1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2007, Myricom Inc. 5 * Copyright (c) 2008, Intel Corporation. 6 * Copyright (c) 2012 The FreeBSD Foundation 7 * Copyright (c) 2016-2021 Mellanox Technologies. 8 * All rights reserved. 9 * 10 * Portions of this software were developed by Bjoern Zeeb 11 * under sponsorship from the FreeBSD Foundation. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include "opt_inet.h" 39 #include "opt_inet6.h" 40 41 #include <sys/param.h> 42 #include <sys/systm.h> 43 #include <sys/kernel.h> 44 #include <sys/malloc.h> 45 #include <sys/mbuf.h> 46 #include <sys/socket.h> 47 #include <sys/socketvar.h> 48 #include <sys/sockbuf.h> 49 #include <sys/sysctl.h> 50 51 #include <net/if.h> 52 #include <net/if_var.h> 53 #include <net/ethernet.h> 54 #include <net/bpf.h> 55 #include <net/vnet.h> 56 57 #include <netinet/in_systm.h> 58 #include <netinet/in.h> 59 #include <netinet/ip6.h> 60 #include <netinet/ip.h> 61 #include <netinet/ip_var.h> 62 #include <netinet/in_pcb.h> 63 #include <netinet6/in6_pcb.h> 64 #include <netinet/tcp.h> 65 #include <netinet/tcp_seq.h> 66 #include <netinet/tcp_lro.h> 67 #include <netinet/tcp_var.h> 68 #include <netinet/tcpip.h> 69 #include <netinet/tcp_hpts.h> 70 #include <netinet/tcp_log_buf.h> 71 #include <netinet/udp.h> 72 #include <netinet6/ip6_var.h> 73 74 #include <machine/in_cksum.h> 75 76 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures"); 77 78 #define TCP_LRO_TS_OPTION \ 79 ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \ 80 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP) 81 82 static void tcp_lro_rx_done(struct lro_ctrl *lc); 83 static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, 84 uint32_t csum, bool use_hash); 85 86 #ifdef TCPHPTS 87 static bool do_bpf_strip_and_compress(struct inpcb *, struct lro_ctrl *, 88 struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **, bool *, bool); 89 90 #endif 91 92 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 93 "TCP LRO"); 94 95 static long tcplro_stacks_wanting_mbufq; 96 counter_u64_t tcp_inp_lro_direct_queue; 97 counter_u64_t tcp_inp_lro_wokeup_queue; 98 counter_u64_t tcp_inp_lro_compressed; 99 counter_u64_t tcp_inp_lro_locks_taken; 100 counter_u64_t tcp_extra_mbuf; 101 counter_u64_t tcp_would_have_but; 102 counter_u64_t tcp_comp_total; 103 counter_u64_t tcp_uncomp_total; 104 counter_u64_t tcp_bad_csums; 105 106 static unsigned tcp_lro_entries = TCP_LRO_ENTRIES; 107 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries, 108 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0, 109 "default number of LRO entries"); 110 111 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH; 112 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold, 113 CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0, 114 "Number of interrups in a row on the same CPU that will make us declare an 'affinity' cpu?"); 115 116 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD, 117 &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport"); 118 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD, 119 &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts"); 120 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD, 121 &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport"); 122 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD, 123 &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken"); 124 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD, 125 &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp"); 126 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD, 127 &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed"); 128 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD, 129 &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set"); 130 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD, 131 &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP"); 132 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD, 133 &tcp_bad_csums, "Number of packets that the common code saw with bad csums"); 134 135 void 136 tcp_lro_reg_mbufq(void) 137 { 138 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1); 139 } 140 141 void 142 tcp_lro_dereg_mbufq(void) 143 { 144 atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1); 145 } 146 147 static __inline void 148 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket, 149 struct lro_entry *le) 150 { 151 152 LIST_INSERT_HEAD(&lc->lro_active, le, next); 153 LIST_INSERT_HEAD(bucket, le, hash_next); 154 } 155 156 static __inline void 157 tcp_lro_active_remove(struct lro_entry *le) 158 { 159 160 LIST_REMOVE(le, next); /* active list */ 161 LIST_REMOVE(le, hash_next); /* hash bucket */ 162 } 163 164 int 165 tcp_lro_init(struct lro_ctrl *lc) 166 { 167 return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0)); 168 } 169 170 int 171 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp, 172 unsigned lro_entries, unsigned lro_mbufs) 173 { 174 struct lro_entry *le; 175 size_t size; 176 unsigned i, elements; 177 178 lc->lro_bad_csum = 0; 179 lc->lro_queued = 0; 180 lc->lro_flushed = 0; 181 lc->lro_mbuf_count = 0; 182 lc->lro_mbuf_max = lro_mbufs; 183 lc->lro_cnt = lro_entries; 184 lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX; 185 lc->lro_length_lim = TCP_LRO_LENGTH_MAX; 186 lc->ifp = ifp; 187 LIST_INIT(&lc->lro_free); 188 LIST_INIT(&lc->lro_active); 189 190 /* create hash table to accelerate entry lookup */ 191 if (lro_entries > lro_mbufs) 192 elements = lro_entries; 193 else 194 elements = lro_mbufs; 195 lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz, 196 HASH_NOWAIT); 197 if (lc->lro_hash == NULL) { 198 memset(lc, 0, sizeof(*lc)); 199 return (ENOMEM); 200 } 201 202 /* compute size to allocate */ 203 size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) + 204 (lro_entries * sizeof(*le)); 205 lc->lro_mbuf_data = (struct lro_mbuf_sort *) 206 malloc(size, M_LRO, M_NOWAIT | M_ZERO); 207 208 /* check for out of memory */ 209 if (lc->lro_mbuf_data == NULL) { 210 free(lc->lro_hash, M_LRO); 211 memset(lc, 0, sizeof(*lc)); 212 return (ENOMEM); 213 } 214 /* compute offset for LRO entries */ 215 le = (struct lro_entry *) 216 (lc->lro_mbuf_data + lro_mbufs); 217 218 /* setup linked list */ 219 for (i = 0; i != lro_entries; i++) 220 LIST_INSERT_HEAD(&lc->lro_free, le + i, next); 221 222 return (0); 223 } 224 225 struct vxlan_header { 226 uint32_t vxlh_flags; 227 uint32_t vxlh_vni; 228 }; 229 230 static inline void * 231 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen) 232 { 233 const struct ether_vlan_header *eh; 234 void *old; 235 uint16_t eth_type; 236 237 if (update_data) 238 memset(parser, 0, sizeof(*parser)); 239 240 old = ptr; 241 242 if (is_vxlan) { 243 const struct vxlan_header *vxh; 244 vxh = ptr; 245 ptr = (uint8_t *)ptr + sizeof(*vxh); 246 if (update_data) { 247 parser->data.vxlan_vni = 248 vxh->vxlh_vni & htonl(0xffffff00); 249 } 250 } 251 252 eh = ptr; 253 if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) { 254 eth_type = eh->evl_proto; 255 if (update_data) { 256 /* strip priority and keep VLAN ID only */ 257 parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK); 258 } 259 /* advance to next header */ 260 ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; 261 mlen -= (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN); 262 } else { 263 eth_type = eh->evl_encap_proto; 264 /* advance to next header */ 265 mlen -= ETHER_HDR_LEN; 266 ptr = (uint8_t *)ptr + ETHER_HDR_LEN; 267 } 268 if (__predict_false(mlen <= 0)) 269 return (NULL); 270 switch (eth_type) { 271 #ifdef INET 272 case htons(ETHERTYPE_IP): 273 parser->ip4 = ptr; 274 if (__predict_false(mlen < sizeof(struct ip))) 275 return (NULL); 276 /* Ensure there are no IPv4 options. */ 277 if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4)) 278 break; 279 /* .. and the packet is not fragmented. */ 280 if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK)) 281 break; 282 ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2); 283 mlen -= sizeof(struct ip); 284 if (update_data) { 285 parser->data.s_addr.v4 = parser->ip4->ip_src; 286 parser->data.d_addr.v4 = parser->ip4->ip_dst; 287 } 288 switch (parser->ip4->ip_p) { 289 case IPPROTO_UDP: 290 if (__predict_false(mlen < sizeof(struct udphdr))) 291 return (NULL); 292 parser->udp = ptr; 293 if (update_data) { 294 parser->data.lro_type = LRO_TYPE_IPV4_UDP; 295 parser->data.s_port = parser->udp->uh_sport; 296 parser->data.d_port = parser->udp->uh_dport; 297 } else { 298 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP); 299 } 300 ptr = ((uint8_t *)ptr + sizeof(*parser->udp)); 301 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; 302 return (ptr); 303 case IPPROTO_TCP: 304 parser->tcp = ptr; 305 if (__predict_false(mlen < sizeof(struct tcphdr))) 306 return (NULL); 307 if (update_data) { 308 parser->data.lro_type = LRO_TYPE_IPV4_TCP; 309 parser->data.s_port = parser->tcp->th_sport; 310 parser->data.d_port = parser->tcp->th_dport; 311 } else { 312 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP); 313 } 314 if (__predict_false(mlen < (parser->tcp->th_off << 2))) 315 return (NULL); 316 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2); 317 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; 318 return (ptr); 319 default: 320 break; 321 } 322 break; 323 #endif 324 #ifdef INET6 325 case htons(ETHERTYPE_IPV6): 326 parser->ip6 = ptr; 327 if (__predict_false(mlen < sizeof(struct ip6_hdr))) 328 return (NULL); 329 ptr = (uint8_t *)ptr + sizeof(*parser->ip6); 330 if (update_data) { 331 parser->data.s_addr.v6 = parser->ip6->ip6_src; 332 parser->data.d_addr.v6 = parser->ip6->ip6_dst; 333 } 334 mlen -= sizeof(struct ip6_hdr); 335 switch (parser->ip6->ip6_nxt) { 336 case IPPROTO_UDP: 337 if (__predict_false(mlen < sizeof(struct udphdr))) 338 return (NULL); 339 parser->udp = ptr; 340 if (update_data) { 341 parser->data.lro_type = LRO_TYPE_IPV6_UDP; 342 parser->data.s_port = parser->udp->uh_sport; 343 parser->data.d_port = parser->udp->uh_dport; 344 } else { 345 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP); 346 } 347 ptr = (uint8_t *)ptr + sizeof(*parser->udp); 348 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; 349 return (ptr); 350 case IPPROTO_TCP: 351 if (__predict_false(mlen < sizeof(struct tcphdr))) 352 return (NULL); 353 parser->tcp = ptr; 354 if (update_data) { 355 parser->data.lro_type = LRO_TYPE_IPV6_TCP; 356 parser->data.s_port = parser->tcp->th_sport; 357 parser->data.d_port = parser->tcp->th_dport; 358 } else { 359 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP); 360 } 361 if (__predict_false(mlen < (parser->tcp->th_off << 2))) 362 return (NULL); 363 ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2); 364 parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; 365 return (ptr); 366 default: 367 break; 368 } 369 break; 370 #endif 371 default: 372 break; 373 } 374 /* Invalid packet - cannot parse */ 375 return (NULL); 376 } 377 378 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID | 379 CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID; 380 381 static inline struct lro_parser * 382 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data) 383 { 384 void *data_ptr; 385 386 /* Try to parse outer headers first. */ 387 data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len); 388 if (data_ptr == NULL || po->total_hdr_len > m->m_len) 389 return (NULL); 390 391 if (update_data) { 392 /* Store VLAN ID, if any. */ 393 if (__predict_false(m->m_flags & M_VLANTAG)) { 394 po->data.vlan_id = 395 htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK); 396 } 397 /* Store decrypted flag, if any. */ 398 if (__predict_false((m->m_pkthdr.csum_flags & 399 CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED)) 400 po->data.lro_flags |= LRO_FLAG_DECRYPTED; 401 } 402 403 switch (po->data.lro_type) { 404 case LRO_TYPE_IPV4_UDP: 405 case LRO_TYPE_IPV6_UDP: 406 /* Check for VXLAN headers. */ 407 if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum) 408 break; 409 410 /* Try to parse inner headers. */ 411 data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true, 412 (m->m_len - ((caddr_t)data_ptr - m->m_data))); 413 if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len) 414 break; 415 416 /* Verify supported header types. */ 417 switch (pi->data.lro_type) { 418 case LRO_TYPE_IPV4_TCP: 419 case LRO_TYPE_IPV6_TCP: 420 return (pi); 421 default: 422 break; 423 } 424 break; 425 case LRO_TYPE_IPV4_TCP: 426 case LRO_TYPE_IPV6_TCP: 427 if (update_data) 428 memset(pi, 0, sizeof(*pi)); 429 return (po); 430 default: 431 break; 432 } 433 return (NULL); 434 } 435 436 static inline int 437 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po) 438 { 439 int len; 440 441 switch (po->data.lro_type) { 442 #ifdef INET 443 case LRO_TYPE_IPV4_TCP: 444 len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) + 445 ntohs(po->ip4->ip_len); 446 break; 447 #endif 448 #ifdef INET6 449 case LRO_TYPE_IPV6_TCP: 450 len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) + 451 ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6); 452 break; 453 #endif 454 default: 455 return (TCP_LRO_CANNOT); 456 } 457 458 /* 459 * If the frame is padded beyond the end of the IP packet, 460 * then trim the extra bytes off: 461 */ 462 if (__predict_true(m->m_pkthdr.len == len)) { 463 return (0); 464 } else if (m->m_pkthdr.len > len) { 465 m_adj(m, len - m->m_pkthdr.len); 466 return (0); 467 } 468 return (TCP_LRO_CANNOT); 469 } 470 471 static struct tcphdr * 472 tcp_lro_get_th(struct mbuf *m) 473 { 474 return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off)); 475 } 476 477 static void 478 lro_free_mbuf_chain(struct mbuf *m) 479 { 480 struct mbuf *save; 481 482 while (m) { 483 save = m->m_nextpkt; 484 m->m_nextpkt = NULL; 485 m_freem(m); 486 m = save; 487 } 488 } 489 490 void 491 tcp_lro_free(struct lro_ctrl *lc) 492 { 493 struct lro_entry *le; 494 unsigned x; 495 496 /* reset LRO free list */ 497 LIST_INIT(&lc->lro_free); 498 499 /* free active mbufs, if any */ 500 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) { 