sys/netinet/tcp_lro.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   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 #include "opt_inet.h"
  37 #include "opt_inet6.h"
  38
  39 #include <sys/param.h>
  40 #include <sys/systm.h>
  41 #include <sys/kernel.h>
  42 #include <sys/malloc.h>
  43 #include <sys/mbuf.h>
  44 #include <sys/socket.h>
  45 #include <sys/socketvar.h>
  46 #include <sys/sockbuf.h>
  47 #include <sys/sysctl.h>
  48
  49 #include <net/if.h>
  50 #include <net/if_var.h>
  51 #include <net/ethernet.h>
  52 #include <net/bpf.h>
  53 #include <net/vnet.h>
  54 #include <net/if_dl.h>
  55 #include <net/if_media.h>
  56 #include <net/if_private.h>
  57 #include <net/if_types.h>
  58 #include <net/infiniband.h>
  59 #include <net/if_lagg.h>
  60
  61 #include <netinet/in_systm.h>
  62 #include <netinet/in.h>
  63 #include <netinet/ip6.h>
  64 #include <netinet/ip.h>
  65 #include <netinet/ip_var.h>
  66 #include <netinet/in_pcb.h>
  67 #include <netinet6/in6_pcb.h>
  68 #include <netinet/tcp.h>
  69 #include <netinet/tcp_seq.h>
  70 #include <netinet/tcp_lro.h>
  71 #include <netinet/tcp_var.h>
  72 #include <netinet/tcpip.h>
  73 #include <netinet/tcp_hpts.h>
  74 #include <netinet/tcp_log_buf.h>
  75 #include <netinet/tcp_fsm.h>
  76 #include <netinet/udp.h>
  77 #include <netinet6/ip6_var.h>
  78
  79 #include <machine/in_cksum.h>
  80
  81 static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures");
  82
  83 #define TCP_LRO_TS_OPTION \
  84     ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \
  85           (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)
  86
  87 static void     tcp_lro_rx_done(struct lro_ctrl *lc);
  88 static int      tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m,
  89                     uint32_t csum, bool use_hash);
  90
  91 #ifdef TCPHPTS
  92 static bool     do_bpf_strip_and_compress(struct tcpcb *, struct lro_ctrl *,
  93                 struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **,
  94                 bool *, bool, bool, struct ifnet *, bool);
  95
  96 #endif
  97
  98 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro,  CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
  99     "TCP LRO");
 100
 101 static long tcplro_stacks_wanting_mbufq;
 102 counter_u64_t tcp_inp_lro_direct_queue;
 103 counter_u64_t tcp_inp_lro_wokeup_queue;
 104 counter_u64_t tcp_inp_lro_compressed;
 105 counter_u64_t tcp_inp_lro_locks_taken;
 106 counter_u64_t tcp_extra_mbuf;
 107 counter_u64_t tcp_would_have_but;
 108 counter_u64_t tcp_comp_total;
 109 counter_u64_t tcp_uncomp_total;
 110 counter_u64_t tcp_bad_csums;
 111
 112 static unsigned tcp_lro_entries = TCP_LRO_ENTRIES;
 113 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries,
 114     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0,
 115     "default number of LRO entries");
 116
 117 static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH;
 118 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold,
 119     CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0,
 120     "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?");
 121
 122 static uint32_t tcp_less_accurate_lro_ts = 0;
 123 SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_less_accurate,
 124     CTLFLAG_MPSAFE, &tcp_less_accurate_lro_ts, 0,
 125     "Do we trade off efficency by doing less timestamp operations for time accuracy?");
 126
 127 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
 128     &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
 129 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
 130     &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
 131 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
 132     &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
 133 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
 134     &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
 135 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
 136     &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
 137 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
 138     &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
 139 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
 140     &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
 141 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
 142     &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
 143 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD,
 144     &tcp_bad_csums, "Number of packets that the common code saw with bad csums");
 145
 146 void
 147 tcp_lro_reg_mbufq(void)
 148 {
