1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2010-2017 Intel Corporation 3 */ 4 5 #include <stdalign.h> 6 #include <stdio.h> 7 #include <string.h> 8 9 #include <rte_common.h> 10 #include <rte_malloc.h> 11 #include <rte_log.h> 12 13 #include "rte_table_hash.h" 14 15 #include "table_log.h" 16 17 #define KEYS_PER_BUCKET 4 18 19 struct bucket { 20 union { 21 uintptr_t next; 22 uint64_t lru_list; 23 }; 24 uint16_t sig[KEYS_PER_BUCKET]; 25 uint32_t key_pos[KEYS_PER_BUCKET]; 26 }; 27 28 #define BUCKET_NEXT(bucket) \ 29 ((void *) ((bucket)->next & (~1LU))) 30 31 #define BUCKET_NEXT_VALID(bucket) \ 32 ((bucket)->next & 1LU) 33 34 #define BUCKET_NEXT_SET(bucket, bucket_next) \ 35 do \ 36 (bucket)->next = (((uintptr_t) ((void *) (bucket_next))) | 1LU);\ 37 while (0) 38 39 #define BUCKET_NEXT_SET_NULL(bucket) \ 40 do \ 41 (bucket)->next = 0; \ 42 while (0) 43 44 #define BUCKET_NEXT_COPY(bucket, bucket2) \ 45 do \ 46 (bucket)->next = (bucket2)->next; \ 47 while (0) 48 49 #ifdef RTE_TABLE_STATS_COLLECT 50 51 #define RTE_TABLE_HASH_EXT_STATS_PKTS_IN_ADD(table, val) \ 52 table->stats.n_pkts_in += val 53 #define RTE_TABLE_HASH_EXT_STATS_PKTS_LOOKUP_MISS(table, val) \ 54 table->stats.n_pkts_lookup_miss += val 55 56 #else 57 58 #define RTE_TABLE_HASH_EXT_STATS_PKTS_IN_ADD(table, val) 59 #define RTE_TABLE_HASH_EXT_STATS_PKTS_LOOKUP_MISS(table, val) 60 61 #endif 62 63 struct grinder { 64 struct bucket *bkt; 65 uint64_t sig; 66 uint64_t match; 67 uint32_t key_index; 68 }; 69 70 struct rte_table_hash { 71 struct rte_table_stats stats; 72 73 /* Input parameters */ 74 uint32_t key_size; 75 uint32_t entry_size; 76 uint32_t n_keys; 77 uint32_t n_buckets; 78 uint32_t n_buckets_ext; 79 rte_table_hash_op_hash f_hash; 80 uint64_t seed; 81 uint32_t key_offset; 82 83 /* Internal */ 84 uint64_t bucket_mask; 85 uint32_t key_size_shl; 86 uint32_t data_size_shl; 87 uint32_t key_stack_tos; 88 uint32_t bkt_ext_stack_tos; 89 90 /* Grinder */ 91 struct grinder grinders[RTE_PORT_IN_BURST_SIZE_MAX]; 92 93 /* Tables */ 94 uint64_t *key_mask; 95 struct bucket *buckets; 96 struct bucket *buckets_ext; 97 uint8_t *key_mem; 98 uint8_t *data_mem; 99 uint32_t *key_stack; 100 uint32_t *bkt_ext_stack; 101 102 /* Table memory */ 103 alignas(RTE_CACHE_LINE_SIZE) uint8_t memory[]; 104 }; 105 106 static int 107 keycmp(void *a, void *b, void *b_mask, uint32_t n_bytes) 108 { 109 uint64_t *a64 = a, *b64 = b, *b_mask64 = b_mask; 110 uint32_t i; 111 112 for (i = 0; i < n_bytes / sizeof(uint64_t); i++) 113 if (a64[i] != (b64[i] & b_mask64[i])) 114 return 1; 115 116 return 0; 117 } 118 119 static void 120 keycpy(void *dst, void *src, void *src_mask, uint32_t n_bytes) 121 { 122 uint64_t *dst64 = dst, *src64 = src, *src_mask64 = src_mask; 123 uint32_t i; 124 125 for (i = 0; i < n_bytes / sizeof(uint64_t); i++) 126 dst64[i] = src64[i] & src_mask64[i]; 127 } 128 129 static int 130 check_params_create(struct rte_table_hash_params *params) 131 { 132 /* name */ 133 if (params->name == NULL) { 134 TABLE_LOG(ERR, "%s: name invalid value", __func__); 135 return -EINVAL; 136 } 137 138 /* key_size */ 139 if ((params->key_size < sizeof(uint64_t)) || 140 (!rte_is_power_of_2(params->key_size))) { 141 TABLE_LOG(ERR, "%s: key_size invalid value", __func__); 142 return -EINVAL; 143 } 144 145 /* n_keys */ 146 if (params->n_keys == 0) { 147 TABLE_LOG(ERR, "%s: n_keys invalid value", __func__); 148 return -EINVAL; 149 } 150 151 /* n_buckets */ 152 if ((params->n_buckets == 0) || 153 (!rte_is_power_of_2(params->n_buckets))) { 154 TABLE_LOG(ERR, "%s: n_buckets invalid value", __func__); 155 return -EINVAL; 156 } 157 158 /* f_hash */ 159 if (params->f_hash == NULL) { 160 TABLE_LOG(ERR, "%s: f_hash invalid value", __func__); 