501 tcp_lro_active_remove(le); 502 lro_free_mbuf_chain(le->m_head); 503 } 504 505 /* free hash table */ 506 free(lc->lro_hash, M_LRO); 507 lc->lro_hash = NULL; 508 lc->lro_hashsz = 0; 509 510 /* free mbuf array, if any */ 511 for (x = 0; x != lc->lro_mbuf_count; x++) 512 m_freem(lc->lro_mbuf_data[x].mb); 513 lc->lro_mbuf_count = 0; 514 515 /* free allocated memory, if any */ 516 free(lc->lro_mbuf_data, M_LRO); 517 lc->lro_mbuf_data = NULL; 518 } 519 520 static uint16_t 521 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th) 522 { 523 const uint16_t *ptr; 524 uint32_t csum; 525 uint16_t len; 526 527 csum = -th->th_sum; /* exclude checksum field */ 528 len = th->th_off; 529 ptr = (const uint16_t *)th; 530 while (len--) { 531 csum += *ptr; 532 ptr++; 533 csum += *ptr; 534 ptr++; 535 } 536 while (csum > 0xffff) 537 csum = (csum >> 16) + (csum & 0xffff); 538 539 return (csum); 540 } 541 542 static uint16_t 543 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum) 544 { 545 uint32_t c; 546 uint16_t cs; 547 548 c = tcp_csum; 549 550 switch (pa->data.lro_type) { 551 #ifdef INET6 552 case LRO_TYPE_IPV6_TCP: 553 /* Compute full pseudo IPv6 header checksum. */ 554 cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0); 555 break; 556 #endif 557 #ifdef INET 558 case LRO_TYPE_IPV4_TCP: 559 /* Compute full pseudo IPv4 header checsum. */ 560 cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP); 561 cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs)); 562 break; 563 #endif 564 default: 565 cs = 0; /* Keep compiler happy. */ 566 break; 567 } 568 569 /* Complement checksum. */ 570 cs = ~cs; 571 c += cs; 572 573 /* Remove TCP header checksum. */ 574 cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp); 575 c += cs; 576 577 /* Compute checksum remainder. */ 578 while (c > 0xffff) 579 c = (c >> 16) + (c & 0xffff); 580 581 return (c); 582 } 583 584 static void 585 tcp_lro_rx_done(struct lro_ctrl *lc) 586 { 587 struct lro_entry *le; 588 589 while ((le = LIST_FIRST(&lc->lro_active)) != NULL) { 590 tcp_lro_active_remove(le); 591 tcp_lro_flush(lc, le); 592 } 593 } 594 595 void 596 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout) 597 { 598 struct lro_entry *le, *le_tmp; 599 uint64_t now, tov; 600 struct bintime bt; 601 602 NET_EPOCH_ASSERT(); 603 if (LIST_EMPTY(&lc->lro_active)) 604 return; 605 606 /* get timeout time and current time in ns */ 607 binuptime(&bt); 608 now = bintime2ns(&bt); 609 tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000)); 610 LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) { 611 if (now >= (bintime2ns(&le->alloc_time) + tov)) { 612 tcp_lro_active_remove(le); 613 tcp_lro_flush(lc, le); 614 } 615 } 616 } 617 618 #ifdef INET 619 static int 620 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4) 621 { 622 uint16_t csum; 623 624 /* Legacy IP has a header checksum that needs to be correct. */ 625 if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) { 626 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) { 627 lc->lro_bad_csum++; 628 return (TCP_LRO_CANNOT); 629 } 630 } else { 631 csum = in_cksum_hdr(ip4); 632 if (__predict_false(csum != 0)) { 633 lc->lro_bad_csum++; 634 return (TCP_LRO_CANNOT); 635 } 636 } 637 return (0); 638 } 639 #endif 640 641 #ifdef TCPHPTS 642 static void 643 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc, 644 const struct lro_entry *le, const struct mbuf *m, 645 int frm, int32_t tcp_data_len, uint32_t th_seq, 646 uint32_t th_ack, uint16_t th_win) 647 { 648 if (tp->t_logstate != TCP_LOG_STATE_OFF) { 649 union tcp_log_stackspecific log; 650 struct timeval tv, btv; 651 uint32_t cts; 652 653 cts = tcp_get_usecs(&tv); 654 memset(&log, 0, sizeof(union tcp_log_stackspecific)); 655 log.u_bbr.flex8 = frm; 656 log.u_bbr.flex1 = tcp_data_len; 657 if (m) 658 log.u_bbr.flex2 = m->m_pkthdr.len; 659 else 660 log.u_bbr.flex2 = 0; 661 log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs; 662 log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len; 663 if (le->m_head) { 664 log.u_bbr.flex5 = le->m_head->m_pkthdr.len; 665 log.u_bbr.delRate = le->m_head->m_flags; 666 log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp; 667 } 668 log.u_bbr.inflight = th_seq; 669 log.u_bbr.delivered = th_ack; 670 log.u_bbr.timeStamp = cts; 671 log.u_bbr.epoch = le->next_seq; 672 log.u_bbr.lt_epoch = le->ack_seq; 673 log.u_bbr.pacing_gain = th_win; 674 log.u_bbr.cwnd_gain = le->window; 675 log.u_bbr.lost = curcpu; 676 log.u_bbr.cur_del_rate = (uintptr_t)m; 677 log.u_bbr.bw_inuse = (uintptr_t)le->m_head; 678 bintime2timeval(&lc->lro_last_queue_time, &btv); 679 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv); 680 log.u_bbr.flex7 = le->compressed; 681 log.u_bbr.pacing_gain = le->uncompressed; 682 if (in_epoch(net_epoch_preempt)) 683 log.u_bbr.inhpts = 1; 684 else 685 log.u_bbr.inhpts = 0; 686 TCP_LOG_EVENTP(tp, NULL, 687 &tp->t_inpcb->inp_socket->so_rcv, 688 &tp->t_inpcb->inp_socket->so_snd, 689 TCP_LOG_LRO, 0, 690 0, &log, false, &tv); 691 } 692 } 693 #endif 694 695 static inline void 696 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum) 697 { 698 uint32_t csum; 699 700 csum = 0xffff - *ptr + value; 701 while (csum > 0xffff) 702 csum = (csum >> 16) + (csum & 0xffff); 703 *ptr = value; 704 *psum = csum; 705 } 706 707 static uint16_t 708 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le, 709 uint16_t payload_len, uint16_t delta_sum) 710 { 711 uint32_t csum; 712 uint16_t tlen; 713 uint16_t temp[5] = {}; 714 715 switch (pa->data.lro_type) { 716 case LRO_TYPE_IPV4_TCP: 717 /* Compute new IPv4 length. */ 718 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len; 719 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]); 720 721 /* Subtract delta from current IPv4 checksum. */ 722 csum = pa->ip4->ip_sum + 0xffff - temp[0]; 723 while (csum > 0xffff) 724 csum = (csum >> 16) + (csum & 0xffff); 725 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]); 726 goto update_tcp_header; 727 728 case LRO_TYPE_IPV6_TCP: 729 /* Compute new IPv6 length. */ 730 tlen = (pa->tcp->th_off << 2) + payload_len; 731 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]); 732 goto update_tcp_header; 733 734 case LRO_TYPE_IPV4_UDP: 735 /* Compute new IPv4 length. */ 736 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len; 737 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]); 738 739 /* Subtract delta from current IPv4 checksum. */ 740 csum = pa->ip4->ip_sum + 0xffff - temp[0]; 741 while (csum > 0xffff) 742 csum = (csum >> 16) + (csum & 0xffff); 743 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]); 744 goto update_udp_header; 745 746 case LRO_TYPE_IPV6_UDP: 747 /* Compute new IPv6 length. */ 748 tlen = sizeof(*pa->udp) + payload_len; 749 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]); 750 goto update_udp_header; 751 752 default: 753 return (0); 754 } 755 756 update_tcp_header: 757 /* Compute current TCP header checksum. */ 758 temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp); 759 