 149         atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
 150 }
 151
 152 void
 153 tcp_lro_dereg_mbufq(void)
 154 {
 155         atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
 156 }
 157
 158 static __inline void
 159 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
 160     struct lro_entry *le)
 161 {
 162
 163         LIST_INSERT_HEAD(&lc->lro_active, le, next);
 164         LIST_INSERT_HEAD(bucket, le, hash_next);
 165 }
 166
 167 static __inline void
 168 tcp_lro_active_remove(struct lro_entry *le)
 169 {
 170
 171         LIST_REMOVE(le, next);          /* active list */
 172         LIST_REMOVE(le, hash_next);     /* hash bucket */
 173 }
 174
 175 int
 176 tcp_lro_init(struct lro_ctrl *lc)
 177 {
 178         return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
 179 }
 180
 181 int
 182 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
 183     unsigned lro_entries, unsigned lro_mbufs)
 184 {
 185         struct lro_entry *le;
 186         size_t size;
 187         unsigned i, elements;
 188
 189         lc->lro_bad_csum = 0;
 190         lc->lro_queued = 0;
 191         lc->lro_flushed = 0;
 192         lc->lro_mbuf_count = 0;
 193         lc->lro_mbuf_max = lro_mbufs;
 194         lc->lro_cnt = lro_entries;
 195         lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
 196         lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
 197         lc->ifp = ifp;
 198         LIST_INIT(&lc->lro_free);
 199         LIST_INIT(&lc->lro_active);
 200
 201         /* create hash table to accelerate entry lookup */
 202         if (lro_entries > lro_mbufs)
 203                 elements = lro_entries;
 204         else
 205                 elements = lro_mbufs;
 206         lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
 207             HASH_NOWAIT);
 208         if (lc->lro_hash == NULL) {
 209                 memset(lc, 0, sizeof(*lc));
 210                 return (ENOMEM);
 211         }
 212
 213         /* compute size to allocate */
 214         size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
 215             (lro_entries * sizeof(*le));
 216         lc->lro_mbuf_data = (struct lro_mbuf_sort *)
 217             malloc(size, M_LRO, M_NOWAIT | M_ZERO);
 218
 219         /* check for out of memory */
 220         if (lc->lro_mbuf_data == NULL) {
 221                 free(lc->lro_hash, M_LRO);
 222                 memset(lc, 0, sizeof(*lc));
 223                 return (ENOMEM);
 224         }
 225         /* compute offset for LRO entries */
 226         le = (struct lro_entry *)
 227             (lc->lro_mbuf_data + lro_mbufs);
 228
 229         /* setup linked list */
 230         for (i = 0; i != lro_entries; i++)
 231                 LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
 232
 233         return (0);
 234 }
 235
 236 struct vxlan_header {
 237         uint32_t        vxlh_flags;
 238         uint32_t        vxlh_vni;
 239 };
 240
 241 static inline void *
 242 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
 243 {
 244         const struct ether_vlan_header *eh;
 245         void *old;
 246         uint16_t eth_type;
 247
 248         if (update_data)
 249                 memset(parser, 0, sizeof(*parser));
 250
 251         old = ptr;
 252
 253         if (is_vxlan) {
 254                 const struct vxlan_header *vxh;
 255                 vxh = ptr;
 256                 ptr = (uint8_t *)ptr + sizeof(*vxh);
 257                 if (update_data) {
 258                         parser->data.vxlan_vni =
 259                             vxh->vxlh_vni & htonl(0xffffff00);
 260                 }
 261         }
 262
 263         eh = ptr;
 264         if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
 265                 eth_type = eh->evl_proto;
 266                 if (update_data) {
 267                         /* strip priority and keep VLAN ID only */
 268                         parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
 269                 }
 270                 /* advance to next header */
 271                 ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
 272                 mlen -= (ETHER_HDR_LEN  + ETHER_VLAN_ENCAP_LEN);
 273         } else {
 274                 eth_type = eh->evl_encap_proto;
 275                 /* advance to next header */
 276                 mlen -= ETHER_HDR_LEN;
 277                 ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
 278         }
 279         if (__predict_false(mlen <= 0))
 280                 return (NULL);
 281         switch (eth_type) {
 282 #ifdef INET
 283         case htons(ETHERTYPE_IP):
