1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2020 Inspur Corporation 3 */ 4 5 #include <rte_malloc.h> 6 #include <rte_mbuf.h> 7 #include <rte_ethdev.h> 8 #include <rte_udp.h> 9 10 #include "gro_vxlan_udp4.h" 11 12 void * 13 gro_vxlan_udp4_tbl_create(uint16_t socket_id, 14 uint16_t max_flow_num, 15 uint16_t max_item_per_flow) 16 { 17 struct gro_vxlan_udp4_tbl *tbl; 18 size_t size; 19 uint32_t entries_num, i; 20 21 entries_num = max_flow_num * max_item_per_flow; 22 entries_num = RTE_MIN(entries_num, GRO_VXLAN_UDP4_TBL_MAX_ITEM_NUM); 23 24 if (entries_num == 0) 25 return NULL; 26 27 tbl = rte_zmalloc_socket(__func__, 28 sizeof(struct gro_vxlan_udp4_tbl), 29 RTE_CACHE_LINE_SIZE, 30 socket_id); 31 if (tbl == NULL) 32 return NULL; 33 34 size = sizeof(struct gro_vxlan_udp4_item) * entries_num; 35 tbl->items = rte_zmalloc_socket(__func__, 36 size, 37 RTE_CACHE_LINE_SIZE, 38 socket_id); 39 if (tbl->items == NULL) { 40 rte_free(tbl); 41 return NULL; 42 } 43 tbl->max_item_num = entries_num; 44 45 size = sizeof(struct gro_vxlan_udp4_flow) * entries_num; 46 tbl->flows = rte_zmalloc_socket(__func__, 47 size, 48 RTE_CACHE_LINE_SIZE, 49 socket_id); 50 if (tbl->flows == NULL) { 51 rte_free(tbl->items); 52 rte_free(tbl); 53 return NULL; 54 } 55 56 for (i = 0; i < entries_num; i++) 57 tbl->flows[i].start_index = INVALID_ARRAY_INDEX; 58 tbl->max_flow_num = entries_num; 59 60 return tbl; 61 } 62 63 void 64 gro_vxlan_udp4_tbl_destroy(void *tbl) 65 { 66 struct gro_vxlan_udp4_tbl *vxlan_tbl = tbl; 67 68 if (vxlan_tbl) { 69 rte_free(vxlan_tbl->items); 70 rte_free(vxlan_tbl->flows); 71 } 72 rte_free(vxlan_tbl); 73 } 74 75 static inline uint32_t 76 find_an_empty_item(struct gro_vxlan_udp4_tbl *tbl) 77 { 78 uint32_t max_item_num = tbl->max_item_num, i; 79 80 for (i = 0; i < max_item_num; i++) 81 if (tbl->items[i].inner_item.firstseg == NULL) 82 return i; 83 return INVALID_ARRAY_INDEX; 84 } 85 86 static inline uint32_t 87 find_an_empty_flow(struct gro_vxlan_udp4_tbl *tbl) 88 { 89 uint32_t max_flow_num = tbl->max_flow_num, i; 90 91 for (i = 0; i < max_flow_num; i++) 92 if (tbl->flows[i].start_index == INVALID_ARRAY_INDEX) 93 return i; 94 return INVALID_ARRAY_INDEX; 95 } 96 97 static inline uint32_t 98 insert_new_item(struct gro_vxlan_udp4_tbl *tbl, 99 struct rte_mbuf *pkt, 100 uint64_t start_time, 101 uint32_t prev_idx, 102 uint16_t frag_offset, 103 uint8_t is_last_frag) 104 { 105 uint32_t item_idx; 106 107 item_idx = find_an_empty_item(tbl); 108 if (unlikely(item_idx == INVALID_ARRAY_INDEX)) 109 return INVALID_ARRAY_INDEX; 110 111 tbl->items[item_idx].inner_item.firstseg = pkt; 112 tbl->items[item_idx].inner_item.lastseg = rte_pktmbuf_lastseg(pkt); 113 tbl->items[item_idx].inner_item.start_time = start_time; 114 tbl->items[item_idx].inner_item.next_pkt_idx = INVALID_ARRAY_INDEX; 115 tbl->items[item_idx].inner_item.frag_offset = frag_offset; 116 tbl->items[item_idx].inner_item.is_last_frag = is_last_frag; 117 tbl->items[item_idx].inner_item.nb_merged = 1; 118 tbl->item_num++; 119 120 /* If the previous packet exists, chain the new one with it. */ 121 if (prev_idx != INVALID_ARRAY_INDEX) { 122 tbl->items[item_idx].inner_item.next_pkt_idx = 123 tbl->items[prev_idx].inner_item.next_pkt_idx; 