760 /* Incorporate the latest ACK into the TCP header. */ 761 pa->tcp->th_ack = le->ack_seq; 762 pa->tcp->th_win = le->window; 763 764 /* Incorporate latest timestamp into the TCP header. */ 765 if (le->timestamp != 0) { 766 uint32_t *ts_ptr; 767 768 ts_ptr = (uint32_t *)(pa->tcp + 1); 769 ts_ptr[1] = htonl(le->tsval); 770 ts_ptr[2] = le->tsecr; 771 } 772 773 /* Compute new TCP header checksum. */ 774 temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp); 775 776 /* Compute new TCP checksum. */ 777 csum = pa->tcp->th_sum + 0xffff - delta_sum + 778 0xffff - temp[0] + 0xffff - temp[3] + temp[2]; 779 while (csum > 0xffff) 780 csum = (csum >> 16) + (csum & 0xffff); 781 782 /* Assign new TCP checksum. */ 783 tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]); 784 785 /* Compute all modififications affecting next checksum. */ 786 csum = temp[0] + temp[1] + 0xffff - temp[2] + 787 temp[3] + temp[4] + delta_sum; 788 while (csum > 0xffff) 789 csum = (csum >> 16) + (csum & 0xffff); 790 791 /* Return delta checksum to next stage, if any. */ 792 return (csum); 793 794 update_udp_header: 795 tlen = sizeof(*pa->udp) + payload_len; 796 /* Assign new UDP length and compute checksum delta. */ 797 tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]); 798 799 /* Check if there is a UDP checksum. */ 800 if (__predict_false(pa->udp->uh_sum != 0)) { 801 /* Compute new UDP checksum. */ 802 csum = pa->udp->uh_sum + 0xffff - delta_sum + 803 0xffff - temp[0] + 0xffff - temp[2]; 804 while (csum > 0xffff) 805 csum = (csum >> 16) + (csum & 0xffff); 806 /* Assign new UDP checksum. */ 807 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]); 808 } 809 810 /* Compute all modififications affecting next checksum. */ 811 csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum; 812 while (csum > 0xffff) 813 csum = (csum >> 16) + (csum & 0xffff); 814 815 /* Return delta checksum to next stage, if any. */ 816 return (csum); 817 } 818 819 static void 820 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le) 821 { 822 /* Check if we need to recompute any checksums. */ 823 if (le->m_head->m_pkthdr.lro_nsegs > 1) { 824 uint16_t csum; 825 826 switch (le->inner.data.lro_type) { 827 case LRO_TYPE_IPV4_TCP: 828 csum = tcp_lro_update_checksum(&le->inner, le, 829 le->m_head->m_pkthdr.lro_tcp_d_len, 830 le->m_head->m_pkthdr.lro_tcp_d_csum); 831 csum = tcp_lro_update_checksum(&le->outer, NULL, 832 le->m_head->m_pkthdr.lro_tcp_d_len + 833 le->inner.total_hdr_len, csum); 834 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | 835 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID; 836 le->m_head->m_pkthdr.csum_data = 0xffff; 837 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED)) 838 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED; 839 break; 840 case LRO_TYPE_IPV6_TCP: 841 csum = tcp_lro_update_checksum(&le->inner, le, 842 le->m_head->m_pkthdr.lro_tcp_d_len, 843 le->m_head->m_pkthdr.lro_tcp_d_csum); 844 csum = tcp_lro_update_checksum(&le->outer, NULL, 845 le->m_head->m_pkthdr.lro_tcp_d_len + 846 le->inner.total_hdr_len, csum); 847 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | 848 CSUM_PSEUDO_HDR; 849 le->m_head->m_pkthdr.csum_data = 0xffff; 850 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED)) 851 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED; 852 break; 853 case LRO_TYPE_NONE: 854 switch (le->outer.data.lro_type) { 855 case LRO_TYPE_IPV4_TCP: 856 csum = tcp_lro_update_checksum(&le->outer, le, 857 le->m_head->m_pkthdr.lro_tcp_d_len, 858 le->m_head->m_pkthdr.lro_tcp_d_csum); 859 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | 860 CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID; 861 le->m_head->m_pkthdr.csum_data = 0xffff; 862 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED)) 863 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED; 864 break; 865 case LRO_TYPE_IPV6_TCP: 866 csum = tcp_lro_update_checksum(&le->outer, le, 867 le->m_head->m_pkthdr.lro_tcp_d_len, 868 le->m_head->m_pkthdr.lro_tcp_d_csum); 869 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | 870 CSUM_PSEUDO_HDR; 871 le->m_head->m_pkthdr.csum_data = 0xffff; 872 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED)) 873 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED; 874 break; 875 default: 876 break; 877 } 878 break; 879 default: 880 break; 881 } 882 } 883 884 /* 885 * Break any chain, this is not set to NULL on the singleton 886 * case m_nextpkt points to m_head. Other case set them 887 * m_nextpkt to NULL in push_and_replace. 888 */ 889 le->m_head->m_nextpkt = NULL; 890 lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs; 891 (*lc->ifp->if_input)(lc->ifp, le->m_head); 892 } 893 894 static void 895 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le, 896 struct mbuf *m, struct tcphdr *th) 897 { 898 uint32_t *ts_ptr; 899 uint16_t tcp_data_len; 900 uint16_t tcp_opt_len; 901 902 ts_ptr = (uint32_t *)(th + 1); 903 tcp_opt_len = (th->th_off << 2); 904 tcp_opt_len -= sizeof(*th); 905 906 /* Check if there is a timestamp option. */ 907 if (tcp_opt_len == 0 || 908 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA || 909 *ts_ptr != TCP_LRO_TS_OPTION)) { 910 /* We failed to find the timestamp option. */ 911 le->timestamp = 0; 912 } else { 913 le->timestamp = 1; 914 le->tsval = ntohl(*(ts_ptr + 1)); 915 le->tsecr = *(ts_ptr + 2); 916 } 917 918 tcp_data_len = m->m_pkthdr.lro_tcp_d_len; 919 920 /* Pull out TCP sequence numbers and window size. */ 921 le->next_seq = ntohl(th->th_seq) + tcp_data_len; 922 le->ack_seq = th->th_ack; 923 le->window = th->th_win; 924 925 /* Setup new data pointers. */ 926 le->m_head = m; 927 le->m_tail = m_last(m); 928 } 929 930 static void 931 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m) 932 { 933 struct lro_parser *pa; 934 935 /* 936 * Push up the stack of the current entry 937 * and replace it with "m". 938 */ 939 struct mbuf *msave; 940 941 /* Grab off the next and save it */ 942 msave = le->m_head->m_nextpkt; 943 le->m_head->m_nextpkt = NULL; 944 945 /* Now push out the old entry */ 946 tcp_flush_out_entry(lc, le); 947 948 /* Re-parse new header, should not fail. */ 949 pa = tcp_lro_parser(m, &le->outer, &le->inner, false); 950 KASSERT(pa != NULL, 951 ("tcp_push_and_replace: LRO parser failed on m=%p\n", m)); 952 953 /* 954 * Now to replace the data properly in the entry 955 * we have to reset the TCP header and 956 * other fields. 957 */ 958 tcp_set_entry_to_mbuf(lc, le, m, pa->tcp); 959 960 /* Restore the next list */ 961 m->m_nextpkt = msave; 962 } 963 964 static void 965 tcp_lro_mbuf_append_pkthdr(struct mbuf *m, const struct mbuf *p) 966 { 967 uint32_t csum; 968 969 if (m->m_pkthdr.lro_nsegs == 1) { 970 /* Compute relative checksum. */ 971 csum = p->m_pkthdr.lro_tcp_d_csum; 972 } else { 973 /* Merge TCP data checksums. */ 974 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum + 975 (uint32_t)p->m_pkthdr.lro_tcp_d_csum; 976 while (csum > 0xffff) 977 csum = (csum >> 16) + (csum & 0xffff); 978 } 979 980 /* Update various counters. */ 981 m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len; 982 m->m_pkthdr.lro_tcp_d_csum = csum; 983 m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len; 984 m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs; 985 } 986 987 static void 988 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le) 989 { 990 /* 991 * Walk through the mbuf chain we 992 * have on tap and compress/condense 993 * as required. 