 284                 parser->ip4 = ptr;
 285                 if (__predict_false(mlen < sizeof(struct ip)))
 286                         return (NULL);
 287                 /* Ensure there are no IPv4 options. */
 288                 if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
 289                         break;
 290                 /* .. and the packet is not fragmented. */
 291                 if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
 292                         break;
 293                 /* .. and the packet has valid src/dst addrs */
 294                 if (__predict_false(parser->ip4->ip_src.s_addr == INADDR_ANY ||
 295                         parser->ip4->ip_dst.s_addr == INADDR_ANY))
 296                         break;
 297                 ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
 298                 mlen -= sizeof(struct ip);
 299                 if (update_data) {
 300                         parser->data.s_addr.v4 = parser->ip4->ip_src;
 301                         parser->data.d_addr.v4 = parser->ip4->ip_dst;
 302                 }
 303                 switch (parser->ip4->ip_p) {
 304                 case IPPROTO_UDP:
 305                         if (__predict_false(mlen < sizeof(struct udphdr)))
 306                                 return (NULL);
 307                         parser->udp = ptr;
 308                         if (update_data) {
 309                                 parser->data.lro_type = LRO_TYPE_IPV4_UDP;
 310                                 parser->data.s_port = parser->udp->uh_sport;
 311                                 parser->data.d_port = parser->udp->uh_dport;
 312                         } else {
 313                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
 314                         }
 315                         ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
 316                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
 317                         return (ptr);
 318                 case IPPROTO_TCP:
 319                         parser->tcp = ptr;
 320                         if (__predict_false(mlen < sizeof(struct tcphdr)))
 321                                 return (NULL);
 322                         if (update_data) {
 323                                 parser->data.lro_type = LRO_TYPE_IPV4_TCP;
 324                                 parser->data.s_port = parser->tcp->th_sport;
 325                                 parser->data.d_port = parser->tcp->th_dport;
 326                         } else {
 327                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
 328                         }
 329                         if (__predict_false(mlen < (parser->tcp->th_off << 2)))
 330                                 return (NULL);
 331                         ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
 332                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
 333                         return (ptr);
 334                 default:
 335                         break;
 336                 }
 337                 break;
 338 #endif
 339 #ifdef INET6
 340         case htons(ETHERTYPE_IPV6):
 341                 parser->ip6 = ptr;
 342                 if (__predict_false(mlen < sizeof(struct ip6_hdr)))
 343                         return (NULL);
 344                 /* Ensure the packet has valid src/dst addrs */
 345                 if (__predict_false(IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_src) ||
 346                         IN6_IS_ADDR_UNSPECIFIED(&parser->ip6->ip6_dst)))
 347                         return (NULL);
 348                 ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
 349                 if (update_data) {
 350                         parser->data.s_addr.v6 = parser->ip6->ip6_src;
 351                         parser->data.d_addr.v6 = parser->ip6->ip6_dst;
 352                 }
 353                 mlen -= sizeof(struct ip6_hdr);
 354                 switch (parser->ip6->ip6_nxt) {
 355                 case IPPROTO_UDP:
 356                         if (__predict_false(mlen < sizeof(struct udphdr)))
 357                                 return (NULL);
 358                         parser->udp = ptr;
 359                         if (update_data) {
 360                                 parser->data.lro_type = LRO_TYPE_IPV6_UDP;
 361                                 parser->data.s_port = parser->udp->uh_sport;
 362                                 parser->data.d_port = parser->udp->uh_dport;
 363                         } else {
 364                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