124 tbl->items[prev_idx].inner_item.next_pkt_idx = item_idx; 125 } 126 127 return item_idx; 128 } 129 130 static inline uint32_t 131 delete_item(struct gro_vxlan_udp4_tbl *tbl, 132 uint32_t item_idx, 133 uint32_t prev_item_idx) 134 { 135 uint32_t next_idx = tbl->items[item_idx].inner_item.next_pkt_idx; 136 137 /* NULL indicates an empty item. */ 138 tbl->items[item_idx].inner_item.firstseg = NULL; 139 tbl->item_num--; 140 if (prev_item_idx != INVALID_ARRAY_INDEX) 141 tbl->items[prev_item_idx].inner_item.next_pkt_idx = next_idx; 142 143 return next_idx; 144 } 145 146 static inline uint32_t 147 insert_new_flow(struct gro_vxlan_udp4_tbl *tbl, 148 struct vxlan_udp4_flow_key *src, 149 uint32_t item_idx) 150 { 151 struct vxlan_udp4_flow_key *dst; 152 uint32_t flow_idx; 153 154 flow_idx = find_an_empty_flow(tbl); 155 if (unlikely(flow_idx == INVALID_ARRAY_INDEX)) 156 return INVALID_ARRAY_INDEX; 157 158 dst = &(tbl->flows[flow_idx].key); 159 160 rte_ether_addr_copy(&(src->inner_key.eth_saddr), 161 &(dst->inner_key.eth_saddr)); 162 rte_ether_addr_copy(&(src->inner_key.eth_daddr), 163 &(dst->inner_key.eth_daddr)); 164 dst->inner_key.ip_src_addr = src->inner_key.ip_src_addr; 165 dst->inner_key.ip_dst_addr = src->inner_key.ip_dst_addr; 166 dst->inner_key.ip_id = src->inner_key.ip_id; 167 168 dst->vxlan_hdr.vx_flags = src->vxlan_hdr.vx_flags; 169 dst->vxlan_hdr.vx_vni = src->vxlan_hdr.vx_vni; 170 rte_ether_addr_copy(&(src->outer_eth_saddr), &(dst->outer_eth_saddr)); 171 rte_ether_addr_copy(&(src->outer_eth_daddr), &(dst->outer_eth_daddr)); 172 dst->outer_ip_src_addr = src->outer_ip_src_addr; 173 dst->outer_ip_dst_addr = src->outer_ip_dst_addr; 174 dst->outer_dst_port = src->outer_dst_port; 175 176 tbl->flows[flow_idx].start_index = item_idx; 177 tbl->flow_num++; 178 179 return flow_idx; 180 } 181 182 static inline int 183 is_same_vxlan_udp4_flow(struct vxlan_udp4_flow_key k1, 184 struct vxlan_udp4_flow_key k2) 185 { 186 /* For VxLAN packet, outer udp src port is calculated from 187 * inner packet RSS hash, udp src port of the first UDP 188 * fragment is different from one of other UDP fragments 189 * even if they are same flow, so we have to skip outer udp 190 * src port comparison here. 191 */ 192 return (rte_is_same_ether_addr(&k1.outer_eth_saddr, 193 &k2.outer_eth_saddr) && 194 rte_is_same_ether_addr(&k1.outer_eth_daddr, 195 &k2.outer_eth_daddr) && 196 (k1.outer_ip_src_addr == k2.outer_ip_src_addr) && 197 (k1.outer_ip_dst_addr == k2.outer_ip_dst_addr) && 198 (k1.outer_dst_port == k2.outer_dst_port) && 199 (k1.vxlan_hdr.vx_flags == k2.vxlan_hdr.vx_flags) && 200 (k1.vxlan_hdr.vx_vni == k2.vxlan_hdr.vx_vni) && 201 is_same_udp4_flow(k1.inner_key, k2.inner_key)); 202 } 203 204 static inline int 205 udp4_check_vxlan_neighbor(struct gro_vxlan_udp4_item *item, 206 uint16_t frag_offset, 207 uint16_t ip_dl) 208 { 209 struct rte_mbuf *pkt = item->inner_item.firstseg; 210 int cmp; 211 uint16_t l2_offset; 212 int ret = 0; 213 214 /* Note: if outer DF bit is set, i.e outer_is_atomic is 0, 215 * we needn't compare outer_ip_id because they are same, 216 * for the case outer_is_atomic is 1, we also have no way 217 * to compare outer_ip_id because the difference between 218 * outer_ip_ids of two received packets isn't always +/-1. 