) bucket_data;
 252
 253                         return 0;
 254                 }
 255         }
 256
 257         /* Bucket full: replace LRU entry */
 258         pos = lru_pos(bucket);
 259         bucket->signature[pos] = signature;
 260         memcpy(bucket->key[pos], key, f->key_size);
 261         memcpy(&bucket->data[pos * f->entry_size], entry, f->entry_size);
 262         lru_update(bucket, pos);
 263         *key_found = 0;
 264         *entry_ptr = (void *) &bucket->data[pos * f->entry_size];
 265
 266         return 0;
 267 }
 268
 269 static int
 270 rte_table_hash_entry_delete_key16_lru(
 271         void *table,
 272         void *key,
 273         int *key_found,
 274         void *entry)
 275 {
 276         struct rte_table_hash *f = (struct rte_table_hash *) table;
 277         struct rte_bucket_4_16 *bucket;
 278         uint64_t signature;
 279         uint32_t bucket_index, i;
 280
 281         signature = f->f_hash(key, f->key_size, f->seed);
 282         bucket_index = signature & (f->n_buckets - 1);
 283         bucket = (struct rte_bucket_4_16 *)
 284                         &f->memory[bucket_index * f->bucket_size];
 285         signature |= RTE_BUCKET_ENTRY_VALID;
 286
 287         /* Key is present in the bucket */
 288         for (i = 0; i < 4; i++) {
 289                 uint64_t bucket_signature = bucket->signature[i];
 290                 uint8_t *bucket_key = (uint8_t *) bucket->key[i];
 291
 292                 if ((bucket_signature == signature) &&
 293                                 (memcmp(key, bucket_key, f->key_size) == 0)) {
 294                         uint8_t *bucket_data = &bucket->data[i * f->entry_size];
 295
 296                         bucket->signature[i] = 0;
 297                         *key_found = 1;
 298                         if (entry)
 299                                 memcpy(entry, bucket_data, f->entry_size);
 300                         return 0;
 301                 }
 302         }
 303
 304         /* Key is not present in the bucket */
 305         *key_found = 0;
 306         return 0;
 307 }
 308
 309 static int
 310 check_params_create_ext(struct rte_table_hash_key16_ext_params *params) {
 311         /* n_entries */
 312         if (params->n_entries == 0) {
 313                 RTE_LOG(ERR, TABLE, "%s: n_entries is zero\n", __func__);
 314                 return -EINVAL;
 315         }
 316
 317         /* n_entries_ext */
 318         if (params->n_entries_ext == 0) {
 319                 RTE_LOG(ERR, TABLE, "%s: n_entries_ext is zero\n", __func__);
 320                 return -EINVAL;
 321         }
 322
 323         /* f_hash */
 324         if (params->f_hash == NULL) {
 325                 RTE_LOG(ERR, TABLE,
 326                         "%s: f_hash function pointer is NULL\n", __func__);
 327                 return -EINVAL;
 328         }
 329
 330         return 0;
 331 }
 332
 333 static void *
 334 rte_table_hash_create_key16_ext(void *params,
 335                 int socket_id,
 336                 uint32_t entry_size)
 337 {
 338         struct rte_table_hash_key16_ext_params *p =
 339                         (struct rte_table_hash_key16_ext_params *) params;
 340         struct rte_table_hash *f;
 341         uint32_t n_buckets, n_buckets_ext, n_entries_per_bucket, key_size,
 342                         bucket_size_cl, stack_size_cl, total_size, i;
 343
 344         /* Check input parameters */
 345         if ((check_params_create_ext(p) != 0) ||
 346                 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
 347                 ((sizeof(struct rte_bucket_4_16) % RTE_CACHE_LINE_SIZE) != 0))
 348                 return NULL;
 349
 350         n_entries_per_bucket = 4;
 351         key_size = 16;
 352
 353         /* Memory allocation */
 354         n_buckets = rte_align32pow2((p->n_entries + n_entries_per_bucket - 1) /
 355                 n_entries_per_bucket);
 356         n_buckets_ext = (p->n_entries_ext + n_entries_per_bucket - 1) /
 357                 n_entries_per_bucket;
 358         bucket_size_cl = (sizeof(struct rte_bucket_4_16) + n_entries_per_bucket
 359                 * entry_size + RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
 360         stack_size_cl = (n_buckets_ext * sizeof(uint32_t) + RTE_CACHE_LINE_SIZE - 1)
 361                 / RTE_CACHE_LINE_SIZE;
 362         total_size = sizeof(struct rte_table_hash) +
 363                 ((n_buckets + n_buckets_ext) * bucket_size_cl + stack_size_cl) *
 364                 RTE_CACHE_LINE_SIZE;
 365
 366         f = rte_zmalloc_socket("TABLE", total_size, RTE_CACHE_LINE_SIZE, socket_id);
 367         if (f == NULL) {
 368                 RTE_LOG(ERR, TABLE,
 369                         "%s: Cannot allocate %u bytes for hash table\n",
 370                         __func__, total_size);
 371                 return NULL;
 372         }
 373         RTE_LOG(INFO, TABLE,
 374                 "%s: Hash table memory footprint is %u bytes\n",
 375                 __func__, total_size);
 376
 377         /* Memory initialization */
 378         f->n_buckets = n_buckets;
 379         f->n_entries_per_bucket = n_entries_per_bucket;
 380         f->key_size = key_size;
 381         f->entry_size = entry_size;
 382         f->bucket_size = bucket_size_cl * RTE_CACHE_LINE_SIZE;
 383         f->signature_offset = p->signature_offset;
 384         f->key_offset = p->key_offset;
 385         f->f_hash = p->f_hash;
 386         f->seed = p->seed;
 387
 388         f->n_buckets_ext = n_buckets_ext;
 389         f->stack_pos = n_buckets_ext;
 390         f->stack = (uint32_t *)
 391                 &f->memory[(n_buckets + n_buckets_ext) * f->bucket_size];
 392
 393         for (i = 0; i < n_buckets_ext; i++)
 394                 f->stack[i] = i;
 395
 396         if (p->key_mask != NULL) {
 397                 f->key_mask[0] = (((uint64_t *)p->key_mask)[0]);
 398                 f->key_mask[1] = (((uint64_t *)p->key_mask)[1]);
 399         } else {
 400                 f->key_mask[0] = 0xFFFFFFFFFFFFFFFFLLU;
 401                 f->key_mask[1] = 0xFFFFFFFFFFFFFFFFLLU;
 402         }
 403
 404         return f;
 405 }
 406
 407 static int
 408 rte_table_hash_free_key16_ext(void *table)
 409 {
 410         struct rte_table_hash *f = (struct rte_table_hash *) table;
 411
 412         /* Check input parameters */
 413         if (f == NULL) {
 414                 RTE_LOG(ERR, TABLE, "%s: table parameter is NULL\n", __func__);
 415                 return -EINVAL;
 416         }
 417
 418         rte_free(f);
 419         return 0;
 420 }
 421
 422 static int
 423 rte_table_hash_entry_add_key16_ext(
 424         void *table,
 425         void *key,
 426         void *entry,
 427         int *key_found,
 428         void **entry_ptr)
 429 {
 430         struct rte_table_hash *f = (struct rte_table_hash *) table;
 431         struct rte_bucket_4_16 *bucket0, *bucket, *bucket_prev;
 432         uint64_t signature;
 433         uint32_t bucket_index, i;
 434
 435         signature = f->f_hash(key, f->key_size, f->seed);
 436         bucket_index = signature & (f->n_buckets - 1);
 437         bucket0 = (struct rte_bucket_4_16 *)
 438                         &f->memory[bucket_index * f->bucket_size];
 439         signature |= RTE_BUCKET_ENTRY_VALID;
 440
 441         /* Key is present in the bucket */
