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