lib/librte_table/rte_table_hash_key8.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                                         8
  46
  47 #ifdef RTE_TABLE_STATS_COLLECT
  48
  49 #define RTE_TABLE_HASH_KEY8_STATS_PKTS_IN_ADD(table, val) \
  50         table->stats.n_pkts_in += val
  51 #define RTE_TABLE_HASH_KEY8_STATS_PKTS_LOOKUP_MISS(table, val) \
  52         table->stats.n_pkts_lookup_miss += val
  53
  54 #else
  55
  56 #define RTE_TABLE_HASH_KEY8_STATS_PKTS_IN_ADD(table, val)
  57 #define RTE_TABLE_HASH_KEY8_STATS_PKTS_LOOKUP_MISS(table, val)
  58
  59 #endif
  60
  61 struct rte_bucket_4_8 {
  62         /* Cache line 0 */
  63         uint64_t signature;
  64         uint64_t lru_list;
  65         struct rte_bucket_4_8 *next;
  66         uint64_t next_valid;
  67
  68         uint64_t key[4];
  69
  70         /* Cache line 1 */
  71         uint8_t data[0];
  72 };
  73
  74 struct rte_table_hash {
  75         struct rte_table_stats stats;
  76
  77         /* Input parameters */
  78         uint32_t n_buckets;
  79         uint32_t n_entries_per_bucket;
  80         uint32_t key_size;
  81         uint32_t entry_size;
  82         uint32_t bucket_size;
  83         uint32_t signature_offset;
  84         uint32_t key_offset;
  85         uint64_t key_mask;
  86         rte_table_hash_op_hash f_hash;
  87         uint64_t seed;
  88
  89         /* Extendible buckets */
  90         uint32_t n_buckets_ext;
  91         uint32_t stack_pos;
  92         uint32_t *stack;
  93
  94         /* Lookup table */
  95         uint8_t memory[0] __rte_cache_aligned;
  96 };
  97
  98 static int
  99 check_params_create_lru(struct rte_table_hash_key8_lru_params *params) {
 100         /* n_entries */
 101         if (params->n_entries == 0) {
 102                 RTE_LOG(ERR, TABLE, "%s: n_entries is zero\n", __func__);
 103                 return -EINVAL;
 104         }
 105
 106         /* f_hash */
 107         if (params->f_hash == NULL) {
 108                 RTE_LOG(ERR, TABLE, "%s: f_hash function pointer is NULL\n",
 109                         __func__);
 110                 return -EINVAL;
 111         }
 112
 113         return 0;
 114 }
 115
 116 static void *
 117 rte_table_hash_create_key8_lru(void *params, int socket_id, uint32_t entry_size)
 118 {
 119         struct rte_table_hash_key8_lru_params *p =
 120                 (struct rte_table_hash_key8_lru_params *) params;
 121         struct rte_table_hash *f;
 122         uint32_t n_buckets, n_entries_per_bucket, key_size, bucket_size_cl;
 123         uint32_t total_size, i;
 124
 125         /* Check input parameters */
 126         if ((check_params_create_lru(p) != 0) ||
 127                 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
 128                 ((sizeof(struct rte_bucket_4_8) % RTE_CACHE_LINE_SIZE) != 0)) {
 129                 return NULL;
 130         }
 131         n_entries_per_bucket = 4;
 132         key_size = 8;
 133
 134         /* Memory allocation */
 135         n_buckets = rte_align32pow2((p->n_entries + n_entries_per_bucket - 1) /
 136                 n_entries_per_bucket);
 137         bucket_size_cl = (sizeof(struct rte_bucket_4_8) + n_entries_per_bucket *
 138                 entry_size + RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
 139         total_size = sizeof(struct rte_table_hash) + n_buckets *
 140                 bucket_size_cl * RTE_CACHE_LINE_SIZE;
 141
 142         f = rte_zmalloc_socket("TABLE", total_size, RTE_CACHE_LINE_SIZE, socket_id);
 143         if (f == NULL) {
 144                 RTE_LOG(ERR, TABLE,
 145                         "%s: Cannot allocate %u bytes for hash table\n",
 146                         __func__, total_size);
 147                 return NULL;
 148         }
 149         RTE_LOG(INFO, TABLE,
 150                 "%s: Hash table memory footprint is %u bytes\n",
 151                 __func__, total_size);
 152
 153         /* Memory initialization */
 154         f->n_buckets = n_buckets;
 155         f->n_entries_per_bucket = n_entries_per_bucket;
 156         f->key_size = key_size;
 157         f->entry_size = entry_size;
 158         f->bucket_size = bucket_size_cl * RTE_CACHE_LINE_SIZE;
 159         f->signature_offset = p->signature_offset;
 160         f->key_offset = p->key_offset;
 161         f->f_hash = p->f_hash;
 162         f->seed = p->seed;
 163
 164         if (p->key_mask != NULL)
 165                 f->key_mask = ((uint64_t *)p->key_mask)[0];
 166         else
 167                 f->key_mask = 0xFFFFFFFFFFFFFFFFLLU;
 168
 169         for (i = 0; i < n_buckets; i++) {
 170                 struct rte_bucket_4_8 *bucket;
 171
 172                 bucket = (struct rte_bucket_4_8 *) &f->memory[i *
 173                         f->bucket_size];
 174                 bucket->lru_list = 0x0000000100020003LLU;
 175         }
 176
 177         return f;
 178 }
 179
 180 static int
 181 rte_table_hash_free_key8_lru(void *table)
 182 {
 183         struct rte_table_hash *f = (struct rte_table_hash *) table;
 184
 185         /* Check input parameters */
 186         if (f == NULL) {
 187                 RTE_LOG(ERR, TABLE, "%s: table parameter is NULL\n", __func__);
 188                 return -EINVAL;
 189         }
 190
 191         rte_free(f);
 192         return 0;
 193 }
 194
 195 static int
 196 rte_table_hash_entry_add_key8_lru(
 197         void *table,
 198         void *key,
 199         void *entry,
 200         int *key_found,
 201         void **entry_ptr)
 202 {
 203         struct rte_table_hash *f = (struct rte_table_hash *) table;
 204         struct rte_bucket_4_8 *bucket;
 205         uint64_t signature, mask, pos;
 206         uint32_t bucket_index, i;
 207
 208         signature = f->f_hash(key, f->key_size, f->seed);
 209         bucket_index = signature & (f->n_buckets - 1);
 210         bucket = (struct rte_bucket_4_8 *)
 211                 &f->memory[bucket_index * f->bucket_size];
 212
 213         /* Key is present in the bucket */
 214         for (i = 0, mask = 1LLU; i < 4; i++, mask <<= 1) {
 215                 uint64_t bucket_signature = bucket->signature;
 216                 uint64_t bucket_key = bucket->key[i];
 217
 218                 if ((bucket_signature & mask) &&
 219                     (*((uint64_t *) key) == bucket_key)) {
 220                         uint8_t *bucket_data = &bucket->data[i * f->entry_size];
 221
 222                         memcpy(bucket_data, entry, f->entry_size);
 223                         lru_update(bucket, i);
 224                         *key_found = 1;
 225                         *entry_ptr = (void *) bucket_data;
 226                         return 0;
 227                 }
 228         }
 229
 230         /* Key is not present in the bucket */
 231         for (i = 0, mask = 1LLU; i < 4; i++, mask <<= 1) {
 232                 uint64_t bucket_signature = bucket->signature;
 233
 234                 if ((bucket_signature & mask) == 0) {
 235                         uint8_t *bucket_data = &bucket->data[i * f->entry_size];
 236
 237                         bucket->signature |= mask;
 238                         bucket->key[i] = *((uint64_t *) key);
 239                         memcpy(bucket_data, entry, f->entry_size);
 240                         lru_update(bucket, i);
 241                         *key_found = 0;
 242                         *entry_ptr = (void *) bucket_data;
 243
 244                         return 0;
 245                 }
 246         }
 247
 248         /* Bucket full: replace LRU entry */
 249         pos = lru_pos(bucket);
 250         bucket->key[pos] = *((uint64_t *) key);
 251         memcpy(&bucket->data[pos * f->entry_size], entry, f->entry_size);
 252         lru_update(bucket, pos);
 253         *key_found      = 0;
 254         *entry_ptr = (void *) &bucket->data[pos * f->entry_size];
 255
 256         return 0;
 257 }
 258
 259 static int
 260 rte_table_hash_entry_delete_key8_lru(
 261         void *table,