 442         for (bucket = bucket0; bucket != NULL; bucket = bucket->next)
 443                 for (i = 0; i < 4; i++) {
 444                         uint64_t bucket_signature = bucket->signature[i];
 445                         uint8_t *bucket_key = (uint8_t *) bucket->key[i];
 446
 447                         if ((bucket_signature == signature) &&
 448                                 (memcmp(key, bucket_key, f->key_size) == 0)) {
 449                                 uint8_t *bucket_data = &bucket->data[i *
 450                                         f->entry_size];
 451
 452                                 memcpy(bucket_data, entry, f->entry_size);
 453                                 *key_found = 1;
 454                                 *entry_ptr = (void *) bucket_data;
 455                                 return 0;
 456                         }
 457                 }
 458
 459         /* Key is not present in the bucket */
 460         for (bucket_prev = NULL, bucket = bucket0; bucket != NULL;
 461                          bucket_prev = bucket, bucket = bucket->next)
 462                 for (i = 0; i < 4; i++) {
 463                         uint64_t bucket_signature = bucket->signature[i];
 464                         uint8_t *bucket_key = (uint8_t *) bucket->key[i];
 465
 466                         if (bucket_signature == 0) {
 467                                 uint8_t *bucket_data = &bucket->data[i *
 468                                         f->entry_size];
 469
 470                                 bucket->signature[i] = signature;
 471                                 memcpy(bucket_key, key, f->key_size);
 472                                 memcpy(bucket_data, entry, f->entry_size);
 473                                 *key_found = 0;
 474                                 *entry_ptr = (void *) bucket_data;
 475
 476                                 return 0;
 477                         }
 478                 }
 479
 480         /* Bucket full: extend bucket */
 481         if (f->stack_pos > 0) {
 482                 bucket_index = f->stack[--f->stack_pos];
 483
 484                 bucket = (struct rte_bucket_4_16 *) &f->memory[(f->n_buckets +
 485                         bucket_index) * f->bucket_size];
 486                 bucket_prev->next = bucket;
 487                 bucket_prev->next_valid = 1;
 488
 489                 bucket->signature[0] = signature;
 490                 memcpy(bucket->key[0], key, f->key_size);
 491                 memcpy(&bucket->data[0], entry, f->entry_size);
 492                 *key_found = 0;
 493                 *entry_ptr = (void *) &bucket->data[0];
 494                 return 0;
 495         }
 496
 497         return -ENOSPC;
 498 }
 499
 500 static int
 501 rte_table_hash_entry_delete_key16_ext(
 502         void *table,
 503         void *key,
 504         int *key_found,
 505         void *entry)
 506 {
 507         struct rte_table_hash *f = (struct rte_table_hash *) table;
 508         struct rte_bucket_4_16 *bucket0, *bucket, *bucket_prev;
 509         uint64_t signature;
 510         uint32_t bucket_index, i;
 511
 512         signature = f->f_hash(key, f->key_size, f->seed);
 513         bucket_index = signature & (f->n_buckets - 1);
 514         bucket0 = (struct rte_bucket_4_16 *)
 515                 &f->memory[bucket_index * f->bucket_size];
 516         signature |= RTE_BUCKET_ENTRY_VALID;
 517
 518         /* Key is present in the bucket */
 519         for (bucket_prev = NULL, bucket = bucket0; bucket != NULL;
 520                 bucket_prev = bucket, bucket = bucket->next)
 521                 for (i = 0; i < 4; i++) {
 522                         uint64_t bucket_signature = bucket->signature[i];
 523                         uint8_t *bucket_key = (uint8_t *) bucket->key[i];
 524
 525                         if ((bucket_signature == signature) &&
 526                                 (memcmp(key, bucket_key, f->key_size) == 0)) {
 527                                 uint8_t *bucket_data = &bucket->data[i *
 528                                         f->entry_size];
 529
 530                                 bucket->signature[i] = 0;
 531                                 *key_found = 1;
 532                                 if (entry)
 533                                         memcpy(entry, bucket_data,
 534                                         f->entry_size);
 535
 536                                 if ((bucket->signature[0] == 0) &&
 537                                         (bucket->signature[1] == 0) &&
 538                                         (bucket->signature[2] == 0) &&
 539                                         (bucket->signature[3] == 0) &&
 540                                         (bucket_prev != NULL)) {
 541                                         bucket_prev->next = bucket->next;
 542                                         bucket_prev->next_valid =
 543                                                 bucket->next_valid;
 544
 545                                         memset(bucket, 0,
 546                                                 sizeof(struct rte_bucket_4_16));
 547                                         bucket_index = (((uint8_t *)bucket -
 548                                                 (uint8_t *)f->memory)/f->bucket_size) - f->n_buckets;
 549                                         f->stack[f->stack_pos++] = bucket_index;
 550                                 }
 551
 552                                 return 0;
 553                         }
 554                 }
 555
 556         /* Key is not present in the bucket */
 557         *key_found = 0;
 558         return 0;
 559 }
 560
 561 #define lookup_key16_cmp(key_in, bucket, pos)                   \
 562 {                                                               \
 563         uint64_t xor[4][2], or[4], signature[4];                \
 564                                                                 \
 565         signature[0] = (~bucket->signature[0]) & 1;             \
 566         signature[1] = (~bucket->signature[1]) & 1;             \
 567         signature[2] = (~bucket->signature[2]) & 1;             \
 568         signature[3] = (~bucket->signature[3]) & 1;             \
 569                                                                 \
 570         xor[0][0] = key_in[0] ^  bucket->key[0][0];             \
 571         xor[0][1] = key_in[1] ^  bucket->key[0][1];             \
 572                                                                 \
 573         xor[1][0] = key_in[0] ^  bucket->key[1][0];             \
 574         xor[1][1] = key_in[1] ^  bucket->key[1][1];             \
 575                                                                 \
 576         xor[2][0] = key_in[0] ^  bucket->key[2][0];             \
 577         xor[2][1] = key_in[1] ^  bucket->key[2][1];             \
 578                                                                 \
 579         xor[3][0] = key_in[0] ^  bucket->key[3][0];             \
 580         xor[3][1] = key_in[1] ^  bucket->key[3][1];             \
 581                                                                 \
 582         or[0] = xor[0][0] | xor[0][1] | signature[0];           \
 583         or[1] = xor[1][0] | xor[1][1] | signature[1];           \
 584         or[2] = xor[2][0] | xor[2][1] | signature[2];           \