 365                         }
 366                         ptr = (uint8_t *)ptr + sizeof(*parser->udp);
 367                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
 368                         return (ptr);
 369                 case IPPROTO_TCP:
 370                         if (__predict_false(mlen < sizeof(struct tcphdr)))
 371                                 return (NULL);
 372                         parser->tcp = ptr;
 373                         if (update_data) {
 374                                 parser->data.lro_type = LRO_TYPE_IPV6_TCP;
 375                                 parser->data.s_port = parser->tcp->th_sport;
 376                                 parser->data.d_port = parser->tcp->th_dport;
 377                         } else {
 378                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
 379                         }
 380                         if (__predict_false(mlen < (parser->tcp->th_off << 2)))
 381                                 return (NULL);
 382                         ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
 383                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
 384                         return (ptr);
 385                 default:
 386                         break;
 387                 }
 388                 break;
 389 #endif
 390         default:
 391                 break;
 392         }
 393         /* Invalid packet - cannot parse */
 394         return (NULL);
 395 }
 396
 397 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
 398     CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
 399
 400 static inline struct lro_parser *
 401 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
 402 {
 403         void *data_ptr;
 404
 405         /* Try to parse outer headers first. */
 406         data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
 407         if (data_ptr == NULL || po->total_hdr_len > m->m_len)
 408                 return (NULL);
 409
 410         if (update_data) {
 411                 /* Store VLAN ID, if any. */
 412                 if (__predict_false(m->m_flags & M_VLANTAG)) {
 413                         po->data.vlan_id =
 414                             htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
 415                 }
 416                 /* Store decrypted flag, if any. */
 417                 if (__predict_false((m->m_pkthdr.csum_flags &
 418                     CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
 419                         po->data.lro_flags |= LRO_FLAG_DECRYPTED;
 420         }
 421
 422         switch (po->data.lro_type) {
 423         case LRO_TYPE_IPV4_UDP:
 424         case LRO_TYPE_IPV6_UDP:
 425                 /* Check for VXLAN headers. */
 426                 if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
 427                         break;
 428
 429                 /* Try to parse inner headers. */
 430                 data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
 431                                                     (m->m_len - ((caddr_t)data_ptr - m->m_data)));
 432                 if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
 433                         break;
 434
 435                 /* Verify supported header types. */
 436                 switch (pi->data.lro_type) {
 437                 case LRO_TYPE_IPV4_TCP:
 438                 case LRO_TYPE_IPV6_TCP:
 439                         return (pi);
 440                 default:
 441                         break;
 442                 }
 443                 break;
 444         case LRO_TYPE_IPV4_TCP:
 445         case LRO_TYPE_IPV6_TCP:
 446                 if (update_data)
 447                         memset(pi, 0, sizeof(*pi));
 448                 return (po);
 449         default:
 450                 break;
 451         }
 452         return (NULL);
 453 }
 454
 455 static inline int
 456 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
 457 {
 458         int len;
 459
 460         switch (po->data.lro_type) {
 461 #ifdef INET
 462         case LRO_TYPE_IPV4_TCP:
 463                 len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
 464                     ntohs(po->ip4->ip_len);
 465                 break;
 466 #endif
 467 #ifdef INET6
 468         case LRO_TYPE_IPV6_TCP:
 469                 len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
 470                     ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
 471                 break;
 472 #endif
 473         default:
 474                 return (TCP_LRO_CANNOT);
 475         }
 476
 477         /*