 262         void *key,
 263         int *key_found,
 264         void *entry)
 265 {
 266         struct rte_table_hash *f = (struct rte_table_hash *) table;
 267         struct rte_bucket_4_8 *bucket;
 268         uint64_t signature, mask;
 269         uint32_t bucket_index, i;
 270
 271         signature = f->f_hash(key, f->key_size, f->seed);
 272         bucket_index = signature & (f->n_buckets - 1);
 273         bucket = (struct rte_bucket_4_8 *)
 274                 &f->memory[bucket_index * f->bucket_size];
 275
 276         /* Key is present in the bucket */
 277         for (i = 0, mask = 1LLU; i < 4; i++, mask <<= 1) {
 278                 uint64_t bucket_signature = bucket->signature;
 279                 uint64_t bucket_key = bucket->key[i];
 280
 281                 if ((bucket_signature & mask) &&
 282                     (*((uint64_t *) key) == bucket_key)) {
 283                         uint8_t *bucket_data = &bucket->data[i * f->entry_size];
 284
 285                         bucket->signature &= ~mask;
 286                         *key_found = 1;
 287                         if (entry)
 288                                 memcpy(entry, bucket_data, f->entry_size);
 289
 290                         return 0;
 291                 }
 292         }
 293
 294         /* Key is not present in the bucket */
 295         *key_found = 0;
 296         return 0;
 297 }
 298
 299 static int
 300 check_params_create_ext(struct rte_table_hash_key8_ext_params *params) {
 301         /* n_entries */
 302         if (params->n_entries == 0) {
 303                 RTE_LOG(ERR, TABLE, "%s: n_entries is zero\n", __func__);
 304                 return -EINVAL;
 305         }
 306
 307         /* n_entries_ext */
 308         if (params->n_entries_ext == 0) {
 309                 RTE_LOG(ERR, TABLE, "%s: n_entries_ext is zero\n", __func__);
 310                 return -EINVAL;
 311         }
 312
 313         /* f_hash */
 314         if (params->f_hash == NULL) {
 315                 RTE_LOG(ERR, TABLE, "%s: f_hash function pointer is NULL\n",
 316                         __func__);
 317                 return -EINVAL;
 318         }
 319
 320         return 0;
 321 }
 322
 323 static void *
 324 rte_table_hash_create_key8_ext(void *params, int socket_id, uint32_t entry_size)
 325 {
 326         struct rte_table_hash_key8_ext_params *p =
 327                 (struct rte_table_hash_key8_ext_params *) params;
 328         struct rte_table_hash *f;
 329         uint32_t n_buckets, n_buckets_ext, n_entries_per_bucket, key_size;
 330         uint32_t bucket_size_cl, stack_size_cl, total_size, i;
 331
 332         /* Check input parameters */
 333         if ((check_params_create_ext(p) != 0) ||
 334                 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
 335                 ((sizeof(struct rte_bucket_4_8) % RTE_CACHE_LINE_SIZE) != 0))
 336                 return NULL;
 337
 338         n_entries_per_bucket = 4;
 339         key_size = 8;
 340
 341         /* Memory allocation */
 342         n_buckets = rte_align32pow2((p->n_entries + n_entries_per_bucket - 1) /
 343                 n_entries_per_bucket);
 344         n_buckets_ext = (p->n_entries_ext + n_entries_per_bucket - 1) /
 345                 n_entries_per_bucket;
 346         bucket_size_cl = (sizeof(struct rte_bucket_4_8) + n_entries_per_bucket *
 347                 entry_size + RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
 348         stack_size_cl = (n_buckets_ext * sizeof(uint32_t) + RTE_CACHE_LINE_SIZE - 1)
 349                 / RTE_CACHE_LINE_SIZE;
 350         total_size = sizeof(struct rte_table_hash) + ((n_buckets +
 351                 n_buckets_ext) * bucket_size_cl + stack_size_cl) *
 352                 RTE_CACHE_LINE_SIZE;
 353
 354         f = rte_zmalloc_socket("TABLE", total_size, RTE_CACHE_LINE_SIZE, socket_id);
 355         if (f == NULL) {
 356                 RTE_LOG(ERR, TABLE,
 357                         "%s: Cannot allocate %u bytes for hash table\n",
 358                         __func__, total_size);
 359                 return NULL;
 360         }
 361         RTE_LOG(INFO, TABLE,
 362                 "%s: Hash table memory footprint is %u bytes\n",
 363                 __func__, total_size);
 364
 365         /* Memory initialization */
 366         f->n_buckets = n_buckets;
 367         f->n_entries_per_bucket = n_entries_per_bucket;
 368         f->key_size = key_size;
 369         f->entry_size = entry_size;
 370         f->bucket_size = bucket_size_cl * RTE_CACHE_LINE_SIZE;
 371         f->signature_offset = p->signature_offset;
 372         f->key_offset = p->key_offset;
 373         f->f_hash = p->f_hash;
 374         f->seed = p->seed;
 375
 376         f->n_buckets_ext = n_buckets_ext;
 377         f->stack_pos = n_buckets_ext;
 378         f->stack = (uint32_t *)
 379                 &f->memory[(n_buckets + n_buckets_ext) * f->bucket_size];
 380
 381         if (p->key_mask != NULL)
 382                 f->key_mask = ((uint64_t *)p->key_mask)[0];
 383         else
 384                 f->key_mask = 0xFFFFFFFFFFFFFFFFLLU;
 385
 386         for (i = 0; i < n_buckets_ext; i++)
 387                 f->stack[i] = i;
 388
 389         return f;
 390 }
 391
 392 static int
 393 rte_table_hash_free_key8_ext(void *table)
 394 {
 395         struct rte_table_hash *f = (struct rte_table_hash *) table;
 396
 397         /* Check input parameters */
 398         if (f == NULL) {
 399                 RTE_LOG(ERR, TABLE, "%s: table parameter is NULL\n", __func__);
 400                 return -EINVAL;
 401         }
 402
 403         rte_free(f);
 404         return 0;
 405 }
 406
 407 static int
 408 rte_table_hash_entry_add_key8_ext(
 409         void *table,
 410         void *key,
 411         void *entry,
 412         int *key_found,
 413         void **entry_ptr)
 414 {
 415         struct rte_table_hash *f = (struct rte_table_hash *) table;
 416         struct rte_bucket_4_8 *bucket0, *bucket, *bucket_prev;
 417         uint64_t signature;
 418         uint32_t bucket_index, i;
 419
 420         signature = f->f_hash(key, f->key_size, f->seed);
 421         bucket_index = signature & (f->n_buckets - 1);
 422         bucket0 = (struct rte_bucket_4_8 *)
 423                 &f->memory[bucket_index * f->bucket_size];
 424
 425         /* Key is present in the bucket */
 426         for (bucket = bucket0; bucket != NULL; bucket = bucket->next) {
 427                 uint64_t mask;
 428
 429                 for (i = 0, mask = 1LLU; i < 4; i++, mask <<= 1) {
 430                         uint64_t bucket_signature = bucket->signature;
 431                         uint64_t bucket_key = bucket->key[i];
 432
 433                         if ((bucket_signature & mask) &&
 434                                         (*((uint64_t *) key) == bucket_key)) {
 435                                 uint8_t *bucket_data = &bucket->data[i *
 436                                         f->entry_size];
 437
 438                                 memcpy(bucket_data, entry, f->entry_size);
 439                                 *key_found = 1;
 440                                 *entry_ptr = (void *) bucket_data;
 441                                 return 0;
 442                         }
 443                 }
 444         }
 445
 446         /* Key is not present in the bucket */
 447         for (bucket_prev = NULL, bucket = bucket0;
 448                 bucket != NULL; bucket_prev = bucket, bucket = bucket->next) {
 449                 uint64_t mask;
 450
 451                 for (i = 0, mask = 1LLU; i < 4; i++, mask <<= 1) {
 452                         uint64_t bucket_signature = bucket->signature;
 453
 454                         if ((bucket_signature & mask) == 0) {
 455                                 uint8_t *bucket_data = &bucket->data[i *
 456                                         f->entry_size];
 457
 458                                 bucket->signature |= mask;