 585         or[3] = xor[3][0] | xor[3][1] | signature[3];           \
 586                                                                 \
 587         pos = 4;                                                \
 588         if (or[0] == 0)                                         \
 589                 pos = 0;                                        \
 590         if (or[1] == 0)                                         \
 591                 pos = 1;                                        \
 592         if (or[2] == 0)                                         \
 593                 pos = 2;                                        \
 594         if (or[3] == 0)                                         \
 595                 pos = 3;                                        \
 596 }
 597
 598 #define lookup1_stage0(pkt0_index, mbuf0, pkts, pkts_mask)      \
 599 {                                                               \
 600         uint64_t pkt_mask;                                      \
 601                                                                 \
 602         pkt0_index = __builtin_ctzll(pkts_mask);                \
 603         pkt_mask = 1LLU << pkt0_index;                          \
 604         pkts_mask &= ~pkt_mask;                                 \
 605                                                                 \
 606         mbuf0 = pkts[pkt0_index];                               \
 607         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf0, 0));   \
 608 }
 609
 610 #define lookup1_stage1(mbuf1, bucket1, f)                       \
 611 {                                                               \
 612         uint64_t signature;                                     \
 613         uint32_t bucket_index;                                  \
 614                                                                 \
 615         signature = RTE_MBUF_METADATA_UINT32(mbuf1, f->signature_offset);\
 616         bucket_index = signature & (f->n_buckets - 1);          \
 617         bucket1 = (struct rte_bucket_4_16 *)                    \
 618                 &f->memory[bucket_index * f->bucket_size];      \
 619         rte_prefetch0(bucket1);                                 \
 620         rte_prefetch0((void *)(((uintptr_t) bucket1) + RTE_CACHE_LINE_SIZE));\
 621 }
 622
 623 #define lookup1_stage1_dosig(mbuf1, bucket1, f)                 \
 624 {                                                               \
 625         uint64_t *key;                                          \
 626         uint64_t signature = 0;                         \
 627         uint32_t bucket_index;                          \
 628         uint64_t hash_key_buffer[2];            \
 629                                                                 \
 630         key = RTE_MBUF_METADATA_UINT64_PTR(mbuf1, f->key_offset);\
 631                                                                 \
 632         hash_key_buffer[0] = key[0] & f->key_mask[0];   \
 633         hash_key_buffer[1] = key[1] & f->key_mask[1];   \
 634         signature = f->f_hash(hash_key_buffer,                  \
 635                         RTE_TABLE_HASH_KEY_SIZE, f->seed);              \
 636                                                                 \
 637         bucket_index = signature & (f->n_buckets - 1);          \
 638         bucket1 = (struct rte_bucket_4_16 *)                    \
 639                 &f->memory[bucket_index * f->bucket_size];      \
 640         rte_prefetch0(bucket1);                                 \
 641         rte_prefetch0((void *)(((uintptr_t) bucket1) + RTE_CACHE_LINE_SIZE));\
 642 }
 643
 644 #define lookup1_stage2_lru(pkt2_index, mbuf2, bucket2,          \
 645                 pkts_mask_out, entries, f)                      \
 646 {                                                               \
 647         void *a;                                                \
 648         uint64_t pkt_mask;                                      \
 649         uint64_t *key;                                          \
 650         uint64_t hash_key_buffer[2];            \
 651         uint32_t pos;                                           \
 652                                                                 \
 653         key = RTE_MBUF_METADATA_UINT64_PTR(mbuf2, f->key_offset);\
 654         hash_key_buffer[0] = key[0] & f->key_mask[0];   \
 655         hash_key_buffer[1] = key[1] & f->key_mask[1];   \
 656                                                                 \
 657         lookup_key16_cmp(hash_key_buffer, bucket2, pos);        \
 658                                                                 \
 659         pkt_mask = (bucket2->signature[pos] & 1LLU) << pkt2_index;\
 660         pkts_mask_out |= pkt_mask;                              \
 661                                                                 \
 662         a = (void *) &bucket2->data[pos * f->entry_size];       \
 663         rte_prefetch0(a);                                       \
 664         entries[pkt2_index] = a;                                \
 665         lru_update(bucket2, pos);                               \
 666 }
 667
 668 #define lookup1_stage2_ext(pkt2_index, mbuf2, bucket2, pkts_mask_out, entries, \
 669         buckets_mask, buckets, keys, f)                         \
 670 {                                                               \
 671         struct rte_bucket_4_16 *bucket_next;                    \
 672         void *a;                                                \
 673         uint64_t pkt_mask, bucket_mask;                         \
 674         uint64_t *key;                                          \
 675         uint64_t hash_key_buffer[2];            \
 676         uint32_t pos;                                           \
 677                                                                 \
 678         key = RTE_MBUF_METADATA_UINT64_PTR(mbuf2, f->key_offset);\
 679         hash_key_buffer[0] = key[0] & f->key_mask[0];   \
 680         hash_key_buffer[1] = key[1] & f->key_mask[1];   \
 681                                                                 \
 682         lookup_key16_cmp(hash_key_buffer, bucket2, pos);        \
 683                                                                 \
 684         pkt_mask = (bucket2->signature[pos] & 1LLU) << pkt2_index;\
 685         pkts_mask_out |= pkt_mask;                              \
 686                                                                 \
 687         a = (void *) &bucket2->data[pos * f->entry_size];       \
 688         rte_prefetch0(a);                                       \
 689         entries[pkt2_index] = a;                                \
 690                                                                 \
 691         bucket_mask = (~pkt_mask) & (bucket2->next_valid << pkt2_index);\
 692         buckets_mask |= bucket_mask;                            \
 693         bucket_next = bucket2->next;                            \
 694         buckets[pkt2_index] = bucket_next;                      \
 695         keys[pkt2_index] = key;                                 \
 696 }
 697
 698 #define lookup_grinder(pkt_index, buckets, keys, pkts_mask_out, entries,\
 699         buckets_mask, f)                                        \
 700 {                                                               \