 478          * If the frame is padded beyond the end of the IP packet,
 479          * then trim the extra bytes off:
 480          */
 481         if (__predict_true(m->m_pkthdr.len == len)) {
 482                 return (0);
 483         } else if (m->m_pkthdr.len > len) {
 484                 m_adj(m, len - m->m_pkthdr.len);
 485                 return (0);
 486         }
 487         return (TCP_LRO_CANNOT);
 488 }
 489
 490 static struct tcphdr *
 491 tcp_lro_get_th(struct mbuf *m)
 492 {
 493         return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off));
 494 }
 495
 496 static void
 497 lro_free_mbuf_chain(struct mbuf *m)
 498 {
 499         struct mbuf *save;
 500
 501         while (m) {
 502                 save = m->m_nextpkt;
 503                 m->m_nextpkt = NULL;
 504                 m_freem(m);
 505                 m = save;
 506         }
 507 }
 508
 509 void
 510 tcp_lro_free(struct lro_ctrl *lc)
 511 {
 512         struct lro_entry *le;
 513         unsigned x;
 514
 515         /* reset LRO free list */
 516         LIST_INIT(&lc->lro_free);
 517
 518         /* free active mbufs, if any */
 519         while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
 520                 tcp_lro_active_remove(le);
 521                 lro_free_mbuf_chain(le->m_head);
 522         }
 523
 524         /* free hash table */
 525         free(lc->lro_hash, M_LRO);
 526         lc->lro_hash = NULL;
 527         lc->lro_hashsz = 0;
 528
 529         /* free mbuf array, if any */
 530         for (x = 0; x != lc->lro_mbuf_count; x++)
 531                 m_freem(lc->lro_mbuf_data[x].mb);
 532         lc->lro_mbuf_count = 0;
 533
 534         /* free allocated memory, if any */
 535         free(lc->lro_mbuf_data, M_LRO);
 536         lc->lro_mbuf_data = NULL;
 537 }
 538
 539 static uint16_t
 540 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
 541 {
 542         const uint16_t *ptr;
 543         uint32_t csum;
 544         uint16_t len;
 545
 546         csum = -th->th_sum;     /* exclude checksum field */
 547         len = th->th_off;
 548         ptr = (const uint16_t *)th;
 549         while (len--) {
 550                 csum += *ptr;
 551                 ptr++;
 552                 csum += *ptr;
 553                 ptr++;
 554         }
 555         while (csum > 0xffff)
 556                 csum = (csum >> 16) + (csum & 0xffff);
 557
 558         return (csum);
 559 }
 560
 561 static uint16_t
 562 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
 563 {
 564         uint32_t c;
 565         uint16_t cs;
 566
 567         c = tcp_csum;
 568
 569         switch (pa->data.lro_type) {
 570 #ifdef INET6
 571         case LRO_TYPE_IPV6_TCP:
 572                 /* Compute full pseudo IPv6 header checksum. */
 573                 cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
 574                 break;
 575 #endif
 576 #ifdef INET
 577         case LRO_TYPE_IPV4_TCP:
 578                 /* Compute full pseudo IPv4 header checsum. */
 579                 cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
 580                 cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
 581                 break;
 582 #endif
 583         default:
 584                 cs = 0;         /* Keep compiler happy. */
 585                 break;
 586         }
 587
 588         /* Complement checksum. */
 589         cs = ~cs;
 590         c += cs;
 591
 592         /* Remove TCP header checksum. */
 593         cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
 594         c += cs;
 595
 596         /* Compute checksum remainder. */
 597         while (c > 0xffff)
 598                 c = (c >> 16) + (c & 0xffff);
 599
 600         return (c);
 601 }
 602
 603 static void
 604 tcp_lro_rx_done(struct lro_ctrl *lc)
 605 {
 606         struct lro_entry *le;
 607
 608         while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
 609                 tcp_lro_active_remove(le);
 610                 tcp_lro_flush(lc, le);
 611         }
 612 }
 613
 614 static void
 615 tcp_lro_flush_active(struct lro_ctrl *lc)
 616 {
 617         struct lro_entry *le;
 618
 619         /*
 620          * Walk through the list of le entries, and
 621          * any one that does have packets flush. This
 622          * is called because we have an inbound packet
 623          * (e.g. SYN) that has to have all others flushed
 624          * in front of it. Note we have to do the remove
 625          * because tcp_lro_flush() assumes that the entry
 626          * is being freed. This is ok it will just get
 627          * reallocated again like it was new.