 459                                 bucket->key[i] = *((uint64_t *) key);
 460                                 memcpy(bucket_data, entry, f->entry_size);
 461                                 *key_found = 0;
 462                                 *entry_ptr = (void *) bucket_data;
 463
 464                                 return 0;
 465                         }
 466                 }
 467         }
 468
 469         /* Bucket full: extend bucket */
 470         if (f->stack_pos > 0) {
 471                 bucket_index = f->stack[--f->stack_pos];
 472
 473                 bucket = (struct rte_bucket_4_8 *) &f->memory[(f->n_buckets +
 474                         bucket_index) * f->bucket_size];
 475                 bucket_prev->next = bucket;
 476                 bucket_prev->next_valid = 1;
 477
 478                 bucket->signature = 1;
 479                 bucket->key[0] = *((uint64_t *) key);
 480                 memcpy(&bucket->data[0], entry, f->entry_size);
 481                 *key_found = 0;
 482                 *entry_ptr = (void *) &bucket->data[0];
 483                 return 0;
 484         }
 485
 486         return -ENOSPC;
 487 }
 488
 489 static int
 490 rte_table_hash_entry_delete_key8_ext(
 491         void *table,
 492         void *key,
 493         int *key_found,
 494         void *entry)
 495 {
 496         struct rte_table_hash *f = (struct rte_table_hash *) table;
 497         struct rte_bucket_4_8 *bucket0, *bucket, *bucket_prev;
 498         uint64_t signature;
 499         uint32_t bucket_index, i;
 500
 501         signature = f->f_hash(key, f->key_size, f->seed);
 502         bucket_index = signature & (f->n_buckets - 1);
 503         bucket0 = (struct rte_bucket_4_8 *)
 504                 &f->memory[bucket_index * f->bucket_size];
 505
 506         /* Key is present in the bucket */
 507         for (bucket_prev = NULL, bucket = bucket0; bucket != NULL;
 508                 bucket_prev = bucket, bucket = bucket->next) {
 509                 uint64_t mask;
 510
 511                 for (i = 0, mask = 1LLU; i < 4; i++, mask <<= 1) {
 512                         uint64_t bucket_signature = bucket->signature;
 513                         uint64_t bucket_key = bucket->key[i];
 514
 515                         if ((bucket_signature & mask) &&
 516                                 (*((uint64_t *) key) == bucket_key)) {
 517                                 uint8_t *bucket_data = &bucket->data[i *
 518                                         f->entry_size];
 519
 520                                 bucket->signature &= ~mask;
 521                                 *key_found = 1;
 522                                 if (entry)
 523                                         memcpy(entry, bucket_data,
 524                                                 f->entry_size);
 525
 526                                 if ((bucket->signature == 0) &&
 527                                     (bucket_prev != NULL)) {
 528                                         bucket_prev->next = bucket->next;
 529                                         bucket_prev->next_valid =
 530                                                 bucket->next_valid;
 531
 532                                         memset(bucket, 0,
 533                                                 sizeof(struct rte_bucket_4_8));
 534                                         bucket_index = (((uint8_t *)bucket -
 535                                                 (uint8_t *)f->memory)/f->bucket_size) - f->n_buckets;
 536                                         f->stack[f->stack_pos++] = bucket_index;
 537                                 }
 538
 539                                 return 0;
 540                         }
 541                 }
 542         }
 543
 544         /* Key is not present in the bucket */
 545         *key_found = 0;
 546         return 0;
 547 }
 548
 549 #define lookup_key8_cmp(key_in, bucket, pos)                    \
 550 {                                                               \
 551         uint64_t xor[4], signature;                             \
 552                                                                 \
 553         signature = ~bucket->signature;                         \
 554                                                                 \
 555         xor[0] = (key_in[0] ^    bucket->key[0]) | (signature & 1);\
 556         xor[1] = (key_in[0] ^    bucket->key[1]) | (signature & 2);\
 557         xor[2] = (key_in[0] ^    bucket->key[2]) | (signature & 4);\
 558         xor[3] = (key_in[0] ^    bucket->key[3]) | (signature & 8);\
 559                                                                 \
 560         pos = 4;                                                \
 561         if (xor[0] == 0)                                        \
 562                 pos = 0;                                        \
 563         if (xor[1] == 0)                                        \
 564                 pos = 1;                                        \
 565         if (xor[2] == 0)                                        \
 566                 pos = 2;                                        \
 567         if (xor[3] == 0)                                        \
 568                 pos = 3;                                        \
 569 }
 570
 571 #define lookup1_stage0(pkt0_index, mbuf0, pkts, pkts_mask)      \
 572 {                                                               \
 573         uint64_t pkt_mask;                                      \
 574                                                                 \
 575         pkt0_index = __builtin_ctzll(pkts_mask);                \
 576         pkt_mask = 1LLU << pkt0_index;                          \
 577         pkts_mask &= ~pkt_mask;                                 \
 578                                                                 \
 579         mbuf0 = pkts[pkt0_index];                               \
 580         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf0, 0));   \
 581 }
 582
 583 #define lookup1_stage1(mbuf1, bucket1, f)                       \
 584 {                                                               \
 585         uint64_t signature;                                     \
 586         uint32_t bucket_index;                                  \
 587                                                                 \
 588         signature = RTE_MBUF_METADATA_UINT32(mbuf1, f->signature_offset);\
 589         bucket_index = signature & (f->n_buckets - 1);          \
 590         bucket1 = (struct rte_bucket_4_8 *)                     \
 591                 &f->memory[bucket_index * f->bucket_size];      \
 592         rte_prefetch0(bucket1);                                 \
 593 }
 594
 595 #define lookup1_stage1_dosig(mbuf1, bucket1, f)                 \
 596 {                                                               \
 597         uint64_t *key;                                          \
 598         uint64_t signature;                                     \
 599         uint32_t bucket_index;                                  \
 600         uint64_t hash_key_buffer;                               \
 601                                                                 \
 602         key = RTE_MBUF_METADATA_UINT64_PTR(mbuf1, f->key_offset);\
 603         hash_key_buffer = *key & f->key_mask;                   \
 604         signature = f->f_hash(&hash_key_buffer,                 \
 605                 RTE_TABLE_HASH_KEY_SIZE, f->seed);              \
 606         bucket_index = signature & (f->n_buckets - 1);          \
 607         bucket1 = (struct rte_bucket_4_8 *)                     \
 608                 &f->memory[bucket_index * f->bucket_size];      \
 609         rte_prefetch0(bucket1);                                 \
 610 }
 611
 612 #define lookup1_stage2_lru(pkt2_index, mbuf2, bucket2,          \
 613         pkts_mask_out, entries, f)                              \
