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