 701         struct rte_bucket_4_16 *bucket, *bucket_next;           \
 702         void *a;                                                \
 703         uint64_t pkt_mask, bucket_mask;                         \
 704         uint64_t *key;                                          \
 705         uint64_t hash_key_buffer[2];            \
 706         uint32_t pos;                                           \
 707                                                                 \
 708         bucket = buckets[pkt_index];                            \
 709         key = keys[pkt_index];                                  \
 710         hash_key_buffer[0] = key[0] & f->key_mask[0];   \
 711         hash_key_buffer[1] = key[1] & f->key_mask[1];   \
 712                                                                 \
 713         lookup_key16_cmp(hash_key_buffer, bucket, pos); \
 714                                                                 \
 715         pkt_mask = (bucket->signature[pos] & 1LLU) << pkt_index;\
 716         pkts_mask_out |= pkt_mask;                              \
 717                                                                 \
 718         a = (void *) &bucket->data[pos * f->entry_size];        \
 719         rte_prefetch0(a);                                       \
 720         entries[pkt_index] = a;                                 \
 721                                                                 \
 722         bucket_mask = (~pkt_mask) & (bucket->next_valid << pkt_index);\
 723         buckets_mask |= bucket_mask;                            \
 724         bucket_next = bucket->next;                             \
 725         rte_prefetch0(bucket_next);                             \
 726         rte_prefetch0((void *)(((uintptr_t) bucket_next) + RTE_CACHE_LINE_SIZE));\
 727         buckets[pkt_index] = bucket_next;                       \
 728         keys[pkt_index] = key;                                  \
 729 }
 730
 731 #define lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01,\
 732                 pkts, pkts_mask)                                \
 733 {                                                               \
 734         uint64_t pkt00_mask, pkt01_mask;                        \
 735                                                                 \
 736         pkt00_index = __builtin_ctzll(pkts_mask);               \
 737         pkt00_mask = 1LLU << pkt00_index;                       \
 738         pkts_mask &= ~pkt00_mask;                               \
 739                                                                 \
 740         mbuf00 = pkts[pkt00_index];                             \
 741         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, 0));  \
 742                                                                 \
 743         pkt01_index = __builtin_ctzll(pkts_mask);               \
 744         pkt01_mask = 1LLU << pkt01_index;                       \
 745         pkts_mask &= ~pkt01_mask;                               \
 746                                                                 \
 747         mbuf01 = pkts[pkt01_index];                             \
 748         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, 0));  \
 749 }
 750
 751 #define lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,\
 752                 mbuf00, mbuf01, pkts, pkts_mask)                \
 753 {                                                               \
 754         uint64_t pkt00_mask, pkt01_mask;                        \
 755                                                                 \
 756         pkt00_index = __builtin_ctzll(pkts_mask);               \
 757         pkt00_mask = 1LLU << pkt00_index;                       \
 758         pkts_mask &= ~pkt00_mask;                               \
 759                                                                 \
 760         mbuf00 = pkts[pkt00_index];                             \
 761         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, 0));  \
 762                                                                 \
 763         pkt01_index = __builtin_ctzll(pkts_mask);               \
 764         if (pkts_mask == 0)                                     \
 765                 pkt01_index = pkt00_index;                      \
 766         pkt01_mask = 1LLU << pkt01_index;                       \
 767         pkts_mask &= ~pkt01_mask;                               \
 768                                                                 \
 769         mbuf01 = pkts[pkt01_index];                             \
 770         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, 0));  \
 771 }
 772
 773 #define lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f)   \
 774 {                                                               \
 775         uint64_t signature10, signature11;                      \
 776         uint32_t bucket10_index, bucket11_index;                \
 777                                                                 \
 778         signature10 = RTE_MBUF_METADATA_UINT32(mbuf10, f->signature_offset);\
 779         bucket10_index = signature10 & (f->n_buckets - 1);      \
 780         bucket10 = (struct rte_bucket_4_16 *)                   \
 781                 &f->memory[bucket10_index * f->bucket_size];    \
 782         rte_prefetch0(bucket10);                                \
 783         rte_prefetch0((void *)(((uintptr_t) bucket10) + RTE_CACHE_LINE_SIZE));\
 784                                                                 \
 785         signature11 = RTE_MBUF_METADATA_UINT32(mbuf11, f->signature_offset);\
 786         bucket11_index = signature11 & (f->n_buckets - 1);      \
 787         bucket11 = (struct rte_bucket_4_16 *)                   \
 788                 &f->memory[bucket11_index * f->bucket_size];    \
 789         rte_prefetch0(bucket11);                                \
 790         rte_prefetch0((void *)(((uintptr_t) bucket11) + RTE_CACHE_LINE_SIZE));\
 791 }
 792
 793 #define lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f)     \
 794 {                                                               \
 795         uint64_t *key10, *key11;                                        \
 796         uint64_t hash_offset_buffer[2];                         \
 797         uint64_t signature10, signature11;                      \
 798         uint32_t bucket10_index, bucket11_index;        \
 799                                                                 \
 800         key10 = RTE_MBUF_METADATA_UINT64_PTR(mbuf10, f->key_offset);\
 801         hash_offset_buffer[0] = key10[0] & f->key_mask[0];      \
 802         hash_offset_buffer[1] = key10[1] & f->key_mask[1];      \
 803         signature10 = f->f_hash(hash_offset_buffer,                     \
 804                         RTE_TABLE_HASH_KEY_SIZE, f->seed);\
 805         bucket10_index = signature10 & (f->n_buckets - 1);      \
 806         bucket10 = (struct rte_bucket_4_16 *)                           \
 807                 &f->memory[bucket10_index * f->bucket_size];    \
 808         rte_prefetch0(bucket10);                                \
 809         rte_prefetch0((void *)(((uintptr_t) bucket10) + RTE_CACHE_LINE_SIZE));\
 810                                                                 \