 614 {                                                               \
 615         void *a;                                                \
 616         uint64_t pkt_mask;                                      \
 617         uint64_t *key;                                          \
 618         uint32_t pos;                                           \
 619         uint64_t hash_key_buffer;                               \
 620                                                                 \
 621         key = RTE_MBUF_METADATA_UINT64_PTR(mbuf2, f->key_offset);\
 622         hash_key_buffer = key[0] & f->key_mask;                 \
 623                                                                 \
 624         lookup_key8_cmp((&hash_key_buffer), bucket2, pos);      \
 625                                                                 \
 626         pkt_mask = ((bucket2->signature >> pos) & 1LLU) << pkt2_index;\
 627         pkts_mask_out |= pkt_mask;                              \
 628                                                                 \
 629         a = (void *) &bucket2->data[pos * f->entry_size];       \
 630         rte_prefetch0(a);                                       \
 631         entries[pkt2_index] = a;                                \
 632         lru_update(bucket2, pos);                               \
 633 }
 634
 635 #define lookup1_stage2_ext(pkt2_index, mbuf2, bucket2, pkts_mask_out,\
 636         entries, buckets_mask, buckets, keys, f)                \
 637 {                                                               \
 638         struct rte_bucket_4_8 *bucket_next;                     \
 639         void *a;                                                \
 640         uint64_t pkt_mask, bucket_mask;                         \
 641         uint64_t *key;                                          \
 642         uint32_t pos;                                           \
 643         uint64_t hash_key_buffer;                               \
 644                                                                 \
 645         key = RTE_MBUF_METADATA_UINT64_PTR(mbuf2, f->key_offset);\
 646         hash_key_buffer = *key & f->key_mask;                   \
 647                                                                 \
 648         lookup_key8_cmp((&hash_key_buffer), bucket2, pos);      \
 649                                                                 \
 650         pkt_mask = ((bucket2->signature >> pos) & 1LLU) << pkt2_index;\
 651         pkts_mask_out |= pkt_mask;                              \
 652                                                                 \
 653         a = (void *) &bucket2->data[pos * f->entry_size];       \
 654         rte_prefetch0(a);                                       \
 655         entries[pkt2_index] = a;                                \
 656                                                                 \
 657         bucket_mask = (~pkt_mask) & (bucket2->next_valid << pkt2_index);\
 658         buckets_mask |= bucket_mask;                            \
 659         bucket_next = bucket2->next;                            \
 660         buckets[pkt2_index] = bucket_next;                      \
 661         keys[pkt2_index] = key;                                 \
 662 }
 663
 664 #define lookup_grinder(pkt_index, buckets, keys, pkts_mask_out, entries,\
 665         buckets_mask, f)                                        \
 666 {                                                               \
 667         struct rte_bucket_4_8 *bucket, *bucket_next;            \
 668         void *a;                                                \
 669         uint64_t pkt_mask, bucket_mask;                         \
 670         uint64_t *key;                                          \
 671         uint32_t pos;                                           \
 672         uint64_t hash_key_buffer;                               \
 673                                                                 \
 674         bucket = buckets[pkt_index];                            \
 675         key = keys[pkt_index];                                  \
 676         hash_key_buffer = (*key) & f->key_mask;                 \
 677                                                                 \
 678         lookup_key8_cmp((&hash_key_buffer), bucket, pos);       \
 679                                                                 \
 680         pkt_mask = ((bucket->signature >> pos) & 1LLU) << pkt_index;\
 681         pkts_mask_out |= pkt_mask;                              \
 682                                                                 \
 683         a = (void *) &bucket->data[pos * f->entry_size];        \
 684         rte_prefetch0(a);                                       \
 685         entries[pkt_index] = a;                                 \
 686                                                                 \
 687         bucket_mask = (~pkt_mask) & (bucket->next_valid << pkt_index);\
 688         buckets_mask |= bucket_mask;                            \
 689         bucket_next = bucket->next;                             \
 690         rte_prefetch0(bucket_next);                             \
 691         buckets[pkt_index] = bucket_next;                       \
 692         keys[pkt_index] = key;                                  \
 693 }
 694
 695 #define lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01,\
 696         pkts, pkts_mask)                                        \
 697 {                                                               \
 698         uint64_t pkt00_mask, pkt01_mask;                        \
 699                                                                 \
 700         pkt00_index = __builtin_ctzll(pkts_mask);               \
 701         pkt00_mask = 1LLU << pkt00_index;                       \
 702         pkts_mask &= ~pkt00_mask;                               \
 703                                                                 \
 704         mbuf00 = pkts[pkt00_index];                             \
 705         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, 0));  \
 706                                                                 \
 707         pkt01_index = __builtin_ctzll(pkts_mask);               \
 708         pkt01_mask = 1LLU << pkt01_index;                       \
 709         pkts_mask &= ~pkt01_mask;                               \
 710                                                                 \
 711         mbuf01 = pkts[pkt01_index];                             \
 712         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, 0));  \
 713 }
 714
 715 #define lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,\
 716         mbuf00, mbuf01, pkts, pkts_mask)                        \
 717 {                                                               \
 718         uint64_t pkt00_mask, pkt01_mask;                        \
 719                                                                 \
 720         pkt00_index = __builtin_ctzll(pkts_mask);               \
 721         pkt00_mask = 1LLU << pkt00_index;                       \
 722         pkts_mask &= ~pkt00_mask;                               \
 723                                                                 \
 724         mbuf00 = pkts[pkt00_index];                             \
 725         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, 0));  \
 726                                                                 \
 727         pkt01_index = __builtin_ctzll(pkts_mask);               \
 728         if (pkts_mask == 0)                                     \
 729                 pkt01_index = pkt00_index;                      \
 730                                                                 \
 731         pkt01_mask = 1LLU << pkt01_index;                       \
 732         pkts_mask &= ~pkt01_mask;                               \
 733                                                                 \