 628          */
 629         LIST_FOREACH(le, &lc->lro_active, next) {
 630                 if (le->m_head != NULL) {
 631                         tcp_lro_active_remove(le);
 632                         tcp_lro_flush(lc, le);
 633                 }
 634         }
 635 }
 636
 637 void
 638 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
 639 {
 640         struct lro_entry *le, *le_tmp;
 641         uint64_t now, tov;
 642         struct bintime bt;
 643
 644         NET_EPOCH_ASSERT();
 645         if (LIST_EMPTY(&lc->lro_active))
 646                 return;
 647
 648         /* get timeout time and current time in ns */
 649         binuptime(&bt);
 650         now = bintime2ns(&bt);
 651         tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
 652         LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
 653                 if (now >= (bintime2ns(&le->alloc_time) + tov)) {
 654                         tcp_lro_active_remove(le);
 655                         tcp_lro_flush(lc, le);
 656                 }
 657         }
 658 }
 659
 660 #ifdef INET
 661 static int
 662 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
 663 {
 664         uint16_t csum;
 665
 666         /* Legacy IP has a header checksum that needs to be correct. */
 667         if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
 668                 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
 669                         lc->lro_bad_csum++;
 670                         return (TCP_LRO_CANNOT);
 671                 }
 672         } else {
 673                 csum = in_cksum_hdr(ip4);
 674                 if (__predict_false(csum != 0)) {
 675                         lc->lro_bad_csum++;
 676                         return (TCP_LRO_CANNOT);
 677                 }
 678         }
 679         return (0);
 680 }
 681 #endif
 682
 683 #ifdef TCPHPTS
 684 static void
 685 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc,
 686     const struct lro_entry *le, const struct mbuf *m,
 687     int frm, int32_t tcp_data_len, uint32_t th_seq,
 688     uint32_t th_ack, uint16_t th_win)
 689 {
 690         if (tcp_bblogging_on(tp)) {
 691                 union tcp_log_stackspecific log;
 692                 struct timeval tv, btv;
 693                 uint32_t cts;
 694
 695                 cts = tcp_get_usecs(&tv);
 696                 memset(&log, 0, sizeof(union tcp_log_stackspecific));
 697                 log.u_bbr.flex8 = frm;
 698                 log.u_bbr.flex1 = tcp_data_len;
 699                 if (m)
 700                         log.u_bbr.flex2 = m->m_pkthdr.len;
 701                 else
 702                         log.u_bbr.flex2 = 0;
 703                 if (le->m_head) {
 704                         log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
 705                         log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
 706                         log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
 707                         log.u_bbr.delRate = le->m_head->m_flags;
 708                         log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
 709                 }
 710                 log.u_bbr.inflight = th_seq;
 711                 log.u_bbr.delivered = th_ack;
 712                 log.u_bbr.timeStamp = cts;
 713                 log.u_bbr.epoch = le->next_seq;
 714                 log.u_bbr.lt_epoch = le->ack_seq;
 715                 log.u_bbr.pacing_gain = th_win;
 716                 log.u_bbr.cwnd_gain = le->window;
 717                 log.u_bbr.lost = curcpu;
 718                 log.u_bbr.cur_del_rate = (uintptr_t)m;
 719                 log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
 720                 bintime2timeval(&lc->lro_last_queue_time, &btv);
 721                 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv);
 722                 log.u_bbr.flex7 = le->compressed;
 723                 log.u_bbr.pacing_gain = le->uncompressed;
 724                 if (in_epoch(net_epoch_preempt))
 725                         log.u_bbr.inhpts = 1;
 726                 else
 727                         log.u_bbr.inhpts = 0;
 728                 TCP_LOG_EVENTP(tp, NULL, &tptosocket(tp)->so_rcv,
 729                     &tptosocket(tp)->so_snd,
 730                     TCP_LOG_LRO, 0, 0, &log, false, &tv);
 731         }
 732 }
 733 #endif
 734
 735 static inline void
 736 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
 737 {
 738         uint32_t csum;
 739
 740         csum = 0xffff - *ptr + value;
 741         while (csum > 0xffff)
 742                 csum = (csum >> 16) + (csum & 0xffff);
 743         *ptr = value;
 744         *psum = csum;
 745 }
 746
 747 static uint16_t
 748 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
 749     uint16_t payload_len, uint16_t delta_sum)
 750 {
 751         uint32_t csum;
 752         uint16_t tlen;
 753         uint16_t temp[5] = {};
 754
 755         switch (pa->data.lro_type) {
 756         case LRO_TYPE_IPV4_TCP:
 757                 /* Compute new IPv4 length. */
 758                 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
 759                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
 760
 761                 /* Subtract delta from current IPv4 checksum. */
 762                 csum = pa->ip4->ip_sum + 0xffff - temp[0];
 763                 while (csum > 0xffff)
 764                         csum = (csum >> 16) + (csum & 0xffff);
