lib/librte_table/rte_table_hash_ext.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
  34 #include <string.h>
  35 #include <stdio.h>
  36
  37 #include <rte_common.h>
  38 #include <rte_mbuf.h>
  39 #include <rte_memory.h>
  40 #include <rte_malloc.h>
  41 #include <rte_log.h>
  42
  43 #include "rte_table_hash.h"
  44
  45 #define KEYS_PER_BUCKET 4
  46
  47 struct bucket {
  48         union {
  49                 uintptr_t next;
  50                 uint64_t lru_list;
  51         };
  52         uint16_t sig[KEYS_PER_BUCKET];
  53         uint32_t key_pos[KEYS_PER_BUCKET];
  54 };
  55
  56 #define BUCKET_NEXT(bucket)                                             \
  57         ((void *) ((bucket)->next & (~1LU)))
  58
  59 #define BUCKET_NEXT_VALID(bucket)                                       \
  60         ((bucket)->next & 1LU)
  61
  62 #define BUCKET_NEXT_SET(bucket, bucket_next)                            \
  63 do                                                                      \
  64         (bucket)->next = (((uintptr_t) ((void *) (bucket_next))) | 1LU);\
  65 while (0)
  66
  67 #define BUCKET_NEXT_SET_NULL(bucket)                                    \
  68 do                                                                      \
  69         (bucket)->next = 0;                                             \
  70 while (0)
  71
  72 #define BUCKET_NEXT_COPY(bucket, bucket2)                               \
  73 do                                                                      \
  74         (bucket)->next = (bucket2)->next;                               \
  75 while (0)
  76
  77 #ifdef RTE_TABLE_STATS_COLLECT
  78
  79 #define RTE_TABLE_HASH_EXT_STATS_PKTS_IN_ADD(table, val) \
  80         table->stats.n_pkts_in += val
  81 #define RTE_TABLE_HASH_EXT_STATS_PKTS_LOOKUP_MISS(table, val) \
  82         table->stats.n_pkts_lookup_miss += val
  83
  84 #else
  85
  86 #define RTE_TABLE_HASH_EXT_STATS_PKTS_IN_ADD(table, val)
  87 #define RTE_TABLE_HASH_EXT_STATS_PKTS_LOOKUP_MISS(table, val)
  88
  89 #endif
  90
  91 struct grinder {
  92         struct bucket *bkt;
  93         uint64_t sig;
  94         uint64_t match;
  95         uint32_t key_index;
  96 };
  97
  98 struct rte_table_hash {
  99         struct rte_table_stats stats;
 100
 101         /* Input parameters */
 102         uint32_t key_size;
 103         uint32_t entry_size;
 104         uint32_t n_keys;
 105         uint32_t n_buckets;
 106         uint32_t n_buckets_ext;
 107         rte_table_hash_op_hash f_hash;
 108         uint64_t seed;
 109         uint32_t signature_offset;
 110         uint32_t key_offset;
 111
 112         /* Internal */
 113         uint64_t bucket_mask;
 114         uint32_t key_size_shl;
 115         uint32_t data_size_shl;
 116         uint32_t key_stack_tos;
 117         uint32_t bkt_ext_stack_tos;
 118
 119         /* Grinder */
 120         struct grinder grinders[RTE_PORT_IN_BURST_SIZE_MAX];
 121
 122         /* Tables */
 123         struct bucket *buckets;
 124         struct bucket *buckets_ext;
 125         uint8_t *key_mem;
 126         uint8_t *data_mem;
 127         uint32_t *key_stack;
 128         uint32_t *bkt_ext_stack;
 129
 130         /* Table memory */
 131         uint8_t memory[0] __rte_cache_aligned;
 132 };
 133
 134 static int
 135 check_params_create(struct rte_table_hash_ext_params *params)
 136 {
 137         uint32_t n_buckets_min;
 138
 139         /* key_size */
 140         if ((params->key_size == 0) ||
 141                 (!rte_is_power_of_2(params->key_size))) {
 142                 RTE_LOG(ERR, TABLE, "%s: key_size invalid value\n", __func__);
 143                 return -EINVAL;
 144         }
 145
 146         /* n_keys */
 147         if ((params->n_keys == 0) ||
 148                 (!rte_is_power_of_2(params->n_keys))) {
 149                 RTE_LOG(ERR, TABLE, "%s: n_keys invalid value\n", __func__);
 150                 return -EINVAL;
 151         }
 152
 153         /* n_buckets */
 154         n_buckets_min = (params->n_keys + KEYS_PER_BUCKET - 1) / params->n_keys;
 155         if ((params->n_buckets == 0) ||
 156                 (!rte_is_power_of_2(params->n_keys)) ||
 157                 (params->n_buckets < n_buckets_min)) {
 158                 RTE_LOG(ERR, TABLE, "%s: n_buckets invalid value\n", __func__);
 159                 return -EINVAL;
 160         }
 161
 162         /* f_hash */
 163         if (params->f_hash == NULL) {
 164                 RTE_LOG(ERR, TABLE, "%s: f_hash invalid value\n", __func__);
 165                 return -EINVAL;
 166         }
 167
 168         return 0;
 169 }
 170
 171 static void *
 172 rte_table_hash_ext_create(void *params, int socket_id, uint32_t entry_size)
 173 {
 174         struct rte_table_hash_ext_params *p =
 175                 (struct rte_table_hash_ext_params *) params;
 176         struct rte_table_hash *t;
 177         uint32_t total_size, table_meta_sz;
 178         uint32_t bucket_sz, bucket_ext_sz, key_sz;
 179         uint32_t key_stack_sz, bkt_ext_stack_sz, data_sz;
 180         uint32_t bucket_offset, bucket_ext_offset, key_offset;
 181         uint32_t key_stack_offset, bkt_ext_stack_offset, data_offset;
 182         uint32_t i;
 183
 184         /* Check input parameters */
 185         if ((check_params_create(p) != 0) ||
 186                 (!rte_is_power_of_2(entry_size)) ||
 187                 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) ||
 188                 (sizeof(struct bucket) != (RTE_CACHE_LINE_SIZE / 2)))
 189                 return NULL;
 190
 191         /* Memory allocation */
 192         table_meta_sz = RTE_CACHE_LINE_ROUNDUP(sizeof(struct rte_table_hash));
 193         bucket_sz = RTE_CACHE_LINE_ROUNDUP(p->n_buckets * sizeof(struct bucket));
 194         bucket_ext_sz =
 195                 RTE_CACHE_LINE_ROUNDUP(p->n_buckets_ext * sizeof(struct bucket));
 196         key_sz = RTE_CACHE_LINE_ROUNDUP(p->n_keys * p->key_size);
 197         key_stack_sz = RTE_CACHE_LINE_ROUNDUP(p->n_keys * sizeof(uint32_t));
 198         bkt_ext_stack_sz =
 199                 RTE_CACHE_LINE_ROUNDUP(p->n_buckets_ext * sizeof(uint32_t));
 200         data_sz = RTE_CACHE_LINE_ROUNDUP(p->n_keys * entry_size);
 201         total_size = table_meta_sz + bucket_sz + bucket_ext_sz + key_sz +
 202                 key_stack_sz + bkt_ext_stack_sz + data_sz;
 203
 204         t = rte_zmalloc_socket("TABLE", total_size, RTE_CACHE_LINE_SIZE, socket_id);
 205         if (t == NULL) {
 206                 RTE_LOG(ERR, TABLE,
 207                         "%s: Cannot allocate %u bytes for hash table\n",
 208                         __func__, total_size);
 209                 return NULL;
 210         }
 211         RTE_LOG(INFO, TABLE, "%s (%u-byte key): Hash table memory footprint is "
 212                 "%u bytes\n", __func__, p->key_size, total_size);
 213
 214         /* Memory initialization */
 215         t->key_size = p->key_size;
 216         t->entry_size = entry_size;
 217         t->n_keys = p->n_keys;
 218         t->n_buckets = p->n_buckets;
 219         t->n_buckets_ext = p->n_buckets_ext;
 220         t->f_hash = p->f_hash;
 221         t->seed = p->seed;
 222         t->signature_offset = p->signature_offset;
 223         t->key_offset = p->key_offset;
 224
 225         /* Internal */