 734         mbuf01 = pkts[pkt01_index];                             \
 735         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, 0));  \
 736 }
 737
 738 #define lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f)   \
 739 {                                                               \
 740         uint64_t signature10, signature11;                      \
 741         uint32_t bucket10_index, bucket11_index;                \
 742                                                                 \
 743         signature10 = RTE_MBUF_METADATA_UINT32(mbuf10, f->signature_offset);\
 744         bucket10_index = signature10 & (f->n_buckets - 1);      \
 745         bucket10 = (struct rte_bucket_4_8 *)                    \
 746                 &f->memory[bucket10_index * f->bucket_size];    \
 747         rte_prefetch0(bucket10);                                \
 748                                                                 \
 749         signature11 = RTE_MBUF_METADATA_UINT32(mbuf11, f->signature_offset);\
 750         bucket11_index = signature11 & (f->n_buckets - 1);      \
 751         bucket11 = (struct rte_bucket_4_8 *)                    \
 752                 &f->memory[bucket11_index * f->bucket_size];    \
 753         rte_prefetch0(bucket11);                                \
 754 }
 755
 756 #define lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f)\
 757 {                                                               \
 758         uint64_t *key10, *key11;                                \
 759         uint64_t hash_offset_buffer10;                          \
 760         uint64_t hash_offset_buffer11;                          \
 761         uint64_t signature10, signature11;                      \
 762         uint32_t bucket10_index, bucket11_index;                \
 763         rte_table_hash_op_hash f_hash = f->f_hash;              \
 764         uint64_t seed = f->seed;                                \
 765         uint32_t key_offset = f->key_offset;                    \
 766                                                                 \
 767         key10 = RTE_MBUF_METADATA_UINT64_PTR(mbuf10, key_offset);\
 768         key11 = RTE_MBUF_METADATA_UINT64_PTR(mbuf11, key_offset);\
 769         hash_offset_buffer10 = *key10 & f->key_mask;            \
 770         hash_offset_buffer11 = *key11 & f->key_mask;            \
 771                                                                 \
 772         signature10 = f_hash(&hash_offset_buffer10,             \
 773                 RTE_TABLE_HASH_KEY_SIZE, seed);                 \
 774         bucket10_index = signature10 & (f->n_buckets - 1);      \
 775         bucket10 = (struct rte_bucket_4_8 *)                    \
 776                 &f->memory[bucket10_index * f->bucket_size];    \
 777         rte_prefetch0(bucket10);                                \
 778                                                                 \
 779         signature11 = f_hash(&hash_offset_buffer11,             \
 780                 RTE_TABLE_HASH_KEY_SIZE, seed);                 \
 781         bucket11_index = signature11 & (f->n_buckets - 1);      \
 782         bucket11 = (struct rte_bucket_4_8 *)                    \
 783                 &f->memory[bucket11_index * f->bucket_size];    \
 784         rte_prefetch0(bucket11);                                \
 785 }
 786
 787 #define lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,\
 788         bucket20, bucket21, pkts_mask_out, entries, f)          \
 789 {                                                               \
 790         void *a20, *a21;                                        \
 791         uint64_t pkt20_mask, pkt21_mask;                        \
 792         uint64_t *key20, *key21;                                \
 793         uint64_t hash_offset_buffer20;                          \
 794         uint64_t hash_offset_buffer21;                          \
 795         uint32_t pos20, pos21;                                  \
 796                                                                 \
 797         key20 = RTE_MBUF_METADATA_UINT64_PTR(mbuf20, f->key_offset);\
 798         key21 = RTE_MBUF_METADATA_UINT64_PTR(mbuf21, f->key_offset);\
 799         hash_offset_buffer20 = *key20 & f->key_mask;            \
 800         hash_offset_buffer21 = *key21 & f->key_mask;            \
 801                                                                 \
 802         lookup_key8_cmp((&hash_offset_buffer20), bucket20, pos20);\
 803         lookup_key8_cmp((&hash_offset_buffer21), bucket21, pos21);\
 804                                                                 \
 805         pkt20_mask = ((bucket20->signature >> pos20) & 1LLU) << pkt20_index;\
 806         pkt21_mask = ((bucket21->signature >> pos21) & 1LLU) << pkt21_index;\
 807         pkts_mask_out |= pkt20_mask | pkt21_mask;               \
 808                                                                 \
 809         a20 = (void *) &bucket20->data[pos20 * f->entry_size];  \
 810         a21 = (void *) &bucket21->data[pos21 * f->entry_size];  \
 811         rte_prefetch0(a20);                                     \
 812         rte_prefetch0(a21);                                     \
 813         entries[pkt20_index] = a20;                             \
 814         entries[pkt21_index] = a21;                             \
 815         lru_update(bucket20, pos20);                            \
 816         lru_update(bucket21, pos21);                            \
 817 }
 818
 819 #define lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21, bucket20, \
 820         bucket21, pkts_mask_out, entries, buckets_mask, buckets, keys, f)\
 821 {                                                               \
 822         struct rte_bucket_4_8 *bucket20_next, *bucket21_next;   \
 823         void *a20, *a21;                                        \
 824         uint64_t pkt20_mask, pkt21_mask, bucket20_mask, bucket21_mask;\
 825         uint64_t *key20, *key21;                                \
 826         uint64_t hash_offset_buffer20;                          \
 827         uint64_t hash_offset_buffer21;                          \
 828         uint32_t pos20, pos21;                                  \
 829                                                                 \
 830         key20 = RTE_MBUF_METADATA_UINT64_PTR(mbuf20, f->key_offset);\
 831         key21 = RTE_MBUF_METADATA_UINT64_PTR(mbuf21, f->key_offset);\
 832         hash_offset_buffer20 = *key20 & f->key_mask;            \
 833         hash_offset_buffer21 = *key21 & f->key_mask;            \
 834                                                                 \
 835         lookup_key8_cmp((&hash_offset_buffer20), bucket20, pos20);\
 836         lookup_key8_cmp((&hash_offset_buffer21), bucket21, pos21);\
 837                                                                 \
 838         pkt20_mask = ((bucket20->signature >> pos20) & 1LLU) << pkt20_index;\
 839         pkt21_mask = ((bucket21->signature >> pos21) & 1LLU) << pkt21_index;\
 840         pkts_mask_out |= pkt20_mask | pkt21_mask;               \
 841                                                                 \
 842         a20 = (void *) &bucket20->data[pos20 * f->entry_size];  \
 843         a21 = (void *) &bucket21->data[pos21 * f->entry_size];  \
 844         rte_prefetch0(a20);                                     \
 845         rte_prefetch0(a21);                                     \
 846         entries[pkt20_index] = a20;                             \
 847         entries[pkt21_index] = a21;                             \
 848                                                                 \
 849         bucket20_mask = (~pkt20_mask) & (bucket20->next_valid << pkt20_index);\
 850         bucket21_mask = (~pkt21_mask) & (bucket21->next_valid << pkt21_index);\
 851         buckets_mask |= bucket20_mask | bucket21_mask;          \