219 * So skip outer_ip_id comparison here. 220 */ 221 222 l2_offset = pkt->outer_l2_len + pkt->outer_l3_len; 223 cmp = udp4_check_neighbor(&item->inner_item, frag_offset, ip_dl, 224 l2_offset); 225 if (cmp > 0) 226 /* Append the new packet. */ 227 ret = 1; 228 else if (cmp < 0) 229 /* Prepend the new packet. */ 230 ret = -1; 231 232 return ret; 233 } 234 235 static inline int 236 merge_two_vxlan_udp4_packets(struct gro_vxlan_udp4_item *item, 237 struct rte_mbuf *pkt, 238 int cmp, 239 uint16_t frag_offset, 240 uint8_t is_last_frag) 241 { 242 if (merge_two_udp4_packets(&item->inner_item, pkt, cmp, frag_offset, 243 is_last_frag, 244 pkt->outer_l2_len + pkt->outer_l3_len)) { 245 return 1; 246 } 247 248 return 0; 249 } 250 251 static inline void 252 update_vxlan_header(struct gro_vxlan_udp4_item *item) 253 { 254 struct rte_ipv4_hdr *ipv4_hdr; 255 struct rte_udp_hdr *udp_hdr; 256 struct rte_mbuf *pkt = item->inner_item.firstseg; 257 uint16_t len; 258 uint16_t frag_offset; 259 260 /* Update the outer IPv4 header. */ 261 len = pkt->pkt_len - pkt->outer_l2_len; 262 ipv4_hdr = rte_pktmbuf_mtod_offset(pkt, struct rte_ipv4_hdr *, 263 pkt->outer_l2_len); 264 ipv4_hdr->total_length = rte_cpu_to_be_16(len); 265 266 /* Update the outer UDP header. */ 267 len -= pkt->outer_l3_len; 268 udp_hdr = (struct rte_udp_hdr *)((char *)ipv4_hdr + pkt->outer_l3_len); 269 udp_hdr->dgram_len = rte_cpu_to_be_16(len); 270 271 /* Update the inner IPv4 header. */ 272 len -= pkt->l2_len; 273 ipv4_hdr = (struct rte_ipv4_hdr *)((char *)udp_hdr + pkt->l2_len); 274 ipv4_hdr->total_length = rte_cpu_to_be_16(len); 275 276 /* Clear MF bit if it is last fragment */ 277 if (item->inner_item.is_last_frag) { 278 frag_offset = rte_be_to_cpu_16(ipv4_hdr->fragment_offset); 279 ipv4_hdr->fragment_offset = 280 rte_cpu_to_be_16(frag_offset & ~RTE_IPV4_HDR_MF_FLAG); 281 } 282 } 283 284 int32_t 285 gro_vxlan_udp4_reassemble(struct rte_mbuf *pkt, 286 struct gro_vxlan_udp4_tbl *tbl, 287 uint64_t start_time) 288 { 289 struct rte_ether_hdr *outer_eth_hdr, *eth_hdr; 290 struct rte_ipv4_hdr *outer_ipv4_hdr, *ipv4_hdr; 291 struct rte_udp_hdr *udp_hdr; 292 struct rte_vxlan_hdr *vxlan_hdr; 293 uint16_t frag_offset; 294 uint8_t is_last_frag; 295 int16_t ip_dl; 296 uint16_t ip_id; 297 298 struct vxlan_udp4_flow_key key; 299 uint32_t cur_idx, prev_idx, item_idx; 300 uint32_t i, max_flow_num, remaining_flow_num; 301 int cmp; 302 uint16_t hdr_len; 303 uint8_t find; 304 305 outer_eth_hdr = rte_pktmbuf_mtod(pkt, struct rte_ether_hdr *); 306 outer_ipv4_hdr = (struct rte_ipv4_hdr *)((char *)outer_eth_hdr + 307 pkt->outer_l2_len); 308 309 udp_hdr = (struct rte_udp_hdr *)((char *)outer_ipv4_hdr + 310 pkt->outer_l3_len); 311 vxlan_hdr = (struct rte_vxlan_hdr *)((char *)udp_hdr + 312 sizeof(struct rte_udp_hdr)); 313 eth_hdr = (struct rte_ether_hdr *)((char *)vxlan_hdr + 314 sizeof(struct rte_vxlan_hdr)); 315 /* l2_len = outer udp hdr len + vxlan hdr len + inner l2 len */ 316 ipv4_hdr = (struct rte_ipv4_hdr *)((char *)udp_hdr + pkt->l2_len); 317 318 /* 319 * Don't process the packet which has non-fragment inner IP. 320 */ 321 if (!is_ipv4_fragment(ipv4_hdr)) 322 return -1; 323 324 hdr_len = pkt->outer_l2_len + pkt->outer_l3_len + pkt->l2_len + 325 pkt->l3_len; 326 /* 327 * Don't process the packet whose payload length is less than or 328 * equal to 0. 