 811         key11 = RTE_MBUF_METADATA_UINT64_PTR(mbuf11, f->key_offset);\
 812         hash_offset_buffer[0] = key11[0] & f->key_mask[0];      \
 813         hash_offset_buffer[1] = key11[1] & f->key_mask[1];      \
 814         signature11 = f->f_hash(hash_offset_buffer,                     \
 815                         RTE_TABLE_HASH_KEY_SIZE, f->seed);\
 816         bucket11_index = signature11 & (f->n_buckets - 1);      \
 817         bucket11 = (struct rte_bucket_4_16 *)                   \
 818                 &f->memory[bucket11_index * f->bucket_size];    \
 819         rte_prefetch0(bucket11);                                \
 820         rte_prefetch0((void *)(((uintptr_t) bucket11) + RTE_CACHE_LINE_SIZE));\
 821 }
 822
 823 #define lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,\
 824                 bucket20, bucket21, pkts_mask_out, entries, f)  \
 825 {                                                               \
 826         void *a20, *a21;                                        \
 827         uint64_t pkt20_mask, pkt21_mask;                        \
 828         uint64_t *key20, *key21;                                \
 829         uint64_t hash_key_buffer20[2];                  \
 830         uint64_t hash_key_buffer21[2];                  \
 831         uint32_t pos20, pos21;                                  \
 832                                                                 \
 833         key20 = RTE_MBUF_METADATA_UINT64_PTR(mbuf20, f->key_offset);\
 834         key21 = RTE_MBUF_METADATA_UINT64_PTR(mbuf21, f->key_offset);\
 835         hash_key_buffer20[0] = key20[0] & f->key_mask[0];       \
 836         hash_key_buffer20[1] = key20[1] & f->key_mask[1];       \
 837         hash_key_buffer21[0] = key21[0] & f->key_mask[0];       \
 838         hash_key_buffer21[1] = key21[1] & f->key_mask[1];       \
 839                                                                 \
 840         lookup_key16_cmp(hash_key_buffer20, bucket20, pos20);   \
 841         lookup_key16_cmp(hash_key_buffer21, bucket21, pos21);   \
 842                                                                 \
 843         pkt20_mask = (bucket20->signature[pos20] & 1LLU) << pkt20_index;\
 844         pkt21_mask = (bucket21->signature[pos21] & 1LLU) << pkt21_index;\
 845         pkts_mask_out |= pkt20_mask | pkt21_mask;                       \
 846                                                                 \
 847         a20 = (void *) &bucket20->data[pos20 * f->entry_size];  \
 848         a21 = (void *) &bucket21->data[pos21 * f->entry_size];  \
 849         rte_prefetch0(a20);                                     \
 850         rte_prefetch0(a21);                                     \
 851         entries[pkt20_index] = a20;                             \
 852         entries[pkt21_index] = a21;                             \
 853         lru_update(bucket20, pos20);                            \
 854         lru_update(bucket21, pos21);                            \
 855 }
 856
 857 #define lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21, bucket20, \
 858         bucket21, pkts_mask_out, entries, buckets_mask, buckets, keys, f) \
 859 {                                                               \
 860         struct rte_bucket_4_16 *bucket20_next, *bucket21_next;  \
 861         void *a20, *a21;                                        \
 862         uint64_t pkt20_mask, pkt21_mask, bucket20_mask, bucket21_mask;\
 863         uint64_t *key20, *key21;                                \
 864         uint64_t hash_key_buffer20[2];                  \
 865         uint64_t hash_key_buffer21[2];                  \
 866         uint32_t pos20, pos21;                                  \
 867                                                                 \
 868         key20 = RTE_MBUF_METADATA_UINT64_PTR(mbuf20, f->key_offset);\
 869         key21 = RTE_MBUF_METADATA_UINT64_PTR(mbuf21, f->key_offset);\
 870         hash_key_buffer20[0] = key20[0] & f->key_mask[0];       \
 871         hash_key_buffer20[1] = key20[1] & f->key_mask[1];       \
 872         hash_key_buffer21[0] = key21[0] & f->key_mask[0];       \
 873         hash_key_buffer21[1] = key21[1] & f->key_mask[1];       \
 874                                                                 \
 875         lookup_key16_cmp(hash_key_buffer20, bucket20, pos20);   \
 876         lookup_key16_cmp(hash_key_buffer21, bucket21, pos21);   \
 877                                                                 \
 878         pkt20_mask = (bucket20->signature[pos20] & 1LLU) << pkt20_index;\
 879         pkt21_mask = (bucket21->signature[pos21] & 1LLU) << pkt21_index;\
 880         pkts_mask_out |= pkt20_mask | pkt21_mask;               \
 881                                                                 \
 882         a20 = (void *) &bucket20->data[pos20 * f->entry_size];  \
 883         a21 = (void *) &bucket21->data[pos21 * f->entry_size];  \
 884         rte_prefetch0(a20);                                     \
 885         rte_prefetch0(a21);                                     \
 886         entries[pkt20_index] = a20;                             \
 887         entries[pkt21_index] = a21;                             \
 888                                                                 \
 889         bucket20_mask = (~pkt20_mask) & (bucket20->next_valid << pkt20_index);\
 890         bucket21_mask = (~pkt21_mask) & (bucket21->next_valid << pkt21_index);\
 891         buckets_mask |= bucket20_mask | bucket21_mask;          \
 892         bucket20_next = bucket20->next;                         \
 893         bucket21_next = bucket21->next;                         \
 894         buckets[pkt20_index] = bucket20_next;                   \
 895         buckets[pkt21_index] = bucket21_next;                   \
 896         keys[pkt20_index] = key20;                              \
 897         keys[pkt21_index] = key21;                              \
 898 }
 899
 900 static int
 901 rte_table_hash_lookup_key16_lru(
 902         void *table,
 903         struct rte_mbuf **pkts,
 904         uint64_t pkts_mask,
 905         uint64_t *lookup_hit_mask,
 906         void **entries)
 907 {
 908         struct rte_table_hash *f = (struct rte_table_hash *) table;
 909         struct rte_bucket_4_16 *bucket10, *bucket11, *bucket20, *bucket21;
 910         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
 911         uint32_t pkt00_index, pkt01_index, pkt10_index;
 912         uint32_t pkt11_index, pkt20_index, pkt21_index;
 913         uint64_t pkts_mask_out = 0;
 914
 915         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
 916         RTE_TABLE_HASH_KEY16_STATS_PKTS_IN_ADD(f, n_pkts_in);
 917
 918         /* Cannot run the pipeline with less than 5 packets */
 919         if (__builtin_popcountll(pkts_mask) < 5) {
 920                 for ( ; pkts_mask; ) {
 921                         struct rte_bucket_4_16 *bucket;
 922                         struct rte_mbuf *mbuf;
 923                         uint32_t pkt_index;
 924
 925                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
 926                         lookup1_stage1(mbuf, bucket, f);
 927                         lookup1_stage2_lru(pkt_index, mbuf, bucket,
 928                                 pkts_mask_out, entries, f);
 929                 }
 930