 852         bucket20_next = bucket20->next;                         \
 853         bucket21_next = bucket21->next;                         \
 854         buckets[pkt20_index] = bucket20_next;                   \
 855         buckets[pkt21_index] = bucket21_next;                   \
 856         keys[pkt20_index] = key20;                              \
 857         keys[pkt21_index] = key21;                              \
 858 }
 859
 860 static int
 861 rte_table_hash_lookup_key8_lru(
 862         void *table,
 863         struct rte_mbuf **pkts,
 864         uint64_t pkts_mask,
 865         uint64_t *lookup_hit_mask,
 866         void **entries)
 867 {
 868         struct rte_table_hash *f = (struct rte_table_hash *) table;
 869         struct rte_bucket_4_8 *bucket10, *bucket11, *bucket20, *bucket21;
 870         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
 871         uint32_t pkt00_index, pkt01_index, pkt10_index,
 872                         pkt11_index, pkt20_index, pkt21_index;
 873         uint64_t pkts_mask_out = 0;
 874
 875         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
 876         RTE_TABLE_HASH_KEY8_STATS_PKTS_IN_ADD(f, n_pkts_in);
 877
 878         /* Cannot run the pipeline with less than 5 packets */
 879         if (__builtin_popcountll(pkts_mask) < 5) {
 880                 for ( ; pkts_mask; ) {
 881                         struct rte_bucket_4_8 *bucket;
 882                         struct rte_mbuf *mbuf;
 883                         uint32_t pkt_index;
 884
 885                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
 886                         lookup1_stage1(mbuf, bucket, f);
 887                         lookup1_stage2_lru(pkt_index, mbuf, bucket,
 888                                         pkts_mask_out, entries, f);
 889                 }
 890
 891                 *lookup_hit_mask = pkts_mask_out;
 892                 RTE_TABLE_HASH_KEY8_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in - __builtin_popcountll(pkts_mask_out));
 893                 return 0;
 894         }
 895
 896         /*
 897          * Pipeline fill
 898          *
 899          */
 900         /* Pipeline stage 0 */
 901         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
 902                 pkts_mask);
 903
 904         /* Pipeline feed */
 905         mbuf10 = mbuf00;
 906         mbuf11 = mbuf01;
 907         pkt10_index = pkt00_index;
 908         pkt11_index = pkt01_index;
 909
 910         /* Pipeline stage 0 */
 911         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
 912                 pkts_mask);
 913
 914         /* Pipeline stage 1 */
 915         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
 916
 917         /*
 918          * Pipeline run
 919          *
 920          */
 921         for ( ; pkts_mask; ) {
 922                 /* Pipeline feed */
 923                 bucket20 = bucket10;
 924                 bucket21 = bucket11;
 925                 mbuf20 = mbuf10;
 926                 mbuf21 = mbuf11;
 927                 mbuf10 = mbuf00;
 928                 mbuf11 = mbuf01;
 929                 pkt20_index = pkt10_index;
 930                 pkt21_index = pkt11_index;
 931                 pkt10_index = pkt00_index;
 932                 pkt11_index = pkt01_index;
 933
 934                 /* Pipeline stage 0 */
 935                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
 936                         mbuf00, mbuf01, pkts, pkts_mask);
 937
 938                 /* Pipeline stage 1 */
 939                 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
 940
 941                 /* Pipeline stage 2 */
 942                 lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
 943                         bucket20, bucket21, pkts_mask_out, entries, f);
 944         }
 945
 946         /*
 947          * Pipeline flush
 948          *
 949          */
 950         /* Pipeline feed */
 951         bucket20 = bucket10;
 952         bucket21 = bucket11;
 953         mbuf20 = mbuf10;
 954         mbuf21 = mbuf11;
 955         mbuf10 = mbuf00;
 956         mbuf11 = mbuf01;
 957         pkt20_index = pkt10_index;
 958         pkt21_index = pkt11_index;
 959         pkt10_index = pkt00_index;
 960         pkt11_index = pkt01_index;
 961
 962         /* Pipeline stage 1 */
 963         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
 964
 965         /* Pipeline stage 2 */
 966         lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
 967                 bucket20, bucket21, pkts_mask_out, entries, f);
 968
 969         /* Pipeline feed */
 970         bucket20 = bucket10;
 971         bucket21 = bucket11;
 972         mbuf20 = mbuf10;
 973         mbuf21 = mbuf11;
 974         pkt20_index = pkt10_index;
 975         pkt21_index = pkt11_index;
 976
 977         /* Pipeline stage 2 */
 978         lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
 979                 bucket20, bucket21, pkts_mask_out, entries, f);
 980
 981         *lookup_hit_mask = pkts_mask_out;
 982         RTE_TABLE_HASH_KEY8_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in - __builtin_popcountll(pkts_mask_out));
 983         return 0;
 984 } /* rte_table_hash_lookup_key8_lru() */
 985
 986 static int
 987 rte_table_hash_lookup_key8_lru_dosig(
 988         void *table,
 989         struct rte_mbuf **pkts,
 990         uint64_t pkts_mask,
 991         uint64_t *lookup_hit_mask,
 992         void **entries)
 993 {
 994         struct rte_table_hash *f = (struct rte_table_hash *) table;
 995         struct rte_bucket_4_8 *bucket10, *bucket11, *bucket20, *bucket21;
 996         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
 997         uint32_t pkt00_index, pkt01_index, pkt10_index;
 998         uint32_t pkt11_index, pkt20_index, pkt21_index;
 999         uint64_t pkts_mask_out = 0;
1000
1001         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
1002         RTE_TABLE_HASH_KEY8_STATS_PKTS_IN_ADD(f, n_pkts_in);
1003
1004         /* Cannot run the pipeline with less than 5 packets */
1005         if (__builtin_popcountll(pkts_mask) < 5) {
1006                 for ( ; pkts_mask; ) {
1007                         struct rte_bucket_4_8 *bucket;
1008                         struct rte_mbuf *mbuf;
1009                         uint32_t pkt_index;
1010
1011                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
1012                         lookup1_stage1_dosig(mbuf, bucket, f);
1013                         lookup1_stage2_lru(pkt_index, mbuf, bucket,
1014                                 pkts_mask_out, entries, f);
1015                 }
1016
1017                 *lookup_hit_mask = pkts_mask_out;
1018                 RTE_TABLE_HASH_KEY8_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in - __builtin_popcountll(pkts_mask_out));
1019                 return 0;
1020         }
1021
1022         /*
1023          * Pipeline fill
1024          *
1025          */
1026         /* Pipeline stage 0 */
1027         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1028                 pkts_mask);
1029
1030         /* Pipeline feed */
1031         mbuf10 = mbuf00;
1032         mbuf11 = mbuf01;
1033         pkt10_index = pkt00_index;
1034         pkt11_index = pkt01_index;
1035
1036         /* Pipeline stage 0 */
1037         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1038                 pkts_mask);
1039
1040         /* Pipeline stage 1 */
1041         lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1042
1043         /*
1044          * Pipeline run
1045          *
1046          */
1047         for ( ; pkts_mask; ) {
1048                 /* Pipeline feed */
1049                 bucket20 = bucket10;
1050                 bucket21 = bucket11;
1051                 mbuf20 = mbuf10;
1052                 mbuf21 = mbuf11;
1053                 mbuf10 = mbuf00;
1054                 mbuf11 = mbuf01;
1055                 pkt20_index = pkt10_index;