329 */ 330 if (pkt->pkt_len <= hdr_len) 331 return -1; 332 333 ip_dl = pkt->pkt_len - hdr_len; 334 335 ip_id = rte_be_to_cpu_16(ipv4_hdr->packet_id); 336 frag_offset = rte_be_to_cpu_16(ipv4_hdr->fragment_offset); 337 is_last_frag = ((frag_offset & RTE_IPV4_HDR_MF_FLAG) == 0) ? 1 : 0; 338 frag_offset = (uint16_t)(frag_offset & RTE_IPV4_HDR_OFFSET_MASK) << 3; 339 340 rte_ether_addr_copy(&(eth_hdr->src_addr), &(key.inner_key.eth_saddr)); 341 rte_ether_addr_copy(&(eth_hdr->dst_addr), &(key.inner_key.eth_daddr)); 342 key.inner_key.ip_src_addr = ipv4_hdr->src_addr; 343 key.inner_key.ip_dst_addr = ipv4_hdr->dst_addr; 344 key.inner_key.ip_id = ip_id; 345 346 key.vxlan_hdr.vx_flags = vxlan_hdr->vx_flags; 347 key.vxlan_hdr.vx_vni = vxlan_hdr->vx_vni; 348 rte_ether_addr_copy(&(outer_eth_hdr->src_addr), &(key.outer_eth_saddr)); 349 rte_ether_addr_copy(&(outer_eth_hdr->dst_addr), &(key.outer_eth_daddr)); 350 key.outer_ip_src_addr = outer_ipv4_hdr->src_addr; 351 key.outer_ip_dst_addr = outer_ipv4_hdr->dst_addr; 352 /* Note: It is unnecessary to save outer_src_port here because it can 353 * be different for VxLAN UDP fragments from the same flow. 354 */ 355 key.outer_dst_port = udp_hdr->dst_port; 356 357 /* Search for a matched flow. */ 358 max_flow_num = tbl->max_flow_num; 359 remaining_flow_num = tbl->flow_num; 360 find = 0; 361 for (i = 0; i < max_flow_num && remaining_flow_num; i++) { 362 if (tbl->flows[i].start_index != INVALID_ARRAY_INDEX) { 363 if (is_same_vxlan_udp4_flow(tbl->flows[i].key, key)) { 364 find = 1; 365 break; 366 } 367 remaining_flow_num--; 368 } 369 } 370 371 /* 372 * Can't find a matched flow. Insert a new flow and store the 373 * packet into the flow. 374 */ 375 if (find == 0) { 376 item_idx = insert_new_item(tbl, pkt, start_time, 377 INVALID_ARRAY_INDEX, frag_offset, 378 is_last_frag); 379 if (unlikely(item_idx == INVALID_ARRAY_INDEX)) 380 return -1; 381 if (insert_new_flow(tbl, &key, item_idx) == 382 INVALID_ARRAY_INDEX) { 383 /* 384 * Fail to insert a new flow, so 385 * delete the inserted packet. 386 */ 387 delete_item(tbl, item_idx, INVALID_ARRAY_INDEX); 388 return -1; 389 } 390 return 0; 391 } 392 393 /* Check all packets in the flow and try to find a neighbor. */ 394 cur_idx = tbl->flows[i].start_index; 395 prev_idx = cur_idx; 396 do { 397 cmp = udp4_check_vxlan_neighbor(&(tbl->items[cur_idx]), 398 frag_offset, ip_dl); 399 if (cmp) { 400 if (merge_two_vxlan_udp4_packets( 401 &(tbl->items[cur_idx]), 402 pkt, cmp, frag_offset, 403 is_last_frag)) { 404 return 1; 405 } 406 /* 407 * Can't merge two packets, as the packet 408 * length will be greater than the max value. 