994 */ 995 uint32_t *ts_ptr; 996 struct mbuf *m; 997 struct tcphdr *th; 998 uint32_t tcp_data_len_total; 999 uint32_t tcp_data_seg_total; 1000 uint16_t tcp_data_len; 1001 uint16_t tcp_opt_len; 1002 1003 /* 1004 * First we must check the lead (m_head) 1005 * we must make sure that it is *not* 1006 * something that should be sent up 1007 * right away (sack etc). 1008 */ 1009 again: 1010 m = le->m_head->m_nextpkt; 1011 if (m == NULL) { 1012 /* Just one left. */ 1013 return; 1014 } 1015 1016 th = tcp_lro_get_th(m); 1017 tcp_opt_len = (th->th_off << 2); 1018 tcp_opt_len -= sizeof(*th); 1019 ts_ptr = (uint32_t *)(th + 1); 1020 1021 if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA || 1022 *ts_ptr != TCP_LRO_TS_OPTION)) { 1023 /* 1024 * Its not the timestamp. We can't 1025 * use this guy as the head. 1026 */ 1027 le->m_head->m_nextpkt = m->m_nextpkt; 1028 tcp_push_and_replace(lc, le, m); 1029 goto again; 1030 } 1031 if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) { 1032 /* 1033 * Make sure that previously seen segements/ACKs are delivered 1034 * before this segment, e.g. FIN. 1035 */ 1036 le->m_head->m_nextpkt = m->m_nextpkt; 1037 tcp_push_and_replace(lc, le, m); 1038 goto again; 1039 } 1040 while((m = le->m_head->m_nextpkt) != NULL) { 1041 /* 1042 * condense m into le, first 1043 * pull m out of the list. 1044 */ 1045 le->m_head->m_nextpkt = m->m_nextpkt; 1046 m->m_nextpkt = NULL; 1047 /* Setup my data */ 1048 tcp_data_len = m->m_pkthdr.lro_tcp_d_len; 1049 th = tcp_lro_get_th(m); 1050 ts_ptr = (uint32_t *)(th + 1); 1051 tcp_opt_len = (th->th_off << 2); 1052 tcp_opt_len -= sizeof(*th); 1053 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len; 1054 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs; 1055 1056 if (tcp_data_seg_total >= lc->lro_ackcnt_lim || 1057 tcp_data_len_total >= lc->lro_length_lim) { 1058 /* Flush now if appending will result in overflow. */ 1059 tcp_push_and_replace(lc, le, m); 1060 goto again; 1061 } 1062 if (tcp_opt_len != 0 && 1063 __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA || 1064 *ts_ptr != TCP_LRO_TS_OPTION)) { 1065 /* 1066 * Maybe a sack in the new one? We need to 1067 * start all over after flushing the 1068 * current le. We will go up to the beginning 1069 * and flush it (calling the replace again possibly 1070 * or just returning). 1071 */ 1072 tcp_push_and_replace(lc, le, m); 1073 goto again; 1074 } 1075 if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) { 1076 tcp_push_and_replace(lc, le, m); 1077 goto again; 1078 } 1079 if (tcp_opt_len != 0) { 1080 uint32_t tsval = ntohl(*(ts_ptr + 1)); 1081 /* Make sure timestamp values are increasing. */ 1082 if (TSTMP_GT(le->tsval, tsval)) { 1083 tcp_push_and_replace(lc, le, m); 1084 goto again; 1085 } 1086 le->tsval = tsval; 1087 le->tsecr = *(ts_ptr + 2); 1088 } 1089 /* Try to append the new segment. */ 1090 if (__predict_false(ntohl(th->th_seq) != le->next_seq || 1091 (tcp_data_len == 0 && 1092 le->ack_seq == th->th_ack && 1093 le->window == th->th_win))) { 1094 /* Out of order packet or duplicate ACK. */ 1095 tcp_push_and_replace(lc, le, m); 1096 goto again; 1097 } 1098 if (tcp_data_len != 0 || 1099 SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) { 1100 le->next_seq += tcp_data_len; 1101 le->ack_seq = th->th_ack; 1102 le->window = th->th_win; 1103 } else if (th->th_ack == le->ack_seq) { 1104 le->window = WIN_MAX(le->window, th->th_win); 1105 } 1106 1107 if (tcp_data_len == 0) { 1108 m_freem(m); 1109 continue; 1110 } 1111 1112 /* Merge TCP data checksum and length to head mbuf. */ 1113 tcp_lro_mbuf_append_pkthdr(le->m_head, m); 1114 1115 /* 1116 * Adjust the mbuf so that m_data points to the first byte of 1117 * the ULP payload. Adjust the mbuf to avoid complications and 1118 * append new segment to existing mbuf chain. 1119 */ 1120 m_adj(m, m->m_pkthdr.len - tcp_data_len); 1121 m_demote_pkthdr(m); 1122 le->m_tail->m_next = m; 1123 le->m_tail = m_last(m); 1124 } 1125 } 1126 1127 #ifdef TCPHPTS 1128 static void 1129 tcp_queue_pkts(struct inpcb *inp, struct tcpcb *tp, struct lro_entry *le) 1130 { 1131 INP_WLOCK_ASSERT(inp); 1132 if (tp->t_in_pkt == NULL) { 1133 /* Nothing yet there */ 1134 tp->t_in_pkt = le->m_head; 1135 tp->t_tail_pkt = le->m_last_mbuf; 1136 } else { 1137 /* Already some there */ 1138 tp->t_tail_pkt->m_nextpkt = le->m_head; 1139 tp->t_tail_pkt = le->m_last_mbuf; 1140 } 1141 le->m_head = NULL; 1142 le->m_last_mbuf = NULL; 1143 } 1144 1145 static struct mbuf * 1146 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le, 1147 struct inpcb *inp, int32_t *new_m) 1148 { 1149 struct tcpcb *tp; 1150 struct mbuf *m; 1151 1152 tp = intotcpcb(inp); 1153 if (__predict_false(tp == NULL)) 1154 return (NULL); 1155 1156 /* Look at the last mbuf if any in queue */ 1157 m = tp->t_tail_pkt; 1158 if (m != NULL && (m->m_flags & M_ACKCMP) != 0) { 1159 if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) { 1160 tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0); 1161 *new_m = 0; 1162 counter_u64_add(tcp_extra_mbuf, 1); 1163 return (m); 1164 } else { 1165 /* Mark we ran out of space */ 1166 inp->inp_flags2 |= INP_MBUF_L_ACKS; 1167 } 1168 } 1169 /* Decide mbuf size. */ 1170 if (inp->inp_flags2 & INP_MBUF_L_ACKS) 1171 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR); 1172 else 1173 m = m_gethdr(M_NOWAIT, MT_DATA); 1174 1175 if (__predict_false(m == NULL)) { 1176 counter_u64_add(tcp_would_have_but, 1); 1177 return (NULL); 1178 } 1179 counter_u64_add(tcp_comp_total, 1); 1180 m->m_flags |= M_ACKCMP; 1181 *new_m = 1; 1182 return (m); 1183 } 1184 1185 static struct inpcb * 1186 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa) 1187 { 1188 struct inpcb *inp; 1189 1190 switch (pa->data.lro_type) { 1191 #ifdef INET6 1192 case LRO_TYPE_IPV6_TCP: 1193 inp = in6_pcblookup(&V_tcbinfo, 1194 &pa->data.s_addr.v6, 1195 pa->data.s_port, 1196 &pa->data.d_addr.v6, 1197 pa->data.d_port, 1198 INPLOOKUP_WLOCKPCB, 1199 ifp); 1200 break; 1201 #endif 1202 #ifdef INET 1203 case LRO_TYPE_IPV4_TCP: 1204 inp = in_pcblookup(&V_tcbinfo, 1205 pa->data.s_addr.v4, 1206 pa->data.s_port, 1207 pa->data.d_addr.v4, 1208 pa->data.d_port, 1209 INPLOOKUP_WLOCKPCB, 1210 ifp); 1211 break; 1212 #endif 1213 default: 1214 inp = NULL; 1215 break; 1216 } 1217 return (inp); 1218 } 1219 1220 static inline bool 1221 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts) 1222 { 1223 /* 1224 * This function returns two bits of valuable information. 