161 return -EINVAL; 162 } 163 164 return 0; 165 } 166 167 static void * 168 rte_table_hash_ext_create(void *params, int socket_id, uint32_t entry_size) 169 { 170 struct rte_table_hash_params *p = params; 171 struct rte_table_hash *t; 172 uint64_t table_meta_sz, key_mask_sz, bucket_sz, bucket_ext_sz, key_sz; 173 uint64_t key_stack_sz, bkt_ext_stack_sz, data_sz, total_size; 174 uint64_t key_mask_offset, bucket_offset, bucket_ext_offset, key_offset; 175 uint64_t key_stack_offset, bkt_ext_stack_offset, data_offset; 176 uint32_t n_buckets_ext, i; 177 178 /* Check input parameters */ 179 if ((check_params_create(p) != 0) || 180 (!rte_is_power_of_2(entry_size)) || 181 ((sizeof(struct rte_table_hash) % RTE_CACHE_LINE_SIZE) != 0) || 182 (sizeof(struct bucket) != (RTE_CACHE_LINE_SIZE / 2))) 183 return NULL; 184 185 /* 186 * Table dimensioning 187 * 188 * Objective: Pick the number of bucket extensions (n_buckets_ext) so that 189 * it is guaranteed that n_keys keys can be stored in the table at any time. 190 * 191 * The worst case scenario takes place when all the n_keys keys fall into 192 * the same bucket. Actually, due to the KEYS_PER_BUCKET scheme, the worst 193 * case takes place when (n_keys - KEYS_PER_BUCKET + 1) keys fall into the 194 * same bucket, while the remaining (KEYS_PER_BUCKET - 1) keys each fall 195 * into a different bucket. This case defeats the purpose of the hash table. 196 * It indicates unsuitable f_hash or n_keys to n_buckets ratio. 197 * 198 * n_buckets_ext = n_keys / KEYS_PER_BUCKET + KEYS_PER_BUCKET - 1 199 */ 200 n_buckets_ext = p->n_keys / KEYS_PER_BUCKET + KEYS_PER_BUCKET - 1; 201 202 /* Memory allocation */ 203 table_meta_sz = RTE_CACHE_LINE_ROUNDUP(sizeof(struct rte_table_hash)); 204 key_mask_sz = RTE_CACHE_LINE_ROUNDUP(p->key_size); 205 bucket_sz = RTE_CACHE_LINE_ROUNDUP(p->n_buckets * sizeof(struct bucket)); 206 bucket_ext_sz = 207 RTE_CACHE_LINE_ROUNDUP(n_buckets_ext * sizeof(struct bucket)); 208 key_sz = RTE_CACHE_LINE_ROUNDUP(p->n_keys * p->key_size); 209 key_stack_sz = RTE_CACHE_LINE_ROUNDUP(p->n_keys * sizeof(uint32_t)); 210 bkt_ext_stack_sz = 211 RTE_CACHE_LINE_ROUNDUP(n_buckets_ext * sizeof(uint32_t)); 212 data_sz = RTE_CACHE_LINE_ROUNDUP(p->n_keys * entry_size); 213 total_size = table_meta_sz + key_mask_sz + bucket_sz + bucket_ext_sz + 214 key_sz + key_stack_sz + bkt_ext_stack_sz + data_sz; 215 216 if (total_size > SIZE_MAX) { 217 TABLE_LOG(ERR, "%s: Cannot allocate %" PRIu64 " bytes" 218 " for hash table %s", 219 __func__, total_size, p->name); 220 return NULL; 221 } 222 223 t = rte_zmalloc_socket(p->name, 224 (size_t)total_size, 225 RTE_CACHE_LINE_SIZE, 226 socket_id); 227 if (t == NULL) { 228 TABLE_LOG(ERR, "%s: Cannot allocate %" PRIu64 " bytes" 229 " for hash table %s", 230 __func__, total_size, p->name); 231 return NULL; 232 } 233 TABLE_LOG(INFO, "%s (%u-byte key): Hash table %s memory " 234 "footprint is %" PRIu64 " bytes", 235 __func__, p->key_size, p->name, total_size); 236 237 /* Memory initialization */ 238 t->key_size = p->key_size; 239 t->entry_size = entry_size; 240 t->n_keys = p->n_keys; 241 t->n_buckets = p->n_buckets; 242 t->n_buckets_ext = n_buckets_ext; 243 t->f_hash = p->f_hash; 244 t->seed = p->seed; 245 t->key_offset = p->key_offset; 246 247 /* Internal */ 248 t->bucket_mask = t->n_buckets - 1; 249 t->key_size_shl = rte_ctz32(p->key_size); 250 t->data_size_shl = rte_ctz32(entry_size); 251 252 /* Tables */ 253 key_mask_offset = 0; 254 bucket_offset = key_mask_offset + key_mask_sz; 255 bucket_ext_offset = bucket_offset + bucket_sz; 256 key_offset = bucket_ext_offset + bucket_ext_sz; 257 key_stack_offset = key_offset + key_sz; 258 bkt_ext_stack_offset = key_stack_offset + key_stack_sz; 