 226         t->bucket_mask = t->n_buckets - 1;
 227         t->key_size_shl = __builtin_ctzl(p->key_size);
 228         t->data_size_shl = __builtin_ctzl(entry_size);
 229
 230         /* Tables */
 231         bucket_offset = 0;
 232         bucket_ext_offset = bucket_offset + bucket_sz;
 233         key_offset = bucket_ext_offset + bucket_ext_sz;
 234         key_stack_offset = key_offset + key_sz;
 235         bkt_ext_stack_offset = key_stack_offset + key_stack_sz;
 236         data_offset = bkt_ext_stack_offset + bkt_ext_stack_sz;
 237
 238         t->buckets = (struct bucket *) &t->memory[bucket_offset];
 239         t->buckets_ext = (struct bucket *) &t->memory[bucket_ext_offset];
 240         t->key_mem = &t->memory[key_offset];
 241         t->key_stack = (uint32_t *) &t->memory[key_stack_offset];
 242         t->bkt_ext_stack = (uint32_t *) &t->memory[bkt_ext_stack_offset];
 243         t->data_mem = &t->memory[data_offset];
 244
 245         /* Key stack */
 246         for (i = 0; i < t->n_keys; i++)
 247                 t->key_stack[i] = t->n_keys - 1 - i;
 248         t->key_stack_tos = t->n_keys;
 249
 250         /* Bucket ext stack */
 251         for (i = 0; i < t->n_buckets_ext; i++)
 252                 t->bkt_ext_stack[i] = t->n_buckets_ext - 1 - i;
 253         t->bkt_ext_stack_tos = t->n_buckets_ext;
 254
 255         return t;
 256 }
 257
 258 static int
 259 rte_table_hash_ext_free(void *table)
 260 {
 261         struct rte_table_hash *t = (struct rte_table_hash *) table;
 262
 263         /* Check input parameters */
 264         if (t == NULL)
 265                 return -EINVAL;
 266
 267         rte_free(t);
 268         return 0;
 269 }
 270
 271 static int
 272 rte_table_hash_ext_entry_add(void *table, void *key, void *entry,
 273         int *key_found, void **entry_ptr)
 274 {
 275         struct rte_table_hash *t = (struct rte_table_hash *) table;
 276         struct bucket *bkt0, *bkt, *bkt_prev;
 277         uint64_t sig;
 278         uint32_t bkt_index, i;
 279
 280         sig = t->f_hash(key, t->key_size, t->seed);
 281         bkt_index = sig & t->bucket_mask;
 282         bkt0 = &t->buckets[bkt_index];
 283         sig = (sig >> 16) | 1LLU;
 284
 285         /* Key is present in the bucket */
 286         for (bkt = bkt0; bkt != NULL; bkt = BUCKET_NEXT(bkt))
 287                 for (i = 0; i < KEYS_PER_BUCKET; i++) {
 288                         uint64_t bkt_sig = (uint64_t) bkt->sig[i];
 289                         uint32_t bkt_key_index = bkt->key_pos[i];
 290                         uint8_t *bkt_key =
 291                                 &t->key_mem[bkt_key_index << t->key_size_shl];
 292
 293                         if ((sig == bkt_sig) && (memcmp(key, bkt_key,
 294                                 t->key_size) == 0)) {
 295                                 uint8_t *data = &t->data_mem[bkt_key_index <<
 296                                         t->data_size_shl];
 297
 298                                 memcpy(data, entry, t->entry_size);
 299                                 *key_found = 1;
 300                                 *entry_ptr = (void *) data;
 301                                 return 0;
 302                         }
 303                 }
 304
 305         /* Key is not present in the bucket */
 306         for (bkt_prev = NULL, bkt = bkt0; bkt != NULL; bkt_prev = bkt,
 307                 bkt = BUCKET_NEXT(bkt))
 308                 for (i = 0; i < KEYS_PER_BUCKET; i++) {
 309                         uint64_t bkt_sig = (uint64_t) bkt->sig[i];
 310
 311                         if (bkt_sig == 0) {
 312                                 uint32_t bkt_key_index;
 313                                 uint8_t *bkt_key, *data;
 314
 315                                 /* Allocate new key */
 316                                 if (t->key_stack_tos == 0) /* No free keys */
 317                                         return -ENOSPC;
 318
 319                                 bkt_key_index = t->key_stack[
 320                                         --t->key_stack_tos];
 321
 322                                 /* Install new key */
 323                                 bkt_key = &t->key_mem[bkt_key_index <<
 324                                         t->key_size_shl];
 325                                 data = &t->data_mem[bkt_key_index <<
 326                                         t->data_size_shl];
 327
 328                                 bkt->sig[i] = (uint16_t) sig;
 329                                 bkt->key_pos[i] = bkt_key_index;
 330                                 memcpy(bkt_key, key, t->key_size);
 331                                 memcpy(data, entry, t->entry_size);
 332
 333                                 *key_found = 0;
 334                                 *entry_ptr = (void *) data;
 335                                 return 0;
 336                         }
 337                 }
 338
 339         /* Bucket full: extend bucket */
 340         if ((t->bkt_ext_stack_tos > 0) && (t->key_stack_tos > 0)) {
 341                 uint32_t bkt_key_index;
 342                 uint8_t *bkt_key, *data;
 343
 344                 /* Allocate new bucket ext */
 345                 bkt_index = t->bkt_ext_stack[--t->bkt_ext_stack_tos];
 346                 bkt = &t->buckets_ext[bkt_index];
 347
 348                 /* Chain the new bucket ext */
 349                 BUCKET_NEXT_SET(bkt_prev, bkt);
 350                 BUCKET_NEXT_SET_NULL(bkt);
 351
 352                 /* Allocate new key */
 353                 bkt_key_index = t->key_stack[--t->key_stack_tos];
 354                 bkt_key = &t->key_mem[bkt_key_index << t->key_size_shl];
 355
 356                 data = &t->data_mem[bkt_key_index << t->data_size_shl];
 357
 358                 /* Install new key into bucket */
 359                 bkt->sig[0] = (uint16_t) sig;
 360                 bkt->key_pos[0] = bkt_key_index;
 361                 memcpy(bkt_key, key, t->key_size);
 362                 memcpy(data, entry, t->entry_size);
 363
 364                 *key_found = 0;
 365                 *entry_ptr = (void *) data;
 366                 return 0;
 367         }
 368
 369         return -ENOSPC;
 370 }
 371
 372 static int
 373 rte_table_hash_ext_entry_delete(void *table, void *key, int *key_found,
 374 void *entry)
 375 {
 376         struct rte_table_hash *t = (struct rte_table_hash *) table;
 377         struct bucket *bkt0, *bkt, *bkt_prev;
 378         uint64_t sig;
 379         uint32_t bkt_index, i;
 380
 381         sig = t->f_hash(key, t->key_size, t->seed);
 382         bkt_index = sig & t->bucket_mask;
 383         bkt0 = &t->buckets[bkt_index];
 384         sig = (sig >> 16) | 1LLU;
 385
 386         /* Key is present in the bucket */
 387         for (bkt_prev = NULL, bkt = bkt0; bkt != NULL; bkt_prev = bkt,
 388                 bkt = BUCKET_NEXT(bkt))
 389                 for (i = 0; i < KEYS_PER_BUCKET; i++) {
 390                         uint64_t bkt_sig = (uint64_t) bkt->sig[i];
 391                         uint32_t bkt_key_index = bkt->key_pos[i];
 392                         uint8_t *bkt_key = &t->key_mem[bkt_key_index <<
 393                                 t->key_size_shl];
 394
 395                         if ((sig == bkt_sig) && (memcmp(key, bkt_key,
 396                                 t->key_size) == 0)) {