1225 * a) Is what is present capable of being ack-compressed, 1226 * we can ack-compress if there is no options or just 1227 * a timestamp option, and of course the th_flags must 1228 * be correct as well. 1229 * b) Our other options present such as SACK. This is 1230 * used to determine if we want to wakeup or not. 1231 */ 1232 bool ret = true; 1233 1234 switch (th->th_off << 2) { 1235 case (sizeof(*th) + TCPOLEN_TSTAMP_APPA): 1236 *ppts = (uint32_t *)(th + 1); 1237 /* Check if we have only one timestamp option. */ 1238 if (**ppts == TCP_LRO_TS_OPTION) 1239 *other_opts = false; 1240 else { 1241 *other_opts = true; 1242 ret = false; 1243 } 1244 break; 1245 case (sizeof(*th)): 1246 /* No options. */ 1247 *ppts = NULL; 1248 *other_opts = false; 1249 break; 1250 default: 1251 *ppts = NULL; 1252 *other_opts = true; 1253 ret = false; 1254 break; 1255 } 1256 /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */ 1257 if ((th->th_flags & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0) 1258 ret = false; 1259 /* If it has data on it we cannot compress it */ 1260 if (m->m_pkthdr.lro_tcp_d_len) 1261 ret = false; 1262 1263 /* ACK flag must be set. */ 1264 if (!(th->th_flags & TH_ACK)) 1265 ret = false; 1266 return (ret); 1267 } 1268 1269 static int 1270 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le) 1271 { 1272 struct inpcb *inp; 1273 struct tcpcb *tp; 1274 struct mbuf **pp, *cmp, *mv_to; 1275 bool bpf_req, should_wake; 1276 1277 /* Check if packet doesn't belongs to our network interface. */ 1278 if ((tcplro_stacks_wanting_mbufq == 0) || 1279 (le->outer.data.vlan_id != 0) || 1280 (le->inner.data.lro_type != LRO_TYPE_NONE)) 1281 return (TCP_LRO_CANNOT); 1282 1283 #ifdef INET6 1284 /* 1285 * Be proactive about unspecified IPv6 address in source. As 1286 * we use all-zero to indicate unbounded/unconnected pcb, 1287 * unspecified IPv6 address can be used to confuse us. 1288 * 1289 * Note that packets with unspecified IPv6 destination is 1290 * already dropped in ip6_input. 1291 */ 1292 if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP && 1293 IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6))) 1294 return (TCP_LRO_CANNOT); 1295 1296 if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP && 1297 IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6))) 1298 return (TCP_LRO_CANNOT); 1299 #endif 1300 /* Lookup inp, if any. */ 1301 inp = tcp_lro_lookup(lc->ifp, 1302 (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner); 1303 if (inp == NULL) 1304 return (TCP_LRO_CANNOT); 1305 1306 counter_u64_add(tcp_inp_lro_locks_taken, 1); 1307 1308 /* Get TCP control structure. */ 1309 tp = intotcpcb(inp); 1310 1311 /* Check if the inp is dead, Jim. */ 1312 if (tp == NULL || 1313 (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) || 1314 (inp->inp_flags2 & INP_FREED)) { 1315 INP_WUNLOCK(inp); 1316 return (TCP_LRO_CANNOT); 1317 } 1318 if ((inp->inp_irq_cpu_set == 0) && (lc->lro_cpu_is_set == 1)) { 1319 inp->inp_irq_cpu = lc->lro_last_cpu; 1320 inp->inp_irq_cpu_set = 1; 1321 } 1322 /* Check if the transport doesn't support the needed optimizations. */ 1323 if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) { 1324 INP_WUNLOCK(inp); 1325 return (TCP_LRO_CANNOT); 1326 } 1327 1328 if (inp->inp_flags2 & INP_MBUF_QUEUE_READY) 1329 should_wake = false; 1330 else 1331 should_wake = true; 1332 /* Check if packets should be tapped to BPF. */ 1333 bpf_req = bpf_peers_present(lc->ifp->if_bpf); 1334 1335 /* Strip and compress all the incoming packets. */ 1336 cmp = NULL; 1337 for (pp = &le->m_head; *pp != NULL; ) { 1338 mv_to = NULL; 1339 if (do_bpf_strip_and_compress(inp, lc, le, pp, 1340 &cmp, &mv_to, &should_wake, bpf_req ) == false) { 1341 /* Advance to next mbuf. */ 1342 pp = &(*pp)->m_nextpkt; 1343 } else if (mv_to != NULL) { 1344 /* We are asked to move pp up */ 1345 pp = &mv_to->m_nextpkt; 1346 } 1347 } 1348 /* Update "m_last_mbuf", if any. */ 1349 if (pp == &le->m_head) 1350 le->m_last_mbuf = *pp; 1351 else 1352 le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt); 1353 1354 /* Check if any data mbufs left. */ 1355 if (le->m_head != NULL) { 1356 counter_u64_add(tcp_inp_lro_direct_queue, 1); 1357 tcp_lro_log(tp, lc, le, NULL, 22, 1, 1358 inp->inp_flags2, inp->inp_in_input, 1); 1359 tcp_queue_pkts(inp, tp, le); 1360 } 1361 if (should_wake) { 1362 /* Wakeup */ 1363 counter_u64_add(tcp_inp_lro_wokeup_queue, 1); 1364 if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0)) 1365 inp = NULL; 1366 } 1367 if (inp != NULL) 1368 INP_WUNLOCK(inp); 1369 return (0); /* Success. */ 1370 } 1371 #endif 1372 1373 void 1374 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le) 1375 { 1376 /* Only optimise if there are multiple packets waiting. */ 1377 #ifdef TCPHPTS 1378 int error; 1379 #endif 1380 1381 NET_EPOCH_ASSERT(); 1382 #ifdef TCPHPTS 1383 CURVNET_SET(lc->ifp->if_vnet); 1384 error = tcp_lro_flush_tcphpts(lc, le); 1385 CURVNET_RESTORE(); 1386 if (error != 0) { 1387 #endif 1388 tcp_lro_condense(lc, le); 1389 tcp_flush_out_entry(lc, le); 1390 #ifdef TCPHPTS 1391 } 1392 #endif 1393 lc->lro_flushed++; 1394 bzero(le, sizeof(*le)); 1395 LIST_INSERT_HEAD(&lc->lro_free, le, next); 1396 } 1397 1398 #ifdef HAVE_INLINE_FLSLL 1399 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1)) 1400 #else 1401 static inline uint64_t 1402 tcp_lro_msb_64(uint64_t x) 1403 { 1404 x |= (x >> 1); 1405 x |= (x >> 2); 1406 x |= (x >> 4); 1407 x |= (x >> 8); 1408 x |= (x >> 16); 1409 x |= (x >> 32); 1410 return (x & ~(x >> 1)); 1411 } 1412 #endif 1413 1414 /* 1415 * The tcp_lro_sort() routine is comparable to qsort(), except it has 1416 * a worst case complexity limit of O(MIN(N,64)*N), where N is the 1417 * number of elements to sort and 64 is the number of sequence bits 1418 * available. The algorithm is bit-slicing the 64-bit sequence number, 1419 * sorting one bit at a time from the most significant bit until the 1420 * least significant one, skipping the constant bits. This is 1421 * typically called a radix sort. 1422 */ 1423 static void 1424 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size) 1425 { 1426 struct lro_mbuf_sort temp; 1427 uint64_t ones; 1428 uint64_t zeros; 1429 uint32_t x; 1430 uint32_t y; 1431 1432 repeat: 1433 /* for small arrays insertion sort is faster */ 1434 if (size <= 12) { 1435 for (x = 1; x < size; x++) { 1436 temp = parray[x]; 1437 for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--) 1438 parray[y] = parray[y - 1]; 1439 parray[y] = temp; 1440 } 1441 return; 1442 } 1443 1444 /* compute sequence bits which are constant */ 1445 ones = 0; 1446 zeros = 0; 1447 for (x = 0; x != size; x++) { 1448 ones |= parray[x].seq; 1449 zeros |= ~parray[x].seq; 1450 } 1451 1452 /* compute bits which are not constant into "ones" */ 1453 ones &= zeros; 1454 if (ones == 0) 1455 return; 1456 1457 /* pick the most significant bit which is not constant */ 1458 ones = tcp_lro_msb_64(ones); 1459 1460 /* 1461 * Move entries having cleared sequence bits to the beginning 1462 * of the array: 1463 */ 1464 for (x = y = 0; y != size; y++) { 1465 /* skip set bits */ 1466 if (parray[y].seq & ones) 1467 continue; 1468 /* swap entries */ 1469 