409 * Insert the packet into the flow. 410 */ 411 if (insert_new_item(tbl, pkt, start_time, prev_idx, 412 frag_offset, is_last_frag) == 413 INVALID_ARRAY_INDEX) 414 return -1; 415 return 0; 416 } 417 418 /* Ensure inserted items are ordered by frag_offset */ 419 if (frag_offset 420 < tbl->items[cur_idx].inner_item.frag_offset) { 421 break; 422 } 423 424 prev_idx = cur_idx; 425 cur_idx = tbl->items[cur_idx].inner_item.next_pkt_idx; 426 } while (cur_idx != INVALID_ARRAY_INDEX); 427 428 /* Can't find neighbor. Insert the packet into the flow. */ 429 if (cur_idx == tbl->flows[i].start_index) { 430 /* Insert it before the first packet of the flow */ 431 item_idx = insert_new_item(tbl, pkt, start_time, 432 INVALID_ARRAY_INDEX, frag_offset, 433 is_last_frag); 434 if (unlikely(item_idx == INVALID_ARRAY_INDEX)) 435 return -1; 436 tbl->items[item_idx].inner_item.next_pkt_idx = cur_idx; 437 tbl->flows[i].start_index = item_idx; 438 } else { 439 if (insert_new_item(tbl, pkt, start_time, prev_idx, 440 frag_offset, is_last_frag 441 ) == INVALID_ARRAY_INDEX) 442 return -1; 443 } 444 445 return 0; 446 } 447 448 static int 449 gro_vxlan_udp4_merge_items(struct gro_vxlan_udp4_tbl *tbl, 450 uint32_t start_idx) 451 { 452 uint16_t frag_offset; 453 uint8_t is_last_frag; 454 int16_t ip_dl; 455 struct rte_mbuf *pkt; 456 int cmp; 457 uint32_t item_idx; 458 uint16_t hdr_len; 459 460 item_idx = tbl->items[start_idx].inner_item.next_pkt_idx; 461 while (item_idx != INVALID_ARRAY_INDEX) { 462 pkt = tbl->items[item_idx].inner_item.firstseg; 463 hdr_len = pkt->outer_l2_len + pkt->outer_l3_len + pkt->l2_len + 464 pkt->l3_len; 465 ip_dl = pkt->pkt_len - hdr_len; 466 frag_offset = tbl->items[item_idx].inner_item.frag_offset; 467 is_last_frag = tbl->items[item_idx].inner_item.is_last_frag; 468 cmp = udp4_check_vxlan_neighbor(&(tbl->items[start_idx]), 469 frag_offset, ip_dl); 470 if (cmp) { 471 if (merge_two_vxlan_udp4_packets( 472 &(tbl->items[start_idx]), 473 pkt, cmp, frag_offset, 474 is_last_frag)) { 475 item_idx = delete_item(tbl, item_idx, 476 INVALID_ARRAY_INDEX); 477 tbl->items[start_idx].inner_item.next_pkt_idx 478 = item_idx; 479 } else 480 return 0; 481 } else 482 return 0; 483 } 484 485 return 0; 486 } 487 488 uint16_t 489 gro_vxlan_udp4_tbl_timeout_flush(struct gro_vxlan_udp4_tbl *tbl, 490 uint64_t flush_timestamp, 491 struct rte_mbuf **out, 492 uint16_t nb_out) 493 { 494 uint16_t k = 0; 495 uint32_t i, j; 496 uint32_t max_flow_num = tbl->max_flow_num; 497 498 for (i = 0; i < max_flow_num; i++) { 499 if (unlikely(tbl->flow_num == 0)) 500 return k; 501 502 j = tbl->flows[i].start_index; 503 while (j != INVALID_ARRAY_INDEX) { 504 if (tbl->items[j].inner_item.start_time <= 505 flush_timestamp) { 506 gro_vxlan_udp4_merge_items(tbl, j); 507 out[k++] = tbl->items[j].inner_item.firstseg; 508 if (tbl->items[j].inner_item.nb_merged > 1) 509 update_vxlan_header(&(tbl->items[j])); 510 /* 511 * Delete the item and get the next packet 512 * index. 513 */ 514 j = delete_item(tbl, j, INVALID_ARRAY_INDEX); 515 tbl->flows[i].start_index = j; 516 if (j == INVALID_ARRAY_INDEX) 517 tbl->flow_num--; 518 519 if (unlikely(k == nb_out)) 520 return k; 521 } else 522 /* 523 * Flushing packets does not strictly follow 524 * timestamp. It does not flush left packets of 525 * the flow this time once it finds one item 526 * whose start_time is greater than 527 * flush_timestamp. So go to check other flows. 528 */ 529 break; 530 } 531 } 532 return k; 533 } 534 535 uint32_t 536 gro_vxlan_udp4_tbl_pkt_count(void *tbl) 537 { 538 struct gro_vxlan_udp4_tbl *gro_tbl = tbl; 539 540 if (gro_tbl) 541 return gro_tbl->item_num; 542 543 return 0; 544 } 545