 931                 *lookup_hit_mask = pkts_mask_out;
 932                 RTE_TABLE_HASH_KEY16_STATS_PKTS_LOOKUP_MISS(f,
 933                         n_pkts_in - __builtin_popcountll(pkts_mask_out));
 934                 return 0;
 935         }
 936
 937         /*
 938          * Pipeline fill
 939          *
 940          */
 941         /* Pipeline stage 0 */
 942         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
 943                 pkts_mask);
 944
 945         /* Pipeline feed */
 946         mbuf10 = mbuf00;
 947         mbuf11 = mbuf01;
 948         pkt10_index = pkt00_index;
 949         pkt11_index = pkt01_index;
 950
 951         /* Pipeline stage 0 */
 952         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
 953                 pkts_mask);
 954
 955         /* Pipeline stage 1 */
 956         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
 957
 958         /*
 959          * Pipeline run
 960          *
 961          */
 962         for ( ; pkts_mask; ) {
 963                 /* Pipeline feed */
 964                 bucket20 = bucket10;
 965                 bucket21 = bucket11;
 966                 mbuf20 = mbuf10;
 967                 mbuf21 = mbuf11;
 968                 mbuf10 = mbuf00;
 969                 mbuf11 = mbuf01;
 970                 pkt20_index = pkt10_index;
 971                 pkt21_index = pkt11_index;
 972                 pkt10_index = pkt00_index;
 973                 pkt11_index = pkt01_index;
 974
 975                 /* Pipeline stage 0 */
 976                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
 977                         mbuf00, mbuf01, pkts, pkts_mask);
 978
 979                 /* Pipeline stage 1 */
 980                 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
 981
 982                 /* Pipeline stage 2 */
 983                 lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
 984                         bucket20, bucket21, pkts_mask_out, entries, f);
 985         }
 986
 987         /*
 988          * Pipeline flush
 989          *
 990          */
 991         /* Pipeline feed */
 992         bucket20 = bucket10;
 993         bucket21 = bucket11;
 994         mbuf20 = mbuf10;
 995         mbuf21 = mbuf11;
 996         mbuf10 = mbuf00;
 997         mbuf11 = mbuf01;
 998         pkt20_index = pkt10_index;
 999         pkt21_index = pkt11_index;
1000         pkt10_index = pkt00_index;
1001         pkt11_index = pkt01_index;
1002
1003         /* Pipeline stage 1 */
1004         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1005
1006         /* Pipeline stage 2 */
1007         lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
1008                 bucket20, bucket21, pkts_mask_out, entries, f);
1009
1010         /* Pipeline feed */
1011         bucket20 = bucket10;
1012         bucket21 = bucket11;
1013         mbuf20 = mbuf10;
1014         mbuf21 = mbuf11;
1015         pkt20_index = pkt10_index;
1016         pkt21_index = pkt11_index;
1017
1018         /* Pipeline stage 2 */
1019         lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
1020                 bucket20, bucket21, pkts_mask_out, entries, f);
1021
1022         *lookup_hit_mask = pkts_mask_out;
1023         RTE_TABLE_HASH_KEY16_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in -
1024                 __builtin_popcountll(pkts_mask_out));
1025         return 0;
1026 } /* rte_table_hash_lookup_key16_lru() */
1027
1028 static int
1029 rte_table_hash_lookup_key16_lru_dosig(
1030         void *table,
1031         struct rte_mbuf **pkts,
1032         uint64_t pkts_mask,
1033         uint64_t *lookup_hit_mask,
1034         void **entries)
1035 {
1036         struct rte_table_hash *f = (struct rte_table_hash *) table;
1037         struct rte_bucket_4_16 *bucket10, *bucket11, *bucket20, *bucket21;
1038         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
1039         uint32_t pkt00_index, pkt01_index, pkt10_index;
1040         uint32_t pkt11_index, pkt20_index, pkt21_index;
1041         uint64_t pkts_mask_out = 0;
1042
1043         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
1044
1045         RTE_TABLE_HASH_KEY16_STATS_PKTS_IN_ADD(f, n_pkts_in);
1046
1047         /* Cannot run the pipeline with less than 5 packets */
1048         if (__builtin_popcountll(pkts_mask) < 5) {
1049                 for ( ; pkts_mask; ) {
1050                         struct rte_bucket_4_16 *bucket;
1051                         struct rte_mbuf *mbuf;
1052                         uint32_t pkt_index;
1053
1054                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
1055                         lookup1_stage1_dosig(mbuf, bucket, f);
1056                         lookup1_stage2_lru(pkt_index, mbuf, bucket,
1057                                 pkts_mask_out, entries, f);
1058                 }
1059
1060                 *lookup_hit_mask = pkts_mask_out;
1061                 RTE_TABLE_HASH_KEY16_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in -
1062                         __builtin_popcountll(pkts_mask_out));
1063                 return 0;
1064         }
1065
1066         /*
1067          * Pipeline fill
1068          *
1069          */
1070         /* Pipeline stage 0 */
1071         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1072                 pkts_mask);
1073
1074         /* Pipeline feed */
1075         mbuf10 = mbuf00;
1076         mbuf11 = mbuf01;
1077         pkt10_index = pkt00_index;
1078         pkt11_index = pkt01_index;
1079
1080         /* Pipeline stage 0 */
1081         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1082                 pkts_mask);
1083
1084         /* Pipeline stage 1 */
1085         lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1086
1087         /*
1088          * Pipeline run
1089          *
1090          */
1091         for ( ; pkts_mask; ) {
1092                 /* Pipeline feed */
1093                 bucket20 = bucket10;
1094                 bucket21 = bucket11;
1095                 mbuf20 = mbuf10;
1096                 mbuf21 = mbuf11;
1097                 mbuf10 = mbuf00;
1098                 mbuf11 = mbuf01;
1099                 pkt20_index = pkt10_index;
1100                 pkt21_index = pkt11_index;
1101                 pkt10_index = pkt00_index;
1102                 pkt11_index = pkt01_index;
1103
1104                 /* Pipeline stage 0 */
1105                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
1106                         mbuf00, mbuf01, pkts, pkts_mask);
1107
1108                 /* Pipeline stage 1 */
1109                 lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1110
1111                 /* Pipeline stage 2 */
1112                 lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
1113                         bucket20, bucket21, pkts_mask_out, entries, f);
1114         }
1115
1116         /*
1117          * Pipeline flush
1118          *
1119          */
1120         /* Pipeline feed */
1121         bucket20 = bucket10;
1122         bucket21 = bucket11;
1123         mbuf20 = mbuf10;
1124         mbuf21 = mbuf11;
1125         mbuf10 = mbuf00;
1126         mbuf11 = mbuf01;
1127         pkt20_index = pkt10_index;
1128         pkt21_index = pkt11_index;
1129         pkt10_index = pkt00_index;
1130         pkt11_index = pkt01_index;
1131
1132         /* Pipeline stage 1 */