 397                                 uint8_t *data = &t->data_mem[bkt_key_index <<
 398                                         t->data_size_shl];
 399
 400                                 /* Uninstall key from bucket */
 401                                 bkt->sig[i] = 0;
 402                                 *key_found = 1;
 403                                 if (entry)
 404                                         memcpy(entry, data, t->entry_size);
 405
 406                                 /* Free key */
 407                                 t->key_stack[t->key_stack_tos++] =
 408                                         bkt_key_index;
 409
 410                                 /*Check if bucket is unused */
 411                                 if ((bkt_prev != NULL) &&
 412                                     (bkt->sig[0] == 0) && (bkt->sig[1] == 0) &&
 413                                     (bkt->sig[2] == 0) && (bkt->sig[3] == 0)) {
 414                                         /* Unchain bucket */
 415                                         BUCKET_NEXT_COPY(bkt_prev, bkt);
 416
 417                                         /* Clear bucket */
 418                                         memset(bkt, 0, sizeof(struct bucket));
 419
 420                                         /* Free bucket back to buckets ext */
 421                                         bkt_index = bkt - t->buckets_ext;
 422                                         t->bkt_ext_stack[t->bkt_ext_stack_tos++]
 423                                                 = bkt_index;
 424                                 }
 425
 426                                 return 0;
 427                         }
 428                 }
 429
 430         /* Key is not present in the bucket */
 431         *key_found = 0;
 432         return 0;
 433 }
 434
 435 static int rte_table_hash_ext_lookup_unoptimized(
 436         void *table,
 437         struct rte_mbuf **pkts,
 438         uint64_t pkts_mask,
 439         uint64_t *lookup_hit_mask,
 440         void **entries,
 441         int dosig)
 442 {
 443         struct rte_table_hash *t = (struct rte_table_hash *) table;
 444         uint64_t pkts_mask_out = 0;
 445
 446         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
 447         RTE_TABLE_HASH_EXT_STATS_PKTS_IN_ADD(t, n_pkts_in);
 448
 449         for ( ; pkts_mask; ) {
 450                 struct bucket *bkt0, *bkt;
 451                 struct rte_mbuf *pkt;
 452                 uint8_t *key;
 453                 uint64_t pkt_mask, sig;
 454                 uint32_t pkt_index, bkt_index, i;
 455
 456                 pkt_index = __builtin_ctzll(pkts_mask);
 457                 pkt_mask = 1LLU << pkt_index;
 458                 pkts_mask &= ~pkt_mask;
 459
 460                 pkt = pkts[pkt_index];
 461                 key = RTE_MBUF_METADATA_UINT8_PTR(pkt, t->key_offset);
 462                 if (dosig)
 463                         sig = (uint64_t) t->f_hash(key, t->key_size, t->seed);
 464                 else
 465                         sig = RTE_MBUF_METADATA_UINT32(pkt,
 466                                 t->signature_offset);
 467
 468                 bkt_index = sig & t->bucket_mask;
 469                 bkt0 = &t->buckets[bkt_index];
 470                 sig = (sig >> 16) | 1LLU;
 471
 472                 /* Key is present in the bucket */
 473                 for (bkt = bkt0; bkt != NULL; bkt = BUCKET_NEXT(bkt))
 474                         for (i = 0; i < KEYS_PER_BUCKET; i++) {
 475                                 uint64_t bkt_sig = (uint64_t) bkt->sig[i];
 476                                 uint32_t bkt_key_index = bkt->key_pos[i];
 477                                 uint8_t *bkt_key = &t->key_mem[bkt_key_index <<
 478                                         t->key_size_shl];
 479
 480                                 if ((sig == bkt_sig) && (memcmp(key, bkt_key,
 481                                         t->key_size) == 0)) {
 482                                         uint8_t *data = &t->data_mem[
 483                                         bkt_key_index << t->data_size_shl];
 484
 485                                         pkts_mask_out |= pkt_mask;
 486                                         entries[pkt_index] = (void *) data;
 487                                         break;
 488                                 }
 489                         }
 490         }
 491
 492         *lookup_hit_mask = pkts_mask_out;
 493         RTE_TABLE_HASH_EXT_STATS_PKTS_LOOKUP_MISS(t, n_pkts_in - __builtin_popcountll(pkts_mask_out));
 494         return 0;
 495 }
 496
 497 /***
 498  *
 499  * mask = match bitmask
 500  * match = at least one match
 501  * match_many = more than one match
 502  * match_pos = position of first match
 503  *
 504  *----------------------------------------
 505  * mask          match   match_many       match_pos
 506  *----------------------------------------
 507  * 0000          0               0                        00
 508  * 0001          1               0                        00
 509  * 0010          1               0                        01
 510  * 0011          1               1                        00
 511  *----------------------------------------
 512  * 0100          1               0                        10
 513  * 0101          1               1                        00
 514  * 0110          1               1                        01
 515  * 0111          1               1                        00
 516  *----------------------------------------
 517  * 1000          1               0                        11
 518  * 1001          1               1                        00
 519  * 1010          1               1                        01
 520  * 1011          1               1                        00
 521  *----------------------------------------
 522  * 1100          1               1                        10
 523  * 1101          1               1                        00
 524  * 1110          1               1                        01
 525  * 1111          1               1                        00
 526  *----------------------------------------
 527  *
 528  * match = 1111_1111_1111_1110
 529  * match_many = 1111_1110_1110_1000
 530  * match_pos = 0001_0010_0001_0011__0001_0010_0001_0000
 531  *
 532  * match = 0xFFFELLU
 533  * match_many = 0xFEE8LLU
 534  * match_pos = 0x12131210LLU
 535  *
 536  ***/
 537
 538 #define LUT_MATCH                                               0xFFFELLU
 539 #define LUT_MATCH_MANY                                          0xFEE8LLU
 540 #define LUT_MATCH_POS                                           0x12131210LLU
 541
 542 #define lookup_cmp_sig(mbuf_sig, bucket, match, match_many, match_pos)  \
 543 {                                                                       \
 544         uint64_t bucket_sig[4], mask[4], mask_all;                      \
 545                                                                         \
 546         bucket_sig[0] = bucket->sig[0];                                 \
 547         bucket_sig[1] = bucket->sig[1];                                 \
 548         bucket_sig[2] = bucket->sig[2];                                 \
 549         bucket_sig[3] = bucket->sig[3];                                 \