1056                 pkt21_index = pkt11_index;
1057                 pkt10_index = pkt00_index;
1058                 pkt11_index = pkt01_index;
1059
1060                 /* Pipeline stage 0 */
1061                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
1062                         mbuf00, mbuf01, pkts, pkts_mask);
1063
1064                 /* Pipeline stage 1 */
1065                 lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1066
1067                 /* Pipeline stage 2 */
1068                 lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
1069                         bucket20, bucket21, pkts_mask_out, entries, f);
1070         }
1071
1072         /*
1073          * Pipeline flush
1074          *
1075          */
1076         /* Pipeline feed */
1077         bucket20 = bucket10;
1078         bucket21 = bucket11;
1079         mbuf20 = mbuf10;
1080         mbuf21 = mbuf11;
1081         mbuf10 = mbuf00;
1082         mbuf11 = mbuf01;
1083         pkt20_index = pkt10_index;
1084         pkt21_index = pkt11_index;
1085         pkt10_index = pkt00_index;
1086         pkt11_index = pkt01_index;
1087
1088         /* Pipeline stage 1 */
1089         lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1090
1091         /* Pipeline stage 2 */
1092         lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
1093                 bucket20, bucket21, pkts_mask_out, entries, f);
1094
1095         /* Pipeline feed */
1096         bucket20 = bucket10;
1097         bucket21 = bucket11;
1098         mbuf20 = mbuf10;
1099         mbuf21 = mbuf11;
1100         pkt20_index = pkt10_index;
1101         pkt21_index = pkt11_index;
1102
1103         /* Pipeline stage 2 */
1104         lookup2_stage2_lru(pkt20_index, pkt21_index, mbuf20, mbuf21,
1105                 bucket20, bucket21, pkts_mask_out, entries, f);
1106
1107         *lookup_hit_mask = pkts_mask_out;
1108         RTE_TABLE_HASH_KEY8_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in - __builtin_popcountll(pkts_mask_out));
1109         return 0;
1110 } /* rte_table_hash_lookup_key8_lru_dosig() */
1111
1112 static int
1113 rte_table_hash_lookup_key8_ext(
1114         void *table,
1115         struct rte_mbuf **pkts,
1116         uint64_t pkts_mask,
1117         uint64_t *lookup_hit_mask,
1118         void **entries)
1119 {
1120         struct rte_table_hash *f = (struct rte_table_hash *) table;
1121         struct rte_bucket_4_8 *bucket10, *bucket11, *bucket20, *bucket21;
1122         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
1123         uint32_t pkt00_index, pkt01_index, pkt10_index;
1124         uint32_t pkt11_index, pkt20_index, pkt21_index;
1125         uint64_t pkts_mask_out = 0, buckets_mask = 0;
1126         struct rte_bucket_4_8 *buckets[RTE_PORT_IN_BURST_SIZE_MAX];
1127         uint64_t *keys[RTE_PORT_IN_BURST_SIZE_MAX];
1128
1129         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
1130         RTE_TABLE_HASH_KEY8_STATS_PKTS_IN_ADD(f, n_pkts_in);
1131
1132         /* Cannot run the pipeline with less than 5 packets */
1133         if (__builtin_popcountll(pkts_mask) < 5) {
1134                 for ( ; pkts_mask; ) {
1135                         struct rte_bucket_4_8 *bucket;
1136                         struct rte_mbuf *mbuf;
1137                         uint32_t pkt_index;
1138
1139                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
1140                         lookup1_stage1(mbuf, bucket, f);
1141                         lookup1_stage2_ext(pkt_index, mbuf, bucket,
1142                                 pkts_mask_out, entries, buckets_mask, buckets,
1143                                 keys, f);
1144                 }
1145
1146                 goto grind_next_buckets;
1147         }
1148
1149         /*
1150          * Pipeline fill
1151          *
1152          */
1153         /* Pipeline stage 0 */
1154         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1155                 pkts_mask);
1156
1157         /* Pipeline feed */
1158         mbuf10 = mbuf00;
1159         mbuf11 = mbuf01;
1160         pkt10_index = pkt00_index;
1161         pkt11_index = pkt01_index;
1162
1163         /* Pipeline stage 0 */
1164         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1165                 pkts_mask);
1166
1167         /* Pipeline stage 1 */
1168         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1169
1170         /*
1171          * Pipeline run
1172          *
1173          */
1174         for ( ; pkts_mask; ) {
1175                 /* Pipeline feed */
1176                 bucket20 = bucket10;
1177                 bucket21 = bucket11;
1178                 mbuf20 = mbuf10;
1179                 mbuf21 = mbuf11;
1180                 mbuf10 = mbuf00;
1181                 mbuf11 = mbuf01;
1182                 pkt20_index = pkt10_index;
1183                 pkt21_index = pkt11_index;
1184                 pkt10_index = pkt00_index;
1185                 pkt11_index = pkt01_index;
1186
1187                 /* Pipeline stage 0 */
1188                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
1189                         mbuf00, mbuf01, pkts, pkts_mask);
1190
1191                 /* Pipeline stage 1 */
1192                 lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1193
1194                 /* Pipeline stage 2 */
1195                 lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1196                         bucket20, bucket21, pkts_mask_out, entries,
1197                         buckets_mask, buckets, keys, f);
1198         }
1199
1200         /*
1201          * Pipeline flush
1202          *
1203          */
1204         /* Pipeline feed */
1205         bucket20 = bucket10;
1206         bucket21 = bucket11;
1207         mbuf20 = mbuf10;
1208         mbuf21 = mbuf11;
1209         mbuf10 = mbuf00;
1210         mbuf11 = mbuf01;
1211         pkt20_index = pkt10_index;
1212         pkt21_index = pkt11_index;
1213         pkt10_index = pkt00_index;
1214         pkt11_index = pkt01_index;
1215
1216         /* Pipeline stage 1 */
1217         lookup2_stage1(mbuf10, mbuf11, bucket10, bucket11, f);
1218
1219         /* Pipeline stage 2 */
1220         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1221                 bucket20, bucket21, pkts_mask_out, entries,
1222                 buckets_mask, buckets, keys, f);
1223
1224         /* Pipeline feed */
1225         bucket20 = bucket10;
1226         bucket21 = bucket11;
1227         mbuf20 = mbuf10;
1228         mbuf21 = mbuf11;
1229         pkt20_index = pkt10_index;
1230         pkt21_index = pkt11_index;
1231
1232         /* Pipeline stage 2 */
1233         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1234                 bucket20, bucket21, pkts_mask_out, entries,
1235                 buckets_mask, buckets, keys, f);
1236
1237 grind_next_buckets:
1238         /* Grind next buckets */
1239         for ( ; buckets_mask; ) {
1240                 uint64_t buckets_mask_next = 0;
1241
1242                 for ( ; buckets_mask; ) {
1243                         uint64_t pkt_mask;
1244                         uint32_t pkt_index;
1245
1246                         pkt_index = __builtin_ctzll(buckets_mask);
1247                         pkt_mask = 1LLU << pkt_index;
1248                         buckets_mask &= ~pkt_mask;
1249
1250                         lookup_grinder(pkt_index, buckets, keys, pkts_mask_out,
1251                                 entries, buckets_mask_next, f);
1252                 }
1253
1254                 buckets_mask = buckets_mask_next;
1255         }
1256
1257         *lookup_hit_mask = pkts_mask_out;
1258         RTE_TABLE_HASH_KEY8_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in - __builtin_popcountll(pkts_mask_out));
1259         return 0;
1260 } /* rte_table_hash_lookup_key8_ext() */
1261
1262 static int
1263 rte_table_hash_lookup_key8_ext_dosig(
1264         void *table,
1265         struct rte_mbuf **pkts,
1266         uint64_t pkts_mask,