1133         lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1134
1135         /* Pipeline stage 2 */
1136         lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
1137                 bucket20, bucket21, pkts_mask_out, entries, f);
1138
1139         /* Pipeline feed */
1140         bucket20 = bucket10;
1141         bucket21 = bucket11;
1142         mbuf20 = mbuf10;
1143         mbuf21 = mbuf11;
1144         pkt20_index = pkt10_index;
1145         pkt21_index = pkt11_index;
1146
1147         /* Pipeline stage 2 */
1148         lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
1149                 bucket20, bucket21, pkts_mask_out, entries, f);
1150
1151         *lookup_hit_mask = pkts_mask_out;
1152         RTE_TABLE_HASH_KEY16_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in -
1153                 __builtin_popcountll(pkts_mask_out));
1154         return 0;
1155 } /* rte_table_hash_lookup_key16_lru_dosig() */
1156
1157 static int
1158 rte_table_hash_lookup_key16_ext(
1159         void *table,
1160         struct rte_mbuf **pkts,
1161         uint64_t pkts_mask,
1162         uint64_t *lookup_hit_mask,
1163         void **entries)
1164 {
1165         struct rte_table_hash *f = (struct rte_table_hash *) table;
1166         struct rte_bucket_4_16 *bucket10, *bucket11, *bucket20, *bucket21;
1167         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
1168         uint32_t pkt00_index, pkt01_index, pkt10_index;
1169         uint32_t pkt11_index, pkt20_index, pkt21_index;
1170         uint64_t pkts_mask_out = 0, buckets_mask = 0;
1171         struct rte_bucket_4_16 *buckets[RTE_PORT_IN_BURST_SIZE_MAX];
1172         uint64_t *keys[RTE_PORT_IN_BURST_SIZE_MAX];
1173
1174         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
1175         RTE_TABLE_HASH_KEY16_STATS_PKTS_IN_ADD(f, n_pkts_in);
1176
1177         /* Cannot run the pipeline with less than 5 packets */
1178         if (__builtin_popcountll(pkts_mask) < 5) {
1179                 for ( ; pkts_mask; ) {
1180                         struct rte_bucket_4_16 *bucket;
1181                         struct rte_mbuf *mbuf;
1182                         uint32_t pkt_index;
1183
1184                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
1185                         lookup1_stage1(mbuf, bucket, f);
1186                         lookup1_stage2_ext(pkt_index, mbuf, bucket,
1187                                 pkts_mask_out, entries, buckets_mask,
1188                                 buckets, keys, f);
1189                 }
1190
1191                 goto grind_next_buckets;
1192         }
1193
1194         /*
1195          * Pipeline fill
1196          *
1197          */
1198         /* Pipeline stage 0 */
1199         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1200                 pkts_mask);
1201
1202         /* Pipeline feed */
1203         mbuf10 = mbuf00;
1204         mbuf11 = mbuf01;
1205         pkt10_index = pkt00_index;
1206         pkt11_index = pkt01_index;
1207
1208         /* Pipeline stage 0 */
1209         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1210                 pkts_mask);
1211
1212         /* Pipeline stage 1 */
1213         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1214
1215         /*
1216          * Pipeline run
1217          *
1218          */
1219         for ( ; pkts_mask; ) {
1220                 /* Pipeline feed */
1221                 bucket20 = bucket10;
1222                 bucket21 = bucket11;
1223                 mbuf20 = mbuf10;
1224                 mbuf21 = mbuf11;
1225                 mbuf10 = mbuf00;
1226                 mbuf11 = mbuf01;
1227                 pkt20_index = pkt10_index;
1228                 pkt21_index = pkt11_index;
1229                 pkt10_index = pkt00_index;
1230                 pkt11_index = pkt01_index;
1231
1232                 /* Pipeline stage 0 */
1233                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
1234                         mbuf00, mbuf01, pkts, pkts_mask);
1235
1236                 /* Pipeline stage 1 */
1237                 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1238
1239                 /* Pipeline stage 2 */
1240                 lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1241                         bucket20, bucket21, pkts_mask_out, entries,
1242                         buckets_mask, buckets, keys, f);
1243         }
1244
1245         /*
1246          * Pipeline flush
1247          *
1248          */
1249         /* Pipeline feed */
1250         bucket20 = bucket10;
1251         bucket21 = bucket11;
1252         mbuf20 = mbuf10;
1253         mbuf21 = mbuf11;
1254         mbuf10 = mbuf00;
1255         mbuf11 = mbuf01;
1256         pkt20_index = pkt10_index;
1257         pkt21_index = pkt11_index;
1258         pkt10_index = pkt00_index;
1259         pkt11_index = pkt01_index;
1260
1261         /* Pipeline stage 1 */
1262         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1263
1264         /* Pipeline stage 2 */
1265         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1266                 bucket20, bucket21, pkts_mask_out, entries,
1267                 buckets_mask, buckets, keys, f);
1268
1269         /* Pipeline feed */
1270         bucket20 = bucket10;
1271         bucket21 = bucket11;
1272         mbuf20 = mbuf10;
1273         mbuf21 = mbuf11;
1274         pkt20_index = pkt10_index;
1275         pkt21_index = pkt11_index;
1276
1277         /* Pipeline stage 2 */
1278         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1279                 bucket20, bucket21, pkts_mask_out, entries,
1280                 buckets_mask, buckets, keys, f);
1281
1282 grind_next_buckets:
1283         /* Grind next buckets */
1284         for ( ; buckets_mask; ) {
1285                 uint64_t buckets_mask_next = 0;
1286
1287                 for ( ; buckets_mask; ) {
1288                         uint64_t pkt_mask;
1289                         uint32_t pkt_index;
1290
1291                         pkt_index = __builtin_ctzll(buckets_mask);
1292                         pkt_mask = 1LLU << pkt_index;
1293                         buckets_mask &= ~pkt_mask;
1294
1295                         lookup_grinder(pkt_index, buckets, keys, pkts_mask_out,
1296                                 entries, buckets_mask_next, f);
1297                 }
1298
1299                 buckets_mask = buckets_mask_next;
1300         }
1301
1302         *lookup_hit_mask = pkts_mask_out;
1303         RTE_TABLE_HASH_KEY16_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in -
1304                 __builtin_popcountll(pkts_mask_out));
1305         return 0;
1306 } /* rte_table_hash_lookup_key16_ext() */
1307
1308 static int
1309 rte_table_hash_lookup_key16_ext_dosig(
1310         void *table,
1311         struct rte_mbuf **pkts,
1312         uint64_t pkts_mask,
1313         uint64_t *lookup_hit_mask,
1314         void **entries)
1315 {
1316         struct rte_table_hash *f = (struct rte_table_hash *) table;
1317         struct rte_bucket_4_16 *bucket10, *bucket11, *bucket20, *bucket21;
1318         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
1319         uint32_t pkt00_index, pkt01_index, pkt10_index;
1320         uint32_t pkt11_index, pkt20_index, pkt21_index;
1321         uint64_t pkts_mask_out = 0, buckets_mask = 0;