 550                                                                         \
 551         bucket_sig[0] ^= mbuf_sig;                                      \
 552         bucket_sig[1] ^= mbuf_sig;                                      \
 553         bucket_sig[2] ^= mbuf_sig;                                      \
 554         bucket_sig[3] ^= mbuf_sig;                                      \
 555                                                                         \
 556         mask[0] = 0;                                                    \
 557         mask[1] = 0;                                                    \
 558         mask[2] = 0;                                                    \
 559         mask[3] = 0;                                                    \
 560                                                                         \
 561         if (bucket_sig[0] == 0)                                         \
 562                 mask[0] = 1;                                            \
 563         if (bucket_sig[1] == 0)                                         \
 564                 mask[1] = 2;                                            \
 565         if (bucket_sig[2] == 0)                                         \
 566                 mask[2] = 4;                                            \
 567         if (bucket_sig[3] == 0)                                         \
 568                 mask[3] = 8;                                            \
 569                                                                         \
 570         mask_all = (mask[0] | mask[1]) | (mask[2] | mask[3]);           \
 571                                                                         \
 572         match = (LUT_MATCH >> mask_all) & 1;                            \
 573         match_many = (LUT_MATCH_MANY >> mask_all) & 1;                  \
 574         match_pos = (LUT_MATCH_POS >> (mask_all << 1)) & 3;             \
 575 }
 576
 577 #define lookup_cmp_key(mbuf, key, match_key, f)                         \
 578 {                                                                       \
 579         uint64_t *pkt_key = RTE_MBUF_METADATA_UINT64_PTR(mbuf, f->key_offset);\
 580         uint64_t *bkt_key = (uint64_t *) key;                           \
 581                                                                         \
 582         switch (f->key_size) {                                          \
 583         case 8:                                                         \
 584         {                                                               \
 585                 uint64_t xor = pkt_key[0] ^ bkt_key[0];                 \
 586                 match_key = 0;                                          \
 587                 if (xor == 0)                                           \
 588                         match_key = 1;                                  \
 589         }                                                               \
 590         break;                                                          \
 591                                                                         \
 592         case 16:                                                        \
 593         {                                                               \
 594                 uint64_t xor[2], or;                                    \
 595                                                                         \
 596                 xor[0] = pkt_key[0] ^ bkt_key[0];                       \
 597                 xor[1] = pkt_key[1] ^ bkt_key[1];                       \
 598                 or = xor[0] | xor[1];                                   \
 599                 match_key = 0;                                          \
 600                 if (or == 0)                                            \
 601                         match_key = 1;                                  \
 602         }                                                               \
 603         break;                                                          \
 604                                                                         \
 605         case 32:                                                        \
 606         {                                                               \
 607                 uint64_t xor[4], or;                                    \
 608                                                                         \
 609                 xor[0] = pkt_key[0] ^ bkt_key[0];                       \
 610                 xor[1] = pkt_key[1] ^ bkt_key[1];                       \
 611                 xor[2] = pkt_key[2] ^ bkt_key[2];                       \
 612                 xor[3] = pkt_key[3] ^ bkt_key[3];                       \
 613                 or = xor[0] | xor[1] | xor[2] | xor[3];                 \
 614                 match_key = 0;                                          \
 615                 if (or == 0)                                            \
 616                         match_key = 1;                                  \
 617         }                                                               \
 618         break;                                                          \
 619                                                                         \
 620         case 64:                                                        \
 621         {                                                               \
 622                 uint64_t xor[8], or;                                    \
 623                                                                         \
 624                 xor[0] = pkt_key[0] ^ bkt_key[0];                       \
 625                 xor[1] = pkt_key[1] ^ bkt_key[1];                       \
 626                 xor[2] = pkt_key[2] ^ bkt_key[2];                       \
 627                 xor[3] = pkt_key[3] ^ bkt_key[3];                       \
 628                 xor[4] = pkt_key[4] ^ bkt_key[4];                       \
 629                 xor[5] = pkt_key[5] ^ bkt_key[5];                       \
 630                 xor[6] = pkt_key[6] ^ bkt_key[6];                       \
 631                 xor[7] = pkt_key[7] ^ bkt_key[7];                       \
 632                 or = xor[0] | xor[1] | xor[2] | xor[3] |                \
 633                         xor[4] | xor[5] | xor[6] | xor[7];              \
 634                 match_key = 0;                                          \
 635                 if (or == 0)                                            \
 636                         match_key = 1;                                  \
 637         }                                                               \
 638         break;                                                          \
 639                                                                         \
 640         default:                                                        \
 641                 match_key = 0;                                          \
 642                 if (memcmp(pkt_key, bkt_key, f->key_size) == 0)         \
 643                         match_key = 1;                                  \
 644         }                                                               \
 645 }
 646