1267         uint64_t *lookup_hit_mask,
1268         void **entries)
1269 {
1270         struct rte_table_hash *f = (struct rte_table_hash *) table;
1271         struct rte_bucket_4_8 *bucket10, *bucket11, *bucket20, *bucket21;
1272         struct rte_mbuf *mbuf00, *mbuf01, *mbuf10, *mbuf11, *mbuf20, *mbuf21;
1273         uint32_t pkt00_index, pkt01_index, pkt10_index;
1274         uint32_t pkt11_index, pkt20_index, pkt21_index;
1275         uint64_t pkts_mask_out = 0, buckets_mask = 0;
1276         struct rte_bucket_4_8 *buckets[RTE_PORT_IN_BURST_SIZE_MAX];
1277         uint64_t *keys[RTE_PORT_IN_BURST_SIZE_MAX];
1278
1279         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
1280         RTE_TABLE_HASH_KEY8_STATS_PKTS_IN_ADD(f, n_pkts_in);
1281
1282         /* Cannot run the pipeline with less than 5 packets */
1283         if (__builtin_popcountll(pkts_mask) < 5) {
1284                 for ( ; pkts_mask; ) {
1285                         struct rte_bucket_4_8 *bucket;
1286                         struct rte_mbuf *mbuf;
1287                         uint32_t pkt_index;
1288
1289                         lookup1_stage0(pkt_index, mbuf, pkts, pkts_mask);
1290                         lookup1_stage1_dosig(mbuf, bucket, f);
1291                         lookup1_stage2_ext(pkt_index, mbuf, bucket,
1292                                 pkts_mask_out, entries, buckets_mask,
1293                                 buckets, keys, f);
1294                 }
1295
1296                 goto grind_next_buckets;
1297         }
1298
1299         /*
1300          * Pipeline fill
1301          *
1302          */
1303         /* Pipeline stage 0 */
1304         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1305                 pkts_mask);
1306
1307         /* Pipeline feed */
1308         mbuf10 = mbuf00;
1309         mbuf11 = mbuf01;
1310         pkt10_index = pkt00_index;
1311         pkt11_index = pkt01_index;
1312
1313         /* Pipeline stage 0 */
1314         lookup2_stage0(pkt00_index, pkt01_index, mbuf00, mbuf01, pkts,
1315                 pkts_mask);
1316
1317         /* Pipeline stage 1 */
1318         lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1319
1320         /*
1321          * Pipeline run
1322          *
1323          */
1324         for ( ; pkts_mask; ) {
1325                 /* Pipeline feed */
1326                 bucket20 = bucket10;
1327                 bucket21 = bucket11;
1328                 mbuf20 = mbuf10;
1329                 mbuf21 = mbuf11;
1330                 mbuf10 = mbuf00;
1331                 mbuf11 = mbuf01;
1332                 pkt20_index = pkt10_index;
1333                 pkt21_index = pkt11_index;
1334                 pkt10_index = pkt00_index;
1335                 pkt11_index = pkt01_index;
1336
1337                 /* Pipeline stage 0 */
1338                 lookup2_stage0_with_odd_support(pkt00_index, pkt01_index,
1339                         mbuf00, mbuf01, pkts, pkts_mask);
1340
1341                 /* Pipeline stage 1 */
1342                 lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1343
1344                 /* Pipeline stage 2 */
1345                 lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1346                         bucket20, bucket21, pkts_mask_out, entries,
1347                         buckets_mask, buckets, keys, f);
1348         }
1349
1350         /*
1351          * Pipeline flush
1352          *
1353          */
1354         /* Pipeline feed */
1355         bucket20 = bucket10;
1356         bucket21 = bucket11;
1357         mbuf20 = mbuf10;
1358         mbuf21 = mbuf11;
1359         mbuf10 = mbuf00;
1360         mbuf11 = mbuf01;
1361         pkt20_index = pkt10_index;
1362         pkt21_index = pkt11_index;
1363         pkt10_index = pkt00_index;
1364         pkt11_index = pkt01_index;
1365
1366         /* Pipeline stage 1 */
1367         lookup2_stage1_dosig(mbuf10, mbuf11, bucket10, bucket11, f);
1368
1369         /* Pipeline stage 2 */
1370         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1371                 bucket20, bucket21, pkts_mask_out, entries,
1372                 buckets_mask, buckets, keys, f);
1373
1374         /* Pipeline feed */
1375         bucket20 = bucket10;
1376         bucket21 = bucket11;
1377         mbuf20 = mbuf10;
1378         mbuf21 = mbuf11;
1379         pkt20_index = pkt10_index;
1380         pkt21_index = pkt11_index;
1381
1382         /* Pipeline stage 2 */
1383         lookup2_stage2_ext(pkt20_index, pkt21_index, mbuf20, mbuf21,
1384                 bucket20, bucket21, pkts_mask_out, entries,
1385                 buckets_mask, buckets, keys, f);
1386
1387 grind_next_buckets:
1388         /* Grind next buckets */
1389         for ( ; buckets_mask; ) {
1390                 uint64_t buckets_mask_next = 0;
1391
1392                 for ( ; buckets_mask; ) {
1393                         uint64_t pkt_mask;
1394                         uint32_t pkt_index;
1395
1396                         pkt_index = __builtin_ctzll(buckets_mask);
1397                         pkt_mask = 1LLU << pkt_index;
1398                         buckets_mask &= ~pkt_mask;
1399
1400                         lookup_grinder(pkt_index, buckets, keys, pkts_mask_out,
1401                                 entries, buckets_mask_next, f);
1402                 }
1403
1404                 buckets_mask = buckets_mask_next;
1405         }
1406
1407         *lookup_hit_mask = pkts_mask_out;
1408         RTE_TABLE_HASH_KEY8_STATS_PKTS_LOOKUP_MISS(f, n_pkts_in - __builtin_popcountll(pkts_mask_out));
1409         return 0;
1410 } /* rte_table_hash_lookup_key8_dosig_ext() */
1411
1412 static int
1413 rte_table_hash_key8_stats_read(void *table, struct rte_table_stats *stats, int clear)
1414 {
1415         struct rte_table_hash *t = (struct rte_table_hash *) table;
1416
1417         if (stats != NULL)
1418                 memcpy(stats, &t->stats, sizeof(t->stats));
1419
1420         if (clear)
1421                 memset(&t->stats, 0, sizeof(t->stats));
1422
1423         return 0;
1424 }
1425
1426 struct rte_table_ops rte_table_hash_key8_lru_ops = {
1427         .f_create = rte_table_hash_create_key8_lru,
1428         .f_free = rte_table_hash_free_key8_lru,
1429         .f_add = rte_table_hash_entry_add_key8_lru,
1430         .f_delete = rte_table_hash_entry_delete_key8_lru,
1431         .f_add_bulk = NULL,
1432         .f_delete_bulk = NULL,
1433         .f_lookup = rte_table_hash_lookup_key8_lru,
1434         .f_stats = rte_table_hash_key8_stats_read,
1435 };
1436
1437 struct rte_table_ops rte_table_hash_key8_lru_dosig_ops = {
1438         .f_create = rte_table_hash_create_key8_lru,
1439         .f_free = rte_table_hash_free_key8_lru,
1440         .f_add = rte_table_hash_entry_add_key8_lru,
1441         .f_delete = rte_table_hash_entry_delete_key8_lru,
1442         .f_add_bulk = NULL,
1443         .f_delete_bulk = NULL,
1444         .f_lookup = rte_table_hash_lookup_key8_lru_dosig,
1445         .f_stats = rte_table_hash_key8_stats_read,
1446 };
1447
1448 struct rte_table_ops rte_table_hash_key8_ext_ops = {
1449         .f_create = rte_table_hash_create_key8_ext,
1450         .f_free = rte_table_hash_free_key8_ext,
1451         .f_add = rte_table_hash_entry_add_key8_ext,
1452         .f_delete = rte_table_hash_entry_delete_key8_ext,
1453         .f_add_bulk = NULL,
1454         .f_delete_bulk = NULL,
1455         .f_lookup = rte_table_hash_lookup_key8_ext,
1456         .f_stats = rte_table_hash_key8_stats_read,
1457 };
1458
1459 struct rte_table_ops rte_table_hash_key8_ext_dosig_ops = {
1460         .f_create = rte_table_hash_create_key8_ext,
1461         .f_free = rte_table_hash_free_key8_ext,
1462         .f_add = rte_table_hash_entry_add_key8_ext,
1463         .f_delete = rte_table_hash_entry_delete_key8_ext,
1464         .f_add_bulk = NULL,
1465         .f_delete_bulk = NULL,
1466         .f_lookup = rte_table_hash_lookup_key8_ext_dosig,
1467         .f_stats = rte_table_hash_key8_stats_read,
1468 };