1322         struct rte_bucket_4_16 *buckets[RTE_PORT_IN_BURST_SIZE_MAX];
1323         uint64_t *keys[RTE_PORT_IN_BURST_SIZE_MAX];
1324
1325         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
1326
1327         RTE_TABLE_HASH_KEY16_STATS_PKTS_IN_ADD(f, n_pkts_in);
1328
1329         /* Cannot run the pipeline with less than 5 packets */
1330         if (__builtin_popcountll(pkts_mask) < 5) {
1331                 for ( ; pkts_mask; ) {
1332                         struct rte_bucket_4_16 *bucket;
1333                         struct rte_mbuf *mbuf;
1334                         uint32_t pkt_index;
1335
1336                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
1337                         lookup1_stage1_dosig(mbuf, bucket, f);
1338                         lookup1_stage2_ext(pkt_index, mbuf, bucket,
1339                                 pkts_mask_out, entries, buckets_mask,
1340                                 buckets, keys, f);
1341                 }
1342
1343                 goto grind_next_buckets;
1344         }
1345
1346         /*
1347          * Pipeline fill
1348          *
1349          */
1350         /* Pipeline stage 0 */
1351         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1352                 pkts_mask);
1353
1354         /* Pipeline feed */
1355         mbuf10 = mbuf00;
1356         mbuf11 = mbuf01;
1357         pkt10_index = pkt00_index;
1358         pkt11_index = pkt01_index;
1359
1360         /* Pipeline stage 0 */
1361         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1362                 pkts_mask);
1363
1364         /* Pipeline stage 1 */
1365         lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1366
1367         /*
1368          * Pipeline run
1369          *
1370          */
1371         for ( ; pkts_mask; ) {
1372                 /* Pipeline feed */
1373                 bucket20 = bucket10;
1374                 bucket21 = bucket11;
1375                 mbuf20 = mbuf10;
1376                 mbuf21 = mbuf11;
1377                 mbuf10 = mbuf00;
1378                 mbuf11 = mbuf01;
1379                 pkt20_index = pkt10_index;
1380                 pkt21_index = pkt11_index;
1381                 pkt10_index = pkt00_index;
1382                 pkt11_index = pkt01_index;
1383
1384                 /* Pipeline stage 0 */
1385                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
1386                         mbuf00, mbuf01, pkts, pkts_mask);
1387
1388                 /* Pipeline stage 1 */
1389                 lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1390
1391                 /* Pipeline stage 2 */
1392                 lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1393                         bucket20, bucket21, pkts_mask_out, entries,
1394                         buckets_mask, buckets, keys, f);
1395         }
1396
1397         /*
1398          * Pipeline flush
1399          *
1400          */
1401         /* Pipeline feed */
1402         bucket20 = bucket10;
1403         bucket21 = bucket11;
1404         mbuf20 = mbuf10;
1405         mbuf21 = mbuf11;
1406         mbuf10 = mbuf00;
1407         mbuf11 = mbuf01;
1408         pkt20_index = pkt10_index;
1409         pkt21_index = pkt11_index;
1410         pkt10_index = pkt00_index;
1411         pkt11_index = pkt01_index;
1412
1413         /* Pipeline stage 1 */
1414         lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1415
1416         /* Pipeline stage 2 */
1417         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1418                 bucket20, bucket21, pkts_mask_out, entries,
1419                 buckets_mask, buckets, keys, f);
1420
1421         /* Pipeline feed */
1422         bucket20 = bucket10;
1423         bucket21 = bucket11;
1424         mbuf20 = mbuf10;
1425         mbuf21 = mbuf11;
1426         pkt20_index = pkt10_index;
1427         pkt21_index = pkt11_index;
1428
1429         /* Pipeline stage 2 */
1430         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1431                 bucket20, bucket21, pkts_mask_out, entries,
1432                 buckets_mask, buckets, keys, f);
1433
1434 grind_next_buckets:
1435         /* Grind next buckets */
1436         for ( ; buckets_mask; ) {
1437                 uint64_t buckets_mask_next = 0;
1438
1439                 for ( ; buckets_mask; ) {
1440                         uint64_t pkt_mask;
1441                         uint32_t pkt_index;
1442
1443                         pkt_index = __builtin_ctzll(buckets_mask);
1444                         pkt_mask = 1LLU << pkt_index;
1445                         buckets_mask &= ~pkt_mask;
1446
1447                         lookup_grinder(pkt_index, buckets, keys, pkts_mask_out,
1448                                 entries, buckets_mask_next, f);
1449                 }
1450
1451                 buckets_mask = buckets_mask_next;
1452         }
1453
1454         *lookup_hit_mask = pkts_mask_out;
1455         RTE_TABLE_HASH_KEY16_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in -
1456                 __builtin_popcountll(pkts_mask_out));
1457         return 0;
1458 } /* rte_table_hash_lookup_key16_ext_dosig() */
1459
1460 static int
1461 rte_table_hash_key16_stats_read(void *table, struct rte_table_stats *stats, int clear)
1462 {
1463         struct rte_table_hash *t = (struct rte_table_hash *) table;
1464
1465         if (stats != NULL)
1466                 memcpy(stats, &t->stats, sizeof(t->stats));
1467
1468         if (clear)
1469                 memset(&t->stats, 0, sizeof(t->stats));
1470
1471         return 0;
1472 }
1473
1474 struct rte_table_ops rte_table_hash_key16_lru_ops = {
1475         .f_create = rte_table_hash_create_key16_lru,
1476         .f_free = rte_table_hash_free_key16_lru,
1477         .f_add = rte_table_hash_entry_add_key16_lru,
1478         .f_delete = rte_table_hash_entry_delete_key16_lru,
1479         .f_add_bulk = NULL,
1480         .f_delete_bulk = NULL,
1481         .f_lookup = rte_table_hash_lookup_key16_lru,
1482         .f_stats = rte_table_hash_key16_stats_read,
1483 };
1484
1485 struct rte_table_ops rte_table_hash_key16_lru_dosig_ops = {
1486         .f_create = rte_table_hash_create_key16_lru,
1487         .f_free = rte_table_hash_free_key16_lru,
1488         .f_add = rte_table_hash_entry_add_key16_lru,
1489         .f_delete = rte_table_hash_entry_delete_key16_lru,
1490         .f_lookup = rte_table_hash_lookup_key16_lru_dosig,
1491         .f_stats = rte_table_hash_key16_stats_read,
1492 };
1493
1494 struct rte_table_ops rte_table_hash_key16_ext_ops = {
1495         .f_create = rte_table_hash_create_key16_ext,
1496         .f_free = rte_table_hash_free_key16_ext,
1497         .f_add = rte_table_hash_entry_add_key16_ext,
1498         .f_delete = rte_table_hash_entry_delete_key16_ext,
1499         .f_add_bulk = NULL,
1500         .f_delete_bulk = NULL,
1501         .f_lookup = rte_table_hash_lookup_key16_ext,
1502         .f_stats = rte_table_hash_key16_stats_read,
1503 };
1504
1505 struct rte_table_ops rte_table_hash_key16_ext_dosig_ops = {
1506         .f_create = rte_table_hash_create_key16_ext,
1507         .f_free = rte_table_hash_free_key16_ext,
1508         .f_add = rte_table_hash_entry_add_key16_ext,
1509         .f_delete = rte_table_hash_entry_delete_key16_ext,
1510         .f_lookup = rte_table_hash_lookup_key16_ext_dosig,
1511         .f_stats = rte_table_hash_key16_stats_read,
1512 };