 647 #define lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index) \
 648 {                                                                       \
 649         uint64_t pkt00_mask, pkt01_mask;                                \
 650         struct rte_mbuf *mbuf00, *mbuf01;                               \
 651                                                                         \
 652         pkt00_index = __builtin_ctzll(pkts_mask);                       \
 653         pkt00_mask = 1LLU << pkt00_index;                               \
 654         pkts_mask &= ~pkt00_mask;                                       \
 655         mbuf00 = pkts[pkt00_index];                                     \
 656                                                                         \
 657         pkt01_index = __builtin_ctzll(pkts_mask);                       \
 658         pkt01_mask = 1LLU << pkt01_index;                               \
 659         pkts_mask &= ~pkt01_mask;                                       \
 660         mbuf01 = pkts[pkt01_index];                                     \
 661                                                                         \
 662         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, 0));          \
 663         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, 0));          \
 664 }
 665
 666 #define lookup2_stage0_with_odd_support(t, g, pkts, pkts_mask, pkt00_index, \
 667         pkt01_index)                                                    \
 668 {                                                                       \
 669         uint64_t pkt00_mask, pkt01_mask;                                \
 670         struct rte_mbuf *mbuf00, *mbuf01;                               \
 671                                                                         \
 672         pkt00_index = __builtin_ctzll(pkts_mask);                       \
 673         pkt00_mask = 1LLU << pkt00_index;                               \
 674         pkts_mask &= ~pkt00_mask;                                       \
 675         mbuf00 = pkts[pkt00_index];                                     \
 676                                                                         \
 677         pkt01_index = __builtin_ctzll(pkts_mask);                       \
 678         if (pkts_mask == 0)                                             \
 679                 pkt01_index = pkt00_index;                              \
 680         pkt01_mask = 1LLU << pkt01_index;                               \
 681         pkts_mask &= ~pkt01_mask;                                       \
 682         mbuf01 = pkts[pkt01_index];                                     \
 683                                                                         \
 684         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, 0));          \
 685         rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, 0));          \
 686 }
 687
 688 #define lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index)            \
 689 {                                                                       \
 690         struct grinder *g10, *g11;                                      \
 691         uint64_t sig10, sig11, bkt10_index, bkt11_index;                \
 692         struct rte_mbuf *mbuf10, *mbuf11;                               \
 693         struct bucket *bkt10, *bkt11, *buckets = t->buckets;            \
 694         uint64_t bucket_mask = t->bucket_mask;                          \
 695         uint32_t signature_offset = t->signature_offset;                \
 696                                                                         \
 697         mbuf10 = pkts[pkt10_index];                                     \
 698         sig10 = (uint64_t) RTE_MBUF_METADATA_UINT32(mbuf10, signature_offset);\
 699         bkt10_index = sig10 & bucket_mask;                              \
 700         bkt10 = &buckets[bkt10_index];                                  \
 701                                                                         \
 702         mbuf11 = pkts[pkt11_index];                                     \
 703         sig11 = (uint64_t) RTE_MBUF_METADATA_UINT32(mbuf11, signature_offset);\
 704         bkt11_index = sig11 & bucket_mask;                              \
 705         bkt11 = &buckets[bkt11_index];                                  \
 706                                                                         \
 707         rte_prefetch0(bkt10);                                           \
 708         rte_prefetch0(bkt11);                                           \
 709                                                                         \
 710         g10 = &g[pkt10_index];                                          \
 711         g10->sig = sig10;                                               \
 712         g10->bkt = bkt10;                                               \
 713                                                                         \
 714         g11 = &g[pkt11_index];                                          \
 715         g11->sig = sig11;                                               \
 716         g11->bkt = bkt11;                                               \
 717 }
 718
 719 #define lookup2_stage1_dosig(t, g, pkts, pkt10_index, pkt11_index)      \
 720 {                                                                       \
 721         struct grinder *g10, *g11;                                      \
 722         uint64_t sig10, sig11, bkt10_index, bkt11_index;                \
 723         struct rte_mbuf *mbuf10, *mbuf11;                               \
 724         struct bucket *bkt10, *bkt11, *buckets = t->buckets;            \
 725         uint8_t *key10, *key11;                                         \
 726         uint64_t bucket_mask = t->bucket_mask;                          \
 727         rte_table_hash_op_hash f_hash = t->f_hash;                      \
 728         uint64_t seed = t->seed;                                        \
 729         uint32_t key_size = t->key_size;                                \
 730         uint32_t key_offset = t->key_offset;                            \
 731                                                                         \
 732         mbuf10 = pkts[pkt10_index];                                     \
 733         key10 = RTE_MBUF_METADATA_UINT8_PTR(mbuf10, key_offset);        \
 734         sig10 = (uint64_t) f_hash(key10, key_size, seed);               \
 735         bkt10_index = sig10 & bucket_mask;                              \
 736         bkt10 = &buckets[bkt10_index];                                  \
 737                                                                         \
 738         mbuf11 = pkts[pkt11_index];                                     \
 739         key11 = RTE_MBUF_METADATA_UINT8_PTR(mbuf11, key_offset);        \
 740         sig11 = (uint64_t) f_hash(key11, key_size, seed);               \
 741         bkt11_index = sig11 & bucket_mask;                              \
 742         bkt11 = &buckets[bkt11_index];                                  \
 743                                                                         \
 744         rte_prefetch0(bkt10);                                           \
 745         rte_prefetch0(bkt11);                                           \