259 data_offset = bkt_ext_stack_offset + bkt_ext_stack_sz; 260 261 t->key_mask = (uint64_t *) &t->memory[key_mask_offset]; 262 t->buckets = (struct bucket *) &t->memory[bucket_offset]; 263 t->buckets_ext = (struct bucket *) &t->memory[bucket_ext_offset]; 264 t->key_mem = &t->memory[key_offset]; 265 t->key_stack = (uint32_t *) &t->memory[key_stack_offset]; 266 t->bkt_ext_stack = (uint32_t *) &t->memory[bkt_ext_stack_offset]; 267 t->data_mem = &t->memory[data_offset]; 268 269 /* Key mask */ 270 if (p->key_mask == NULL) 271 memset(t->key_mask, 0xFF, p->key_size); 272 else 273 memcpy(t->key_mask, p->key_mask, p->key_size); 274 275 /* Key stack */ 276 for (i = 0; i < t->n_keys; i++) 277 t->key_stack[i] = t->n_keys - 1 - i; 278 t->key_stack_tos = t->n_keys; 279 280 /* Bucket ext stack */ 281 for (i = 0; i < t->n_buckets_ext; i++) 282 t->bkt_ext_stack[i] = t->n_buckets_ext - 1 - i; 283 t->bkt_ext_stack_tos = t->n_buckets_ext; 284 285 return t; 286 } 287 288 static int 289 rte_table_hash_ext_free(void *table) 290 { 291 struct rte_table_hash *t = table; 292 293 /* Check input parameters */ 294 if (t == NULL) 295 return -EINVAL; 296 297 rte_free(t); 298 return 0; 299 } 300 301 static int 302 rte_table_hash_ext_entry_add(void *table, void *key, void *entry, 303 int *key_found, void **entry_ptr) 304 { 305 struct rte_table_hash *t = table; 306 struct bucket *bkt0, *bkt, *bkt_prev; 307 uint64_t sig; 308 uint32_t bkt_index, i; 309 310 sig = t->f_hash(key, t->key_mask, t->key_size, t->seed); 311 bkt_index = sig & t->bucket_mask; 312 bkt0 = &t->buckets[bkt_index]; 313 sig = (sig >> 16) | 1LLU; 314 315 /* Key is present in the bucket */ 316 for (bkt = bkt0; bkt != NULL; bkt = BUCKET_NEXT(bkt)) 317 for (i = 0; i < KEYS_PER_BUCKET; i++) { 318 uint64_t bkt_sig = (uint64_t) bkt->sig[i]; 319 uint32_t bkt_key_index = bkt->key_pos[i]; 320 uint8_t *bkt_key = 321 &t->key_mem[bkt_key_index << t->key_size_shl]; 322 323 if ((sig == bkt_sig) && (keycmp(bkt_key, key, t->key_mask, 324 t->key_size) == 0)) { 325 uint8_t *data = &t->data_mem[bkt_key_index << 326 t->data_size_shl]; 327 328 memcpy(data, entry, t->entry_size); 329 *key_found = 1; 330 *entry_ptr = (void *) data; 331 return 0; 332 } 333 } 334 335 /* Key is not present in the bucket */ 336 for (bkt_prev = NULL, bkt = bkt0; bkt != NULL; bkt_prev = bkt, 337 bkt = BUCKET_NEXT(bkt)) 338 for (i = 0; i < KEYS_PER_BUCKET; i++) { 339 uint64_t bkt_sig = (uint64_t) bkt->sig[i]; 340 341 if (bkt_sig == 0) { 342 uint32_t bkt_key_index; 343 uint8_t *bkt_key, *data; 344 345 /* Allocate new key */ 346 if (t->key_stack_tos == 0) /* No free keys */ 347 return -ENOSPC; 348 349 bkt_key_index = t->key_stack[ 350 --t->key_stack_tos]; 351 352 /* Install new key */ 353 bkt_key = &t->key_mem[bkt_key_index << 354 t->key_size_shl]; 355 data = &t->data_mem[bkt_key_index << 356 t->data_size_shl]; 357 358 bkt->sig[i] = (uint16_t) sig; 359 bkt->key_pos[i] = bkt_key_index; 360 keycpy(bkt_key, key, t->key_mask, t->key_size); 361 memcpy(data, entry, t->entry_size); 362 363 *key_found = 0; 364 *entry_ptr = (void *) data; 365 return 0; 366 } 367 } 368 369 /* Bucket full: extend bucket */ 370 if ((t->bkt_ext_stack_tos > 0) && (t->key_stack_tos > 0)) { 371 uint32_t bkt_key_index; 372 uint8_t *bkt_key, *data; 373 374 /* Allocate new bucket ext */ 375 bkt_index = t->bkt_ext_stack[--t->bkt_ext_stack_tos]; 376 bkt = &t->buckets_ext[bkt_index]; 377 378 /* Chain the new bucket ext */ 379 BUCKET_NEXT_SET(bkt_prev, bkt); 380 BUCKET_NEXT_SET_NULL(bkt); 381 382 /* Allocate new key */ 383 bkt_key_index = t->key_stack[--t->key_stack_tos]; 384 bkt_key = &t->key_mem[bkt_key_index << t->key_size_shl]; 385 386 data = &t->data_mem[bkt_key_index << t->data_size_shl]; 387 