temp = parray[x]; 1470 parray[x] = parray[y]; 1471 parray[y] = temp; 1472 x++; 1473 } 1474 1475 KASSERT(x != 0 && x != size, ("Memory is corrupted\n")); 1476 1477 /* sort zeros */ 1478 tcp_lro_sort(parray, x); 1479 1480 /* sort ones */ 1481 parray += x; 1482 size -= x; 1483 goto repeat; 1484 } 1485 1486 void 1487 tcp_lro_flush_all(struct lro_ctrl *lc) 1488 { 1489 uint64_t seq; 1490 uint64_t nseq; 1491 unsigned x; 1492 1493 NET_EPOCH_ASSERT(); 1494 /* check if no mbufs to flush */ 1495 if (lc->lro_mbuf_count == 0) 1496 goto done; 1497 if (lc->lro_cpu_is_set == 0) { 1498 if (lc->lro_last_cpu == curcpu) { 1499 lc->lro_cnt_of_same_cpu++; 1500 /* Have we reached the threshold to declare a cpu? */ 1501 if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh) 1502 lc->lro_cpu_is_set = 1; 1503 } else { 1504 lc->lro_last_cpu = curcpu; 1505 lc->lro_cnt_of_same_cpu = 0; 1506 } 1507 } 1508 CURVNET_SET(lc->ifp->if_vnet); 1509 1510 /* get current time */ 1511 binuptime(&lc->lro_last_queue_time); 1512 1513 /* sort all mbufs according to stream */ 1514 tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count); 1515 1516 /* input data into LRO engine, stream by stream */ 1517 seq = 0; 1518 for (x = 0; x != lc->lro_mbuf_count; x++) { 1519 struct mbuf *mb; 1520 1521 /* get mbuf */ 1522 mb = lc->lro_mbuf_data[x].mb; 1523 1524 /* get sequence number, masking away the packet index */ 1525 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24); 1526 1527 /* check for new stream */ 1528 if (seq != nseq) { 1529 seq = nseq; 1530 1531 /* flush active streams */ 1532 tcp_lro_rx_done(lc); 1533 } 1534 1535 /* add packet to LRO engine */ 1536 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) { 1537 /* input packet to network layer */ 1538 (*lc->ifp->if_input)(lc->ifp, mb); 1539 lc->lro_queued++; 1540 lc->lro_flushed++; 1541 } 1542 } 1543 CURVNET_RESTORE(); 1544 done: 1545 /* flush active streams */ 1546 tcp_lro_rx_done(lc); 1547 1548 #ifdef TCPHPTS 1549 tcp_run_hpts(); 1550 #endif 1551 lc->lro_mbuf_count = 0; 1552 } 1553 1554 #ifdef TCPHPTS 1555 static void 1556 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m, 1557 uint32_t *ts_ptr, uint16_t iptos) 1558 { 1559 /* 1560 * Given a TCP ACK, summarize it down into the small TCP ACK 1561 * entry. 1562 */ 1563 ae->timestamp = m->m_pkthdr.rcv_tstmp; 1564 if (m->m_flags & M_TSTMP_LRO) 1565 ae->flags = TSTMP_LRO; 1566 else if (m->m_flags & M_TSTMP) 1567 ae->flags = TSTMP_HDWR; 1568 ae->seq = ntohl(th->th_seq); 1569 ae->ack = ntohl(th->th_ack); 1570 ae->flags |= th->th_flags; 1571 if (ts_ptr != NULL) { 1572 ae->ts_value = ntohl(ts_ptr[1]); 1573 ae->ts_echo = ntohl(ts_ptr[2]); 1574 ae->flags |= HAS_TSTMP; 1575 } 1576 ae->win = ntohs(th->th_win); 1577 ae->codepoint = iptos; 1578 } 1579 1580 /* 1581 * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets 1582 * and strip all, but the IPv4/IPv6 header. 1583 */ 1584 static bool 1585 do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc, 1586 struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to, 1587 bool *should_wake, bool bpf_req) 1588 { 1589 union { 1590 void *ptr; 1591 struct ip *ip4; 1592 struct ip6_hdr *ip6; 1593 } l3; 1594 struct mbuf *m; 1595 struct mbuf *nm; 1596 struct tcphdr *th; 1597 struct tcp_ackent *ack_ent; 1598 uint32_t *ts_ptr; 1599 int32_t n_mbuf; 1600 bool other_opts, can_compress; 1601 uint8_t lro_type; 1602 uint16_t iptos; 1603 int tcp_hdr_offset; 1604 int idx; 1605 1606 /* Get current mbuf. */ 1607 m = *pp; 1608 1609 /* Let the BPF see the packet */ 1610 if (__predict_false(bpf_req)) 1611 ETHER_BPF_MTAP(lc->ifp, m); 1612 1613 tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off; 1614 lro_type = le->inner.data.lro_type; 1615 switch (lro_type) { 1616 case LRO_TYPE_NONE: 1617 lro_type = le->outer.data.lro_type; 1618 switch (lro_type) { 1619 case LRO_TYPE_IPV4_TCP: 1620 tcp_hdr_offset -= sizeof(*le->outer.ip4); 1621 m->m_pkthdr.lro_etype = ETHERTYPE_IP; 1622 break; 1623 case LRO_TYPE_IPV6_TCP: 1624 tcp_hdr_offset -= sizeof(*le->outer.ip6); 1625 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6; 1626 break; 1627 default: 1628 goto compressed; 1629 } 1630 break; 1631 case LRO_TYPE_IPV4_TCP: 1632 tcp_hdr_offset -= sizeof(*le->outer.ip4); 1633 m->m_pkthdr.lro_etype = ETHERTYPE_IP; 1634 break; 1635 case LRO_TYPE_IPV6_TCP: 1636 tcp_hdr_offset -= sizeof(*le->outer.ip6); 1637 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6; 1638 break; 1639 default: 1640 goto compressed; 1641 } 1642 1643 MPASS(tcp_hdr_offset >= 0); 1644 1645 m_adj(m, tcp_hdr_offset); 1646 m->m_flags |= M_LRO_EHDRSTRP; 1647 m->m_flags &= ~M_ACKCMP; 1648 m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset; 1649 1650 th = tcp_lro_get_th(m); 1651 1652 th->th_sum = 0; /* TCP checksum is valid. */ 1653 1654 /* Check if ACK can be compressed */ 1655 can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts); 1656 1657 /* Now lets look at the should wake states */ 1658 if ((other_opts == true) && 1659 ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) { 1660 /* 1661 * If there are other options (SACK?) and the 1662 * tcp endpoint has not expressly told us it does 1663 * not care about SACKS, then we should wake up. 1664 */ 1665 *should_wake = true; 1666 } 1667 /* Is the ack compressable? */ 1668 if (can_compress == false) 1669 goto done; 1670 /* Does the TCP endpoint support ACK compression? */ 1671 if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0) 1672 goto done; 1673 1674 /* Lets get the TOS/traffic class field */ 1675 l3.ptr = mtod(m, void *); 1676 switch (lro_type) { 1677 case LRO_TYPE_IPV4_TCP: 1678 iptos = l3.ip4->ip_tos; 1679 break; 1680 case LRO_TYPE_IPV6_TCP: 1681 iptos = IPV6_TRAFFIC_CLASS(l3.ip6); 1682 break; 1683 default: 1684 iptos = 0; /* Keep compiler happy. */ 1685 break; 1686 } 1687 /* Now lets get space if we don't have some already */ 1688 if (*cmp == NULL) { 1689 new_one: 1690 nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf); 1691 if (__predict_false(nm == NULL)) 1692 goto done; 1693 *cmp = nm; 1694 if (n_mbuf) { 1695 /* 1696 * Link in the new cmp ack to our in-order place, 1697 * first set our cmp ack's next to where we are. 1698 */ 1699 nm->m_nextpkt = m; 1700 (*pp) = nm; 1701 /* 1702 * Set it up so mv_to is advanced to our 1703 * compressed ack. This way the caller can 1704 * advance pp to the right place. 