 746                                                                         \
 747         g10 = &g[pkt10_index];                                          \
 748         g10->sig = sig10;                                               \
 749         g10->bkt = bkt10;                                               \
 750                                                                         \
 751         g11 = &g[pkt11_index];                                          \
 752         g11->sig = sig11;                                               \
 753         g11->bkt = bkt11;                                               \
 754 }
 755
 756 #define lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many)\
 757 {                                                                       \
 758         struct grinder *g20, *g21;                                      \
 759         uint64_t sig20, sig21;                                          \
 760         struct bucket *bkt20, *bkt21;                                   \
 761         uint8_t *key20, *key21, *key_mem = t->key_mem;                  \
 762         uint64_t match20, match21, match_many20, match_many21;          \
 763         uint64_t match_pos20, match_pos21;                              \
 764         uint32_t key20_index, key21_index, key_size_shl = t->key_size_shl;\
 765                                                                         \
 766         g20 = &g[pkt20_index];                                          \
 767         sig20 = g20->sig;                                               \
 768         bkt20 = g20->bkt;                                               \
 769         sig20 = (sig20 >> 16) | 1LLU;                                   \
 770         lookup_cmp_sig(sig20, bkt20, match20, match_many20, match_pos20);\
 771         match20 <<= pkt20_index;                                        \
 772         match_many20 |= BUCKET_NEXT_VALID(bkt20);                       \
 773         match_many20 <<= pkt20_index;                                   \
 774         key20_index = bkt20->key_pos[match_pos20];                      \
 775         key20 = &key_mem[key20_index << key_size_shl];                  \
 776                                                                         \
 777         g21 = &g[pkt21_index];                                          \
 778         sig21 = g21->sig;                                               \
 779         bkt21 = g21->bkt;                                               \
 780         sig21 = (sig21 >> 16) | 1LLU;                                   \
 781         lookup_cmp_sig(sig21, bkt21, match21, match_many21, match_pos21);\
 782         match21 <<= pkt21_index;                                        \
 783         match_many21 |= BUCKET_NEXT_VALID(bkt21);                       \
 784         match_many21 <<= pkt21_index;                                   \
 785         key21_index = bkt21->key_pos[match_pos21];                      \
 786         key21 = &key_mem[key21_index << key_size_shl];                  \
 787                                                                         \
 788         rte_prefetch0(key20);                                           \
 789         rte_prefetch0(key21);                                           \
 790                                                                         \
 791         pkts_mask_match_many |= match_many20 | match_many21;            \
 792                                                                         \
 793         g20->match = match20;                                           \
 794         g20->key_index = key20_index;                                   \
 795                                                                         \
 796         g21->match = match21;                                           \
 797         g21->key_index = key21_index;                                   \
 798 }
 799
 800 #define lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out, \
 801         entries)                                                        \
 802 {                                                                       \
 803         struct grinder *g30, *g31;                                      \
 804         struct rte_mbuf *mbuf30, *mbuf31;                               \
 805         uint8_t *key30, *key31, *key_mem = t->key_mem;                  \
 806         uint8_t *data30, *data31, *data_mem = t->data_mem;              \
 807         uint64_t match30, match31, match_key30, match_key31, match_keys;\
 808         uint32_t key30_index, key31_index;                              \
 809         uint32_t key_size_shl = t->key_size_shl;                        \
 810         uint32_t data_size_shl = t->data_size_shl;                      \
 811                                                                         \
 812         mbuf30 = pkts[pkt30_index];                                     \
 813         g30 = &g[pkt30_index];                                          \
 814         match30 = g30->match;                                           \
 815         key30_index = g30->key_index;                                   \
 816         key30 = &key_mem[key30_index << key_size_shl];                  \
 817         lookup_cmp_key(mbuf30, key30, match_key30, t);                  \
 818         match_key30 <<= pkt30_index;                                    \
 819         match_key30 &= match30;                                         \
 820         data30 = &data_mem[key30_index << data_size_shl];               \
 821         entries[pkt30_index] = data30;                                  \
 822                                                                         \
 823         mbuf31 = pkts[pkt31_index];                                     \
 824         g31 = &g[pkt31_index];                                          \
 825         match31 = g31->match;                                           \
 826         key31_index = g31->key_index;                                   \
 827         key31 = &key_mem[key31_index << key_size_shl];                  \
 828         lookup_cmp_key(mbuf31, key31, match_key31, t);                  \
 829         match_key31 <<= pkt31_index;                                    \
 830         match_key31 &= match31;                                         \
 831         data31 = &data_mem[key31_index << data_size_shl];               \
 832         entries[pkt31_index] = data31;                                  \
 833                                                                         \
 834         rte_prefetch0(data30);                                          \
 835         rte_prefetch0(data31);                                          \
 836                                                                         \
 837         match_keys = match_key30 | match_key31;                         \
 838         pkts_mask_out |= match_keys;                                    \
 839 }
 840
 841 /***
 842 * The lookup function implements a 4-stage pipeline, with each stage processing
 843 * two different packets. The purpose of pipelined implementation is to hide the
 844 * latency of prefetching the data structures and loosen the data dependency
 845 * between instructions.