388 /* Install new key into bucket */ 389 bkt->sig[0] = (uint16_t) sig; 390 bkt->key_pos[0] = bkt_key_index; 391 keycpy(bkt_key, key, t->key_mask, t->key_size); 392 memcpy(data, entry, t->entry_size); 393 394 *key_found = 0; 395 *entry_ptr = (void *) data; 396 return 0; 397 } 398 399 return -ENOSPC; 400 } 401 402 static int 403 rte_table_hash_ext_entry_delete(void *table, void *key, int *key_found, 404 void *entry) 405 { 406 struct rte_table_hash *t = table; 407 struct bucket *bkt0, *bkt, *bkt_prev; 408 uint64_t sig; 409 uint32_t bkt_index, i; 410 411 sig = t->f_hash(key, t->key_mask, t->key_size, t->seed); 412 bkt_index = sig & t->bucket_mask; 413 bkt0 = &t->buckets[bkt_index]; 414 sig = (sig >> 16) | 1LLU; 415 416 /* Key is present in the bucket */ 417 for (bkt_prev = NULL, bkt = bkt0; bkt != NULL; bkt_prev = bkt, 418 bkt = BUCKET_NEXT(bkt)) 419 for (i = 0; i < KEYS_PER_BUCKET; i++) { 420 uint64_t bkt_sig = (uint64_t) bkt->sig[i]; 421 uint32_t bkt_key_index = bkt->key_pos[i]; 422 uint8_t *bkt_key = &t->key_mem[bkt_key_index << 423 t->key_size_shl]; 424 425 if ((sig == bkt_sig) && (keycmp(bkt_key, key, t->key_mask, 426 t->key_size) == 0)) { 427 uint8_t *data = &t->data_mem[bkt_key_index << 428 t->data_size_shl]; 429 430 /* Uninstall key from bucket */ 431 bkt->sig[i] = 0; 432 *key_found = 1; 433 if (entry) 434 memcpy(entry, data, t->entry_size); 435 436 /* Free key */ 437 t->key_stack[t->key_stack_tos++] = 438 bkt_key_index; 439 440 /*Check if bucket is unused */ 441 if ((bkt_prev != NULL) && 442 (bkt->sig[0] == 0) && (bkt->sig[1] == 0) && 443 (bkt->sig[2] == 0) && (bkt->sig[3] == 0)) { 444 /* Unchain bucket */ 445 BUCKET_NEXT_COPY(bkt_prev, bkt); 446 447 /* Clear bucket */ 448 memset(bkt, 0, sizeof(struct bucket)); 449 450 /* Free bucket back to buckets ext */ 451 bkt_index = bkt - t->buckets_ext; 452 t->bkt_ext_stack[t->bkt_ext_stack_tos++] 453 = bkt_index; 454 } 455 456 return 0; 457 } 458 } 459 460 /* Key is not present in the bucket */ 461 *key_found = 0; 462 return 0; 463 } 464 465 static int rte_table_hash_ext_lookup_unoptimized( 466 void *table, 467 struct rte_mbuf **pkts, 468 uint64_t pkts_mask, 469 uint64_t *lookup_hit_mask, 470 void **entries) 471 { 472 struct rte_table_hash *t = (struct rte_table_hash *) table; 473 uint64_t pkts_mask_out = 0; 474 475 __rte_unused uint32_t n_pkts_in = rte_popcount64(pkts_mask); 476 477 for ( ; pkts_mask; ) { 478 struct bucket *bkt0, *bkt; 479 struct rte_mbuf *pkt; 480 uint8_t *key; 481 uint64_t pkt_mask, sig; 482 uint32_t pkt_index, bkt_index, i; 483 484 pkt_index = rte_ctz64(pkts_mask); 485 pkt_mask = 1LLU << pkt_index; 486 pkts_mask &= ~pkt_mask; 487 488 pkt = pkts[pkt_index]; 489 key = RTE_MBUF_METADATA_UINT8_PTR(pkt, t->key_offset); 490 sig = (uint64_t) t->f_hash(key, t->key_mask, t->key_size, t->seed); 491 492 bkt_index = sig & t->bucket_mask; 493 bkt0 = &t->buckets[bkt_index]; 494 sig = (sig >> 16) | 1LLU; 495 496 /* Key is present in the bucket */ 497 for (bkt = bkt0; bkt != NULL; bkt = BUCKET_NEXT(bkt)) 498 for (i = 0; i < KEYS_PER_BUCKET; i++) { 499 uint64_t bkt_sig = (uint64_t) bkt->sig[i]; 500 uint32_t bkt_key_index = bkt->key_pos[i]; 501 uint8_t *bkt_key = &t->key_mem[bkt_key_index << 502 t->key_size_shl]; 503 504 if ((sig == bkt_sig) && (keycmp(bkt_key, key, 505 t->key_mask, t->key_size) == 0)) { 506 uint8_t *data = &t->data_mem[ 507 bkt_key_index << t->data_size_shl]; 508 509 pkts_mask_out |= pkt_mask; 510 entries[pkt_index] = (void *) data; 511 break; 512 } 513 } 514 } 515 516 *lookup_hit_mask = pkts_mask_out; 517 return 0; 518 } 519 520 /* 521 * mask = match bitmask 522 * match = at least one match 523 * match_many = more than one match 524 * match_pos = position of first match 525 * 526 *---------------------------------------- 527 * mask match match_many match_pos 528 *---------------------------------------- 529 * 0000 0 0 00 530 * 0001 1 0 00 531 * 0010 1 0 01 532 * 0011 1 1 00 533 *---------------------------------------- 534 * 0100 1 0 10 535 * 0101 1 1 00 536 * 0110 1 1 01 537 * 0111 1 1 00 538 *---------------------------------------- 539 * 1000 1 0 11 540 * 1001 1 1 00 541 * 1010 1 1 01 542 * 1011 1 1 00 543 *---------------------------------------- 544 * 1100 1 1 10 545 * 1101 1 1 00 546 * 1110 1 1 01 547 * 1111 1 1 00 548 *---------------------------------------- 549 * 550 * match = 1111_1111_1111_1110 551 * match_many = 1111_1110_1110_1000 552 * match_pos = 0001_0010_0001_0011__0001_0010_0001_0000 553 * 554 * match = 0xFFFELLU 555 * match_many = 0xFEE8LLU 556 * match_pos = 0x12131210LLU 557 */ 558 559 #define LUT_MATCH 0xFFFELLU 560 #define LUT_MATCH_MANY 0xFEE8LLU 561 #define LUT_MATCH_POS 0x12131210LLU 562 563 #define lookup_cmp_sig(mbuf_sig, bucket, match, match_many, match_pos) \ 564 { \ 565 uint64_t bucket_sig[4], mask[4], mask_all; \ 566 \ 567 bucket_sig[0] = bucket->sig[0]; \ 568 bucket_sig[1] = bucket->sig[1]; \ 569 bucket_sig[2] = bucket->sig[2]; \ 570 bucket_sig[3] = bucket->sig[3]; \ 571 \ 572 bucket_sig[0] ^= mbuf_sig; \ 573 bucket_sig[1] ^= mbuf_sig; \ 574 bucket_sig[2] ^= mbuf_sig; \ 575 bucket_sig[3] ^= mbuf_sig; \ 576 \ 577 mask[0] = 0; \ 578 mask[1] = 0; \ 579 mask[2] = 0; \ 580 mask[3] = 0; \ 581 \ 582 if (bucket_sig[0] == 0) \ 583 mask[0] = 1; \ 584 if (bucket_sig[1] == 0) \ 585 mask[1] = 2; \ 586 if (bucket_sig[2] == 0) \ 587 mask[2] = 4; \ 588 if (bucket_sig[3] == 0) \ 589 mask[3] = 8; \ 590 \ 591 mask_all = (mask[0] | mask[1]) | (mask[2] | mask[3]); \ 592 \ 593 match = (LUT_MATCH >> mask_all) & 1; \ 594 match_many = (LUT_MATCH_MANY >> mask_all) & 1; \ 595 match_pos = (LUT_MATCH_POS >> (mask_all << 1)) & 3; \ 596 } 597 598 #define lookup_cmp_key(mbuf, key, match_key, f) \ 599 { \ 600 uint64_t *pkt_key = RTE_MBUF_METADATA_UINT64_PTR(mbuf, f->key_offset);\ 601 uint64_t *bkt_key = (uint64_t *) key; \ 602 uint64_t *key_mask = f->key_mask; \ 603 \ 604 switch (f->key_size) { \ 605 case 8: \ 606 { \ 607 uint64_t xor = (pkt_key[0] & key_mask[0]) ^ bkt_key[0]; \ 608 match_key = 0; \ 609 if (xor == 0) \ 610 match_key = 1; \ 611 } \ 612 break; \ 613 \ 614 case 16: \ 615 { \ 616 uint64_t xor[2], or; \ 617 \ 618 xor[0] = (pkt_key[0] & key_mask[0]) ^ bkt_key[0]; \ 619 xor[1] = (pkt_key[1] & key_mask[1]) ^ bkt_key[1]; \ 620 or = xor[0] | xor[1]; \ 621 match_key = 0; \ 622 if (or == 0) \ 623 match_key = 1; \ 624 } \ 625 break; \ 626 \ 627 case 32: \ 628 { \ 629 uint64_t xor[4], or; \ 630 \ 631 xor[0] = (pkt_key[0] & key_mask[0]) ^ bkt_key[0]; \ 632 xor[1] = (pkt_key[1] & key_mask[1]) ^ bkt_key[1]; \ 633 xor[2] = (pkt_key[2] & key_mask[2]) ^ bkt_key[2]; \ 634 xor[3] = (pkt_key[3] & key_mask[3]) ^ bkt_key[3]; \ 635 or = xor[0] | xor[1] | xor[2] | xor[3]; \ 636 match_key = 0; \ 637 if (or == 0) \ 638 match_key = 1; \ 639 } \ 640 break; \ 641 \ 642 case 64: \ 643 { \ 644 uint64_t xor[8], or; \ 645 \ 646 xor[0] = (pkt_key[0] & key_mask[0]) ^ bkt_key[0]; \ 647 xor[1] = (pkt_key[1] & key_mask[1]) ^ bkt_key[1]; \ 648 xor[2] = (pkt_key[2] & key_mask[2]) ^ bkt_key[2]; \ 649 xor[3] = (pkt_key[3] & key_mask[3]) ^ bkt_key[3]; \ 650 xor[4] = (pkt_key[4] & key_mask[4]) ^ bkt_key[4]; \ 651 xor[5] = (pkt_key[5] & key_mask[5]) ^ bkt_key[5]; \ 652 xor[6] = (pkt_key[6] & key_mask[6]) ^ bkt_key[6]; \ 653 xor[7] = (pkt_key[7] & key_mask[7]) ^ bkt_key[7]; \ 654 or = xor[0] | xor[1] | xor[2] | xor[3] | \ 655 xor[4] | xor[5] | xor[6] | xor[7]; \ 656 match_key = 0; \ 657 if (or == 0) \ 658 match_key = 1; \ 659 } \ 660 break; \ 661 \ 662 default: \ 663 match_key = 0; \ 664 if (keycmp(bkt_key, pkt_key, key_mask, f->key_size) == 0) \ 665 match_key = 1; \ 666 } \ 667 } 668 669 #define lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index) \ 670 { \ 671 uint64_t pkt00_mask, pkt01_mask; \ 672 struct rte_mbuf *mbuf00, *mbuf01; \ 673 uint32_t key_offset = t->key_offset; \ 674 \ 675 pkt00_index = rte_ctz64(pkts_mask); \ 676 pkt00_mask = 1LLU << pkt00_index; \ 677 pkts_mask &= ~pkt00_mask; \ 678 mbuf00 = pkts[pkt00_index]; \ 679 \ 680 pkt01_index = rte_ctz64(pkts_mask); \ 681 pkt01_mask = 1LLU << pkt01_index; \ 682 pkts_mask &= ~pkt01_mask; \ 683 mbuf01 = pkts[pkt01_index]; \ 684 \ 685 rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, key_offset));\ 686 rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, key_offset));\ 687 } 688 689 #define lookup2_stage0_with_odd_support(t, g, pkts, pkts_mask, pkt00_index, \ 690 pkt01_index) \ 691 { \ 692 uint64_t pkt00_mask, pkt01_mask; \ 693 struct rte_mbuf *mbuf00, *mbuf01; \ 694 uint32_t key_offset = t->key_offset; \ 695 \ 696 pkt00_index = rte_ctz64(pkts_mask); \ 697 pkt00_mask = 1LLU << pkt00_index; \ 698 pkts_mask &= ~pkt00_mask; \ 699 mbuf00 = pkts[pkt00_index]; \ 700 \ 701 pkt01_index = rte_ctz64(pkts_mask); \ 702 if (pkts_mask == 0) \ 703 pkt01_index = pkt00_index; \ 704 pkt01_mask = 1LLU << pkt01_index; \ 705 pkts_mask &= ~pkt01_mask; \ 706 mbuf01 = pkts[pkt01_index]; \ 707 \ 708 rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf00, key_offset));\ 709 rte_prefetch0(RTE_MBUF_METADATA_UINT8_PTR(mbuf01, key_offset));\ 710 } 711 712 #define lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index) \ 713 { \ 714 struct grinder *g10, *g11; \ 715 uint64_t sig10, sig11, bkt10_index, bkt11_index; \ 716 struct rte_mbuf *mbuf10, *mbuf11; \ 717 struct bucket *bkt10, *bkt11, *buckets = t->buckets; \ 718 uint8_t *key10, *key11; \ 719 uint64_t bucket_mask = t->bucket_mask; \ 720 rte_table_hash_op_hash f_hash = t->f_hash; \ 721 uint64_t seed = t->seed; \ 722 uint32_t key_size = t->key_size; \ 723 uint32_t key_offset = t->key_offset; \ 724 \ 725 mbuf10 = pkts[pkt10_index]; \ 726 key10 = RTE_MBUF_METADATA_UINT8_PTR(mbuf10, key_offset); \ 727 sig10 = (uint64_t) f_hash(key10, t->key_mask, key_size, seed); \ 728 bkt10_index = sig10 & bucket_mask; \ 729 bkt10 = &buckets[bkt10_index]; \ 730 \ 731 mbuf11 = pkts[pkt11_index]; \ 732 key11 = RTE_MBUF_METADATA_UINT8_PTR(mbuf11, key_offset); \ 733 sig11 = (uint64_t) f_hash(key11, t->key_mask, key_size, seed); \ 734 bkt11_index = sig11 & bucket_mask; \ 735 bkt11 = &buckets[bkt11_index]; \ 736 \ 737 rte_prefetch0(bkt10); \ 738 rte_prefetch0(bkt11); \ 739 \ 740 g10 = &g[pkt10_index]; \ 741 g10->sig = sig10; \ 742 g10->bkt = bkt10; \ 743 \ 744 g11 = &g[pkt11_index]; \ 745 g11->sig = sig11; \ 746 g11->bkt = bkt11; \ 747 } 748 749 #define lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many)\ 750 { \ 751 struct grinder *g20, *g21; \ 752 uint64_t sig20, sig21; \ 753 struct bucket *bkt20, *bkt21; \ 754 uint8_t *key20, *key21, *key_mem = t->key_mem; \ 755 uint64_t match20, match21, match_many20, match_many21; \ 756 uint64_t match_pos20, match_pos21; \ 757 uint32_t key20_index, key21_index, key_size_shl = t->key_size_shl;\ 758 \ 759 g20 = &g[pkt20_index]; \ 760 sig20 = g20->sig; \ 761 bkt20 = g20->bkt; \ 762 sig20 = (sig20 >> 16) | 1LLU; \ 763 lookup_cmp_sig(sig20, bkt20, match20, match_many20, match_pos20);\ 764 match20 <<= pkt20_index; \ 765 match_many20 |= BUCKET_NEXT_VALID(bkt20); \ 766 match_many20 <<= pkt20_index; \ 767 key20_index = bkt20->key_pos[match_pos20]; \ 768 key20 = &key_mem[key20_index << key_size_shl]; \ 769 \ 770 g21 = &g[pkt21_index]; \ 771 sig21 = g21->sig; \ 772 bkt21 = g21->bkt; \ 773 sig21 = (sig21 >> 16) | 1LLU; \ 774 lookup_cmp_sig(sig21, bkt21, match21, match_many21, match_pos21);\ 775 match21 <<= pkt21_index; \ 776 match_many21 |= BUCKET_NEXT_VALID(bkt21); \ 777 match_many21 <<= pkt21_index; \ 778 key21_index = bkt21->key_pos[match_pos21]; \ 779 key21 = &key_mem[key21_index << key_size_shl]; \ 780 \ 781 rte_prefetch0(key20); \ 782 rte_prefetch0(key21); \ 783 \ 784 pkts_mask_match_many |= match_many20 | match_many21; \ 785 \ 786 g20->match = match20; \ 787 g20->key_index = key20_index; \ 788 \ 789 g21->match = match21; \ 790 g21->key_index = key21_index; \ 791 } 792 793 #define lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out, \ 794 entries) \ 795 { \ 796 struct grinder *g30, *g31; \ 797 struct rte_mbuf *mbuf30, *mbuf31; \ 798 uint8_t *key30, *key31, *key_mem = t->key_mem; \ 799 uint8_t *data30, *data31, *data_mem = t->data_mem; \ 800 uint64_t match30, match31, match_key30, match_key31, match_keys;\ 801 uint32_t key30_index, key31_index; \ 802 uint32_t key_size_shl = t->key_size_shl; \ 803 uint32_t data_size_shl = t->data_size_shl; \ 804 \ 805 mbuf30 = pkts[pkt30_index]; \ 806 g30 = &g[pkt30_index]; \ 807 match30 = g30->match; \ 808 key30_index = g30->key_index; \ 809 key30 = &key_mem[key30_index << key_size_shl]; \ 810 lookup_cmp_key(mbuf30, key30, match_key30, t); \ 811 match_key30 <<= pkt30_index; \ 812 match_key30 &= match30; \ 813 data30 = &data_mem[key30_index << data_size_shl]; \ 814 entries[pkt30_index] = data30; \ 815 \ 816 mbuf31 = pkts[pkt31_index]; \ 817 g31 = &g[pkt31_index]; \ 818 match31 = g31->match; \ 819 key31_index = g31->key_index; \ 820 key31 = &key_mem[key31_index << key_size_shl]; \ 821 lookup_cmp_key(mbuf31, key31, match_key31, t); \ 822 match_key31 <<= pkt31_index; \ 823 match_key31 &= match31; \ 824 data31 = &data_mem[key31_index << data_size_shl]; \ 825 entries[pkt31_index] = data31; \ 826 \ 827 rte_prefetch0(data30); \ 828 rte_prefetch0(data31); \ 829 \ 830 match_keys = match_key30 | match_key31; \ 831 pkts_mask_out |= match_keys; \ 832 } 833 834 /* 835 * The lookup function implements a 4-stage pipeline, with each stage processing 836 * two different packets. The purpose of pipelined implementation is to hide the 837 * latency of prefetching the data structures and loosen the data dependency 838 * between instructions. 839 * 840 * p00 _______ p10 _______ p20 _______ p30 _______ 841 *----->| |----->| |----->| |----->| |-----> 842 * | 0 | | 1 | | 2 | | 3 | 843 *----->|_______|----->|_______|----->|_______|----->|_______|-----> 844 * p01 p11 p21 p31 845 * 846 * The naming convention is: 847 * pXY = packet Y of stage X, X = 0 .. 3, Y = 0 .. 1 848 */ 849 static int rte_table_hash_ext_lookup( 850 void *table, 851 struct rte_mbuf **pkts, 852 uint64_t pkts_mask, 853 uint64_t *lookup_hit_mask, 854 void **entries) 855 { 856 struct rte_table_hash *t = (struct rte_table_hash *) table; 857 struct grinder *g = t->grinders; 858 uint64_t pkt00_index, pkt01_index, pkt10_index, pkt11_index; 859 uint64_t pkt20_index, pkt21_index, pkt30_index, pkt31_index; 860 uint64_t pkts_mask_out = 0, pkts_mask_match_many = 0; 861 int status = 0; 862 863 __rte_unused uint32_t n_pkts_in = rte_popcount64(pkts_mask); 864 RTE_TABLE_HASH_EXT_STATS_PKTS_IN_ADD(t, n_pkts_in); 865 866 /* Cannot run the pipeline with less than 7 packets */ 867 if (rte_popcount64(pkts_mask) < 7) { 868 status = rte_table_hash_ext_lookup_unoptimized(table, pkts, 869 pkts_mask, lookup_hit_mask, entries); 870 RTE_TABLE_HASH_EXT_STATS_PKTS_LOOKUP_MISS(t, n_pkts_in - 871 rte_popcount64(*lookup_hit_mask)); 872 return status; 873 } 874 875 /* Pipeline stage 0 */ 876 lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index); 877 