 765                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
 766                 goto update_tcp_header;
 767
 768         case LRO_TYPE_IPV6_TCP:
 769                 /* Compute new IPv6 length. */
 770                 tlen = (pa->tcp->th_off << 2) + payload_len;
 771                 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
 772                 goto update_tcp_header;
 773
 774         case LRO_TYPE_IPV4_UDP:
 775                 /* Compute new IPv4 length. */
 776                 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
 777                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
 778
 779                 /* Subtract delta from current IPv4 checksum. */
 780                 csum = pa->ip4->ip_sum + 0xffff - temp[0];
 781                 while (csum > 0xffff)
 782                         csum = (csum >> 16) + (csum & 0xffff);
 783                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
 784                 goto update_udp_header;
 785
 786         case LRO_TYPE_IPV6_UDP:
 787                 /* Compute new IPv6 length. */
 788                 tlen = sizeof(*pa->udp) + payload_len;
 789                 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
 790                 goto update_udp_header;
 791
 792         default:
 793                 return (0);
 794         }
 795
 796 update_tcp_header:
 797         /* Compute current TCP header checksum. */
 798         temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
 799
 800         /* Incorporate the latest ACK into the TCP header. */
 801         pa->tcp->th_ack = le->ack_seq;
 802         pa->tcp->th_win = le->window;
 803
 804         /* Incorporate latest timestamp into the TCP header. */
 805         if (le->timestamp != 0) {
 806                 uint32_t *ts_ptr;
 807
 808                 ts_ptr = (uint32_t *)(pa->tcp + 1);
 809                 ts_ptr[1] = htonl(le->tsval);
 810                 ts_ptr[2] = le->tsecr;
 811         }
 812
 813         /* Compute new TCP header checksum. */
 814         temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
 815
 816         /* Compute new TCP checksum. */
 817         csum = pa->tcp->th_sum + 0xffff - delta_sum +
 818             0xffff - temp[0] + 0xffff - temp[3] + temp[2];
 819         while (csum > 0xffff)
 820                 csum = (csum >> 16) + (csum & 0xffff);
 821
 822         /* Assign new TCP checksum. */
 823         tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
 824
 825         /* Compute all modififications affecting next checksum. */
 826         csum = temp[0] + temp[1] + 0xffff - temp[2] +
 827             temp[3] + temp[4] + delta_sum;
 828         while (csum > 0xffff)
 829                 csum = (csum >> 16) + (csum & 0xffff);
 830
 831         /* Return delta checksum to next stage, if any. */
 832         return (csum);
 833
 834 update_udp_header:
 835         tlen = sizeof(*pa->udp) + payload_len;
 836         /* Assign new UDP length and compute checksum delta. */
 837         tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
 838
 839         /* Check if there is a UDP checksum. */
 840         if (__predict_false(pa->udp->uh_sum != 0)) {
 841                 /* Compute new UDP checksum. */
 842                 csum = pa->udp->uh_sum + 0xffff - delta_sum +
 843                     0xffff - temp[0] + 0xffff - temp[2];
 844                 while (csum > 0xffff)
 845                         csum = (csum >> 16) + (csum & 0xffff);
 846                 /* Assign new UDP checksum. */
 847                 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
 848         }
 849
 850         /* Compute all modififications affecting next checksum. */
 851         csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
 852         while (csum > 0xffff)
 853                 csum = (csum >> 16) + (csum & 0xffff);
 854
 855         /* Return delta checksum to next stage, if any. */
 856         return (csum);
 857 }
 858
 859 static void
 860 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
 861 {
 862         /* Check if we need to recompute any checksums. */
 863         if (le->needs_merge) {
 864                 uint16_t csum;
 865
 866                 switch (le->inner.data.lro_type) {
 867                 case LRO_TYPE_IPV4_TCP:
 868                         csum = tcp_lro_update_checksum(&le->inner, le,
 869                             le->m_head->m_pkthdr.lro_tcp_d_len,
 870                             le->m_head->m_pkthdr.lro_tcp_d_csum);
 871                         csum = tcp_lro_update_checksum(&le->outer, NULL,
 872                             le->m_head->m_pkthdr.lro_tcp_d_len +
 873                             le->inner.total_hdr_len, csum);
 874                         le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 875                             CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
 876                         le->m_head->m_pkthdr.csum_data = 0xffff;
 877                         if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
 878                                 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
 879                         break;
 880                 case LRO_TYPE_IPV6_TCP:
 881                         csum = tcp_lro_update_checksum(&le->inner, le,
 882                             le->m_head->m_pkthdr.lro_tcp_d_len,