1705 */ 1706 *mv_to = nm; 1707 /* 1708 * Advance it here locally as well. 1709 */ 1710 pp = &nm->m_nextpkt; 1711 } 1712 } else { 1713 /* We have one already we are working on */ 1714 nm = *cmp; 1715 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) { 1716 /* We ran out of space */ 1717 inp->inp_flags2 |= INP_MBUF_L_ACKS; 1718 goto new_one; 1719 } 1720 } 1721 MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent)); 1722 counter_u64_add(tcp_inp_lro_compressed, 1); 1723 le->compressed++; 1724 /* We can add in to the one on the tail */ 1725 ack_ent = mtod(nm, struct tcp_ackent *); 1726 idx = (nm->m_len / sizeof(struct tcp_ackent)); 1727 build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos); 1728 1729 /* Bump the size of both pkt-hdr and len */ 1730 nm->m_len += sizeof(struct tcp_ackent); 1731 nm->m_pkthdr.len += sizeof(struct tcp_ackent); 1732 compressed: 1733 /* Advance to next mbuf before freeing. */ 1734 *pp = m->m_nextpkt; 1735 m->m_nextpkt = NULL; 1736 m_freem(m); 1737 return (true); 1738 done: 1739 counter_u64_add(tcp_uncomp_total, 1); 1740 le->uncompressed++; 1741 return (false); 1742 } 1743 #endif 1744 1745 static struct lro_head * 1746 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser) 1747 { 1748 u_long hash; 1749 1750 if (M_HASHTYPE_ISHASH(m)) { 1751 hash = m->m_pkthdr.flowid; 1752 } else { 1753 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++) 1754 hash += parser->data.raw[i]; 1755 } 1756 return (&lc->lro_hash[hash % lc->lro_hashsz]); 1757 } 1758 1759 static int 1760 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash) 1761 { 1762 struct lro_parser pi; /* inner address data */ 1763 struct lro_parser po; /* outer address data */ 1764 struct lro_parser *pa; /* current parser for TCP stream */ 1765 struct lro_entry *le; 1766 struct lro_head *bucket; 1767 struct tcphdr *th; 1768 int tcp_data_len; 1769 int tcp_opt_len; 1770 int error; 1771 uint16_t tcp_data_sum; 1772 1773 #ifdef INET 1774 /* Quickly decide if packet cannot be LRO'ed */ 1775 if (__predict_false(V_ipforwarding != 0)) 1776 return (TCP_LRO_CANNOT); 1777 #endif 1778 #ifdef INET6 1779 /* Quickly decide if packet cannot be LRO'ed */ 1780 if (__predict_false(V_ip6_forwarding != 0)) 1781 return (TCP_LRO_CANNOT); 1782 #endif 1783 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) != 1784 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) || 1785 (m->m_pkthdr.csum_data != 0xffff)) { 1786 /* 1787 * The checksum either did not have hardware offload 1788 * or it was a bad checksum. We can't LRO such 1789 * a packet. 1790 */ 1791 counter_u64_add(tcp_bad_csums, 1); 1792 return (TCP_LRO_CANNOT); 1793 } 1794 /* We expect a contiguous header [eh, ip, tcp]. */ 1795 pa = tcp_lro_parser(m, &po, &pi, true); 1796 if (__predict_false(pa == NULL)) 1797 return (TCP_LRO_NOT_SUPPORTED); 1798 1799 /* We don't expect any padding. */ 1800 error = tcp_lro_trim_mbuf_chain(m, pa); 1801 if (__predict_false(error != 0)) 1802 return (error); 1803 1804 #ifdef INET 1805 switch (pa->data.lro_type) { 1806 case LRO_TYPE_IPV4_TCP: 1807 error = tcp_lro_rx_ipv4(lc, m, pa->ip4); 1808 if (__predict_false(error != 0)) 1809 return (error); 1810 break; 1811 default: 1812 break; 1813 } 1814 #endif 1815 /* If no hardware or arrival stamp on the packet add timestamp */ 1816 if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) { 1817 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time); 1818 m->m_flags |= M_TSTMP_LRO; 1819 } 1820 1821 /* Get pointer to TCP header. */ 1822 th = pa->tcp; 1823 1824 /* Don't process SYN packets. */ 1825 if (__predict_false(th->th_flags & TH_SYN)) 1826 return (TCP_LRO_CANNOT); 1827 1828 /* Get total TCP header length and compute payload length. */ 1829 tcp_opt_len = (th->th_off << 2); 1830 tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th - 1831 (uint8_t *)m->m_data) - tcp_opt_len; 1832 tcp_opt_len -= sizeof(*th); 1833 1834 /* Don't process invalid TCP headers. */ 1835 if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0)) 1836 return (TCP_LRO_CANNOT); 1837 1838 /* Compute TCP data only checksum. */ 1839 if (tcp_data_len == 0) 1840 tcp_data_sum = 0; /* no data, no checksum */ 1841 else if (__predict_false(csum != 0)) 1842 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum); 1843 else 1844 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum); 1845 1846 /* Save TCP info in mbuf. */ 1847 m->m_nextpkt = NULL; 1848 m->m_pkthdr.rcvif = lc->ifp; 1849 m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum; 1850 m->m_pkthdr.lro_tcp_d_len = tcp_data_len; 1851 m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data); 1852 m->m_pkthdr.lro_nsegs = 1; 1853 1854 /* Get hash bucket. */ 1855 if (!use_hash) { 1856 bucket = &lc->lro_hash[0]; 1857 } else { 1858 bucket = tcp_lro_rx_get_bucket(lc, m, pa); 1859 } 1860 1861 /* Try to find a matching previous segment. */ 1862 LIST_FOREACH(le, bucket, hash_next) { 1863 /* Compare addresses and ports. */ 1864 if (lro_address_compare(&po.data, &le->outer.data) == false || 1865 lro_address_compare(&pi.data, &le->inner.data) == false) 1866 continue; 1867 1868 /* Check if no data and old ACK. */ 1869 if (tcp_data_len == 0 && 1870 SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) { 1871 m_freem(m); 1872 return (0); 1873 } 1874 1875 /* Mark "m" in the last spot. */ 1876 le->m_last_mbuf->m_nextpkt = m; 1877 /* Now set the tail to "m". */ 1878 le->m_last_mbuf = m; 1879 return (0); 1880 } 1881 1882 /* Try to find an empty slot. */ 1883 if (LIST_EMPTY(&lc->lro_free)) 1884 return (TCP_LRO_NO_ENTRIES); 1885 1886 /* Start a new segment chain. */ 1887 le = LIST_FIRST(&lc->lro_free); 1888 LIST_REMOVE(le, next); 1889 tcp_lro_active_insert(lc, bucket, le); 1890 1891 /* Make sure the headers are set. */ 1892 le->inner = pi; 1893 le->outer = po; 1894 1895 /* Store time this entry was allocated. */ 1896 le->alloc_time = lc->lro_last_queue_time; 1897 1898 tcp_set_entry_to_mbuf(lc, le, m, th); 1899 1900 /* Now set the tail to "m". */ 1901 le->m_last_mbuf = m; 1902 1903 return (0); 1904 } 1905 1906 int 1907 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum) 1908 { 1909 int error; 1910 1911 if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) != 1912 ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) || 1913 (m->m_pkthdr.csum_data != 0xffff)) { 1914 /* 1915 * The checksum either did not have hardware offload 1916 * or it was a bad checksum. We can't LRO such 1917 * a packet. 1918 */ 1919 counter_u64_add(tcp_bad_csums, 1); 1920 return (TCP_LRO_CANNOT); 1921 } 1922 /* get current time */ 1923 binuptime(&lc->lro_last_queue_time); 1924 CURVNET_SET(lc->ifp->if_vnet); 1925 error = tcp_lro_rx_common(lc, m, csum, true); 1926 CURVNET_RESTORE(); 1927 1928 return (error); 1929 } 1930 1931 void 1932 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb) 1933 { 1934 NET_EPOCH_ASSERT(); 1935 /* sanity checks */ 1936 if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL || 1937 lc->lro_mbuf_max == 0)) { 1938 /* packet drop */ 1939 m_freem(mb); 1940 return; 1941 } 1942 1943 /* check if packet is not LRO capable */ 1944 if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) { 1945 /* input packet to network layer */ 1946 (*lc->ifp->if_input) (lc->ifp, mb); 1947 return; 1948 } 1949 1950 /* create sequence number */ 1951 lc->lro_mbuf_data[lc->lro_mbuf_count].seq = 1952 (((uint64_t)M_HASHTYPE_GET(mb)) << 56) | 1953 (((uint64_t)mb->m_pkthdr.flowid) << 24) | 1954 ((uint64_t)lc->lro_mbuf_count); 1955 1956 /* enter mbuf */ 1957 lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb; 1958 1959 /* flush if array is full */ 1960 if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max)) 1961 tcp_lro_flush_all(lc); 1962 } 1963 1964 /* end */ 1965