878 /* Pipeline feed */ 879 pkt10_index = pkt00_index; 880 pkt11_index = pkt01_index; 881 882 /* Pipeline stage 0 */ 883 lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index); 884 885 /* Pipeline stage 1 */ 886 lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index); 887 888 /* Pipeline feed */ 889 pkt20_index = pkt10_index; 890 pkt21_index = pkt11_index; 891 pkt10_index = pkt00_index; 892 pkt11_index = pkt01_index; 893 894 /* Pipeline stage 0 */ 895 lookup2_stage0(t, g, pkts, pkts_mask, pkt00_index, pkt01_index); 896 897 /* Pipeline stage 1 */ 898 lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index); 899 900 /* Pipeline stage 2 */ 901 lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many); 902 903 /* 904 * Pipeline run 905 * 906 */ 907 for ( ; pkts_mask; ) { 908 /* Pipeline feed */ 909 pkt30_index = pkt20_index; 910 pkt31_index = pkt21_index; 911 pkt20_index = pkt10_index; 912 pkt21_index = pkt11_index; 913 pkt10_index = pkt00_index; 914 pkt11_index = pkt01_index; 915 916 /* Pipeline stage 0 */ 917 lookup2_stage0_with_odd_support(t, g, pkts, pkts_mask, 918 pkt00_index, pkt01_index); 919 920 /* Pipeline stage 1 */ 921 lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index); 922 923 /* Pipeline stage 2 */ 924 lookup2_stage2(t, g, pkt20_index, pkt21_index, 925 pkts_mask_match_many); 926 927 /* Pipeline stage 3 */ 928 lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, 929 pkts_mask_out, entries); 930 } 931 932 /* Pipeline feed */ 933 pkt30_index = pkt20_index; 934 pkt31_index = pkt21_index; 935 pkt20_index = pkt10_index; 936 pkt21_index = pkt11_index; 937 pkt10_index = pkt00_index; 938 pkt11_index = pkt01_index; 939 940 /* Pipeline stage 1 */ 941 lookup2_stage1(t, g, pkts, pkt10_index, pkt11_index); 942 943 /* Pipeline stage 2 */ 944 lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many); 945 946 /* Pipeline stage 3 */ 947 lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out, 948 entries); 949 950 /* Pipeline feed */ 951 pkt30_index = pkt20_index; 952 pkt31_index = pkt21_index; 953 pkt20_index = pkt10_index; 954 pkt21_index = pkt11_index; 955 956 /* Pipeline stage 2 */ 957 lookup2_stage2(t, g, pkt20_index, pkt21_index, pkts_mask_match_many); 958 959 /* Pipeline stage 3 */ 960 lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out, 961 entries); 962 963 /* Pipeline feed */ 964 pkt30_index = pkt20_index; 965 pkt31_index = pkt21_index; 966 967 /* Pipeline stage 3 */ 968 lookup2_stage3(t, g, pkts, pkt30_index, pkt31_index, pkts_mask_out, 969 entries); 970 971 /* Slow path */ 972 pkts_mask_match_many &= ~pkts_mask_out; 973 if (pkts_mask_match_many) { 974 uint64_t pkts_mask_out_slow = 0; 975 976 status = rte_table_hash_ext_lookup_unoptimized(table, pkts, 977 pkts_mask_match_many, &pkts_mask_out_slow, entries); 978 pkts_mask_out |= pkts_mask_out_slow; 979 } 980 981 *lookup_hit_mask = pkts_mask_out; 982 RTE_TABLE_HASH_EXT_STATS_PKTS_LOOKUP_MISS(t, n_pkts_in - rte_popcount64(pkts_mask_out)); 983 return status; 984 } 985 986 static int 987 rte_table_hash_ext_stats_read(void *table, struct rte_table_stats *stats, int clear) 988 { 989 struct rte_table_hash *t = table; 990 991 if (stats != NULL) 992 memcpy(stats, &t->stats, sizeof(t->stats)); 993 994 if (clear) 995 memset(&t->stats, 0, sizeof(t->stats)); 996 997 return 0; 998 } 999 1000 struct rte_table_ops rte_table_hash_ext_ops = { 1001 .f_create = rte_table_hash_ext_create, 1002 .f_free = rte_table_hash_ext_free, 1003 .f_add = rte_table_hash_ext_entry_add, 1004 .f_delete = rte_table_hash_ext_entry_delete, 1005 .f_add_bulk = NULL, 1006 .f_delete_bulk = NULL, 1007 .f_lookup = rte_table_hash_ext_lookup, 1008 .f_stats = rte_table_hash_ext_stats_read, 1009 }; 1010