 846 *
 847 *  p00  _______   p10  _______   p20  _______   p30  _______
 848 *----->|       |----->|       |----->|       |----->|       |----->
 849 *      |   0   |      |   1   |      |   2   |      |   3   |
 850 *----->|_______|----->|_______|----->|_______|----->|_______|----->
 851 *  p01            p11            p21            p31
 852 *
 853 * The naming convention is:
 854 *    pXY = packet Y of stage X, X = 0 .. 3, Y = 0 .. 1
 855 *
 856 ***/
 857 static int rte_table_hash_ext_lookup(
 858         void *table,
 859         struct rte_mbuf **pkts,
 860         uint64_t pkts_mask,
 861         uint64_t *lookup_hit_mask,
 862         void **entries)
 863 {
 864         struct rte_table_hash *t = (struct rte_table_hash *) table;
 865         struct grinder *g = t->grinders;
 866         uint64_t pkt00_index, pkt01_index, pkt10_index, pkt11_index;
 867         uint64_t pkt20_index, pkt21_index, pkt30_index, pkt31_index;
 868         uint64_t pkts_mask_out = 0, pkts_mask_match_many = 0;
 869         int status = 0;
 870
 871         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
 872         RTE_TABLE_HASH_EXT_STATS_PKTS_IN_ADD(t, n_pkts_in);
 873
 874         /* Cannot run the pipeline with less than 7 packets */
 875         if (__builtin_popcountll(pkts_mask) < 7)
 876                 return rte_table_hash_ext_lookup_unoptimized(table, pkts,
 877                         pkts_mask, lookup_hit_mask, entries, 0);
 878
 879         /* Pipeline stage 0 */
 880         lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index);
 881
 882         /* Pipeline feed */
 883         pkt10_index = pkt00_index;
 884         pkt11_index = pkt01_index;
 885
 886         /* Pipeline stage 0 */
 887         lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index);
 888
 889         /* Pipeline stage 1 */
 890         lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index);
 891
 892         /* Pipeline feed */
 893         pkt20_index = pkt10_index;
 894         pkt21_index = pkt11_index;
 895         pkt10_index = pkt00_index;
 896         pkt11_index = pkt01_index;
 897
 898         /* Pipeline stage 0 */
 899         lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index);
 900
 901         /* Pipeline stage 1 */
 902         lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index);
 903
 904         /* Pipeline stage 2 */
 905         lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many);
 906
 907         /*
 908         * Pipeline run
 909         *
 910         */
 911         for ( ; pkts_mask; ) {
 912                 /* Pipeline feed */
 913                 pkt30_index = pkt20_index;
 914                 pkt31_index = pkt21_index;
 915                 pkt20_index = pkt10_index;
 916                 pkt21_index = pkt11_index;
 917                 pkt10_index = pkt00_index;
 918                 pkt11_index = pkt01_index;
 919
 920                 /* Pipeline stage 0 */
 921                 lookup2_stage0_with_odd_support(t, g, pkts, pkts_mask,
 922                         pkt00_index, pkt01_index);
 923
 924                 /* Pipeline stage 1 */
 925                 lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index);
 926
 927                 /* Pipeline stage 2 */
 928                 lookup2_stage2(t, g, pkt20_index, pkt21_index,
 929                         pkts_mask_match_many);
 930
 931                 /* Pipeline stage 3 */
 932                 lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index,
 933                         pkts_mask_out, entries);
 934         }
 935
 936         /* Pipeline feed */
 937         pkt30_index = pkt20_index;
 938         pkt31_index = pkt21_index;
 939         pkt20_index = pkt10_index;
 940         pkt21_index = pkt11_index;
 941         pkt10_index = pkt00_index;
 942         pkt11_index = pkt01_index;
 943
 944         /* Pipeline stage 1 */
 945         lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index);
 946
 947         /* Pipeline stage 2 */
 948         lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many);
 949
 950         /* Pipeline stage 3 */
 951         lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out,
 952                 entries);
 953
 954         /* Pipeline feed */
 955         pkt30_index = pkt20_index;
 956         pkt31_index = pkt21_index;
 957         pkt20_index = pkt10_index;
 958         pkt21_index = pkt11_index;
 959
 960         /* Pipeline stage 2 */
 961         lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many);
 962
 963         /* Pipeline stage 3 */
 964         lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out,
 965                 entries);
 966
 967         /* Pipeline feed */
 968         pkt30_index = pkt20_index;
 969         pkt31_index = pkt21_index;
 970
 971         /* Pipeline stage 3 */
 972         lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out,
 973                 entries);
 974
 975         /* Slow path */
 976         pkts_mask_match_many &= ~pkts_mask_out;
 977         if (pkts_mask_match_many) {
 978                 uint64_t pkts_mask_out_slow = 0;
 979
 980                 status = rte_table_hash_ext_lookup_unoptimized(table, pkts,
 981                         pkts_mask_match_many, &pkts_mask_out_slow, entries, 0);
 982                 pkts_mask_out |= pkts_mask_out_slow;
 983         }
 984
 985         *lookup_hit_mask = pkts_mask_out;
 986         RTE_TABLE_HASH_EXT_STATS_PKTS_LOOKUP_MISS(t, n_pkts_in - __builtin_popcountll(pkts_mask_out));
 987         return status;
 988 }
 989
 990 static int rte_table_hash_ext_lookup_dosig(
 991         void *table,
 992         struct rte_mbuf **pkts,
 993         uint64_t pkts_mask,
 994         uint64_t *lookup_hit_mask,
 995         void **entries)
 996 {
 997         struct rte_table_hash *t = (struct rte_table_hash *) table;
 998         struct grinder *g = t->grinders;
 999         uint64_t pkt00_index, pkt01_index, pkt10_index, pkt11_index;
1000         uint64_t pkt20_index, pkt21_index, pkt30_index, pkt31_index;
1001         uint64_t pkts_mask_out = 0, pkts_mask_match_many = 0;
1002         int status = 0;
1003
1004         __rte_unused uint32_t n_pkts_in = __builtin_popcountll(pkts_mask);
1005         RTE_TABLE_HASH_EXT_STATS_PKTS_IN_ADD(t, n_pkts_in);
1006
1007         /* Cannot run the pipeline with less than 7 packets */
1008         if (__builtin_popcountll(pkts_mask) < 7)
1009                 return rte_table_hash_ext_lookup_unoptimized(table, pkts,
1010                         pkts_mask, lookup_hit_mask, entries, 1);
1011
1012         /* Pipeline stage 0 */
1013         lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index);
1014
1015         /* Pipeline feed */
1016         pkt10_index = pkt00_index;
1017         pkt11_index = pkt01_index;
1018
1019         /* Pipeline stage 0 */
1020         lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index);
1021
1022         /* Pipeline stage 1 */
1023         lookup2_stage1_dosig(t, g, pkts, pkt10_index, pkt11_index);
1024
1025         /* Pipeline feed */
1026         pkt20_index = pkt10_index;
1027         pkt21_index = pkt11_index;
1028         pkt10_index = pkt00_index;
1029         pkt11_index = pkt01_index;
1030
1031         /* Pipeline stage 0 */
1032         lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index);
1033
1034         /* Pipeline stage 1 */
1035         lookup2_stage1_dosig(t, g, pkts, pkt10_index, pkt11_index);
1036
1037         /* Pipeline stage 2 */
1038         lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many);
1039
1040         /*
1041         * Pipeline run
1042         *
1043         */
1044         for ( ; pkts_mask; ) {
1045                 /* Pipeline feed */
1046                 pkt30_index = pkt20_index;
1047                 pkt31_index = pkt21_index;
1048                 pkt20_index = pkt10_index;
1049                 pkt21_index = pkt11_index;
1050                 pkt10_index = pkt00_index;
1051                 pkt11_index = pkt01_index;
1052
1053                 /* Pipeline stage 0 */
1054                 lookup2_stage0_with_odd_support(t, g, pkts, pkts_mask,
1055                         pkt00_index, pkt01_index);
1056
1057                 /* Pipeline stage 1 */
1058                 lookup2_stage1_dosig(t, g, pkts, pkt10_index, pkt11_index);
1059
1060                 /* Pipeline stage 2 */
1061                 lookup2_stage2(t, g, pkt20_index, pkt21_index,
1062                         pkts_mask_match_many);
1063
1064                 /* Pipeline stage 3 */
1065                 lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index,
1066                         pkts_mask_out, entries);
1067         }
1068
1069         /* Pipeline feed */
1070         pkt30_index = pkt20_index;
1071         pkt31_index = pkt21_index;
1072         pkt20_index = pkt10_index;
1073         pkt21_index = pkt11_index;
1074         pkt10_index = pkt00_index;
1075         pkt11_index = pkt01_index;
1076
1077         /* Pipeline stage 1 */
1078         lookup2_stage1_dosig(t, g, pkts, pkt10_index, pkt11_index);
1079
1080         /* Pipeline stage 2 */
1081         lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many);
1082
1083         /* Pipeline stage 3 */
1084         lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out,
1085                 entries);
1086
1087         /* Pipeline feed */
1088         pkt30_index = pkt20_index;
1089         pkt31_index = pkt21_index;
1090         pkt20_index = pkt10_index;
1091         pkt21_index = pkt11_index;
1092
1093         /* Pipeline stage 2 */
1094         lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many);
1095
1096         /* Pipeline stage 3 */
1097         lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out,
1098                 entries);
1099
1100         /* Pipeline feed */
1101         pkt30_index = pkt20_index;
1102         pkt31_index = pkt21_index;
1103
1104         /* Pipeline stage 3 */
1105         lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out,
1106                 entries);
1107
1108         /* Slow path */
1109         pkts_mask_match_many &= ~pkts_mask_out;
1110         if (pkts_mask_match_many) {
1111                 uint64_t pkts_mask_out_slow = 0;
1112
1113                 status = rte_table_hash_ext_lookup_unoptimized(table, pkts,
1114                         pkts_mask_match_many, &pkts_mask_out_slow, entries, 1);
1115                 pkts_mask_out |= pkts_mask_out_slow;
1116         }
1117
1118         *lookup_hit_mask = pkts_mask_out;
1119         RTE_TABLE_HASH_EXT_STATS_PKTS_LOOKUP_MISS(t, n_pkts_in - __builtin_popcountll(pkts_mask_out));
1120         return status;
1121 }
1122
1123 static int
1124 rte_table_hash_ext_stats_read(void *table, struct rte_table_stats *stats, int clear)
1125 {
1126         struct rte_table_hash *t = (struct rte_table_hash *) table;
1127
1128         if (stats != NULL)
1129                 memcpy(stats, &t->stats, sizeof(t->stats));
1130
1131         if (clear)
1132                 memset(&t->stats, 0, sizeof(t->stats));
1133
1134         return 0;
1135 }
1136
1137 struct rte_table_ops rte_table_hash_ext_ops      = {
1138         .f_create = rte_table_hash_ext_create,
1139         .f_free = rte_table_hash_ext_free,
1140         .f_add = rte_table_hash_ext_entry_add,
1141         .f_delete = rte_table_hash_ext_entry_delete,
1142         .f_add_bulk = NULL,
1143         .f_delete_bulk = NULL,
1144         .f_lookup = rte_table_hash_ext_lookup,
1145         .f_stats = rte_table_hash_ext_stats_read,
1146 };
1147
1148 struct rte_table_ops rte_table_hash_ext_dosig_ops  = {
1149         .f_create = rte_table_hash_ext_create,
1150         .f_free = rte_table_hash_ext_free,
1151         .f_add = rte_table_hash_ext_entry_add,
1152         .f_delete = rte_table_hash_ext_entry_delete,
1153         .f_add_bulk = NULL,
1154         .f_delete_bulk = NULL,
1155         .f_lookup = rte_table_hash_ext_lookup_dosig,
1156         .f_stats = rte_table_hash_ext_stats_read,
1157 };