 883                             le->m_head->m_pkthdr.lro_tcp_d_csum);
 884                         csum = tcp_lro_update_checksum(&le->outer, NULL,
 885                             le->m_head->m_pkthdr.lro_tcp_d_len +
 886                             le->inner.total_hdr_len, csum);
 887                         le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 888                             CSUM_PSEUDO_HDR;
 889                         le->m_head->m_pkthdr.csum_data = 0xffff;
 890                         if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
 891                                 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
 892                         break;
 893                 case LRO_TYPE_NONE:
 894                         switch (le->outer.data.lro_type) {
 895                         case LRO_TYPE_IPV4_TCP:
 896                                 csum = tcp_lro_update_checksum(&le->outer, le,
 897                                     le->m_head->m_pkthdr.lro_tcp_d_len,
 898                                     le->m_head->m_pkthdr.lro_tcp_d_csum);
 899                                 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 900                                     CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
 901                                 le->m_head->m_pkthdr.csum_data = 0xffff;
 902                                 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
 903                                         le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
 904                                 break;
 905                         case LRO_TYPE_IPV6_TCP:
 906                                 csum = tcp_lro_update_checksum(&le->outer, le,
 907                                     le->m_head->m_pkthdr.lro_tcp_d_len,
 908                                     le->m_head->m_pkthdr.lro_tcp_d_csum);
 909                                 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 910                                     CSUM_PSEUDO_HDR;
 911                                 le->m_head->m_pkthdr.csum_data = 0xffff;
 912                                 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
 913                                         le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
 914                                 break;
 915                         default:
 916                                 break;
 917                         }
 918                         break;
 919                 default:
 920                         break;
 921                 }
 922         }
 923
 924         /*
 925          * Break any chain, this is not set to NULL on the singleton
 926          * case m_nextpkt points to m_head. Other case set them
 927          * m_nextpkt to NULL in push_and_replace.
 928          */
 929         le->m_head->m_nextpkt = NULL;
 930         lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
 931         (*lc->ifp->if_input)(lc->ifp, le->m_head);
 932 }
 933
 934 static void
 935 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
 936     struct mbuf *m, struct tcphdr *th)
 937 {
 938         uint32_t *ts_ptr;
 939         uint16_t tcp_data_len;
 940         uint16_t tcp_opt_len;
 941
 942         ts_ptr = (uint32_t *)(th + 1);
 943         tcp_opt_len = (th->th_off << 2);
 944         tcp_opt_len -= sizeof(*th);
 945
 946         /* Check if there is a timestamp option. */
 947         if (tcp_opt_len == 0 ||
 948             __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
 949             *ts_ptr != TCP_LRO_TS_OPTION)) {
 950                 /* We failed to find the timestamp option. */
 951                 le->timestamp = 0;
 952         } else {
 953                 le->timestamp = 1;
 954                 le->tsval = ntohl(*(ts_ptr + 1));
 955                 le->tsecr = *(ts_ptr + 2);
 956         }
 957
 958         tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
 959
 960         /* Pull out TCP sequence numbers and window size. */
 961         le->next_seq = ntohl(th->th_seq) + tcp_data_len;
 962         le->ack_seq = th->th_ack;
 963         le->window = th->th_win;
 964         le->flags = tcp_get_flags(th);
 965         le->needs_merge = 0;
 966
 967         /* Setup new data pointers. */
 968         le->m_head = m;
 969         le->m_tail = m_last(m);
 970 }
 971
 972 static void
 973 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
 974 {
 975         struct lro_parser *pa;
 976
 977         /*
 978          * Push up the stack of the current entry
 979          * and replace it with "m".
 980          */
 981         struct mbuf *msave;
 982
 983         /* Grab off the next and save it */
 984         msave = le->m_head->m_nextpkt;
 985         le->m_head->m_nextpkt = NULL;
 986
 987         /* Now push out the old entry */
 988         tcp_flush_out_entry(lc, le);
 989
 990         /* Re-parse new header, should not fail. */
 991         pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
 992         KASSERT(pa != NULL,
 993             ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
 994
 995         /*
 996          * Now to replace the data properly in the entry
 997          * we have to reset the TCP header and
 998          * other fields.
 999          */
1000         tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
1001
1002         /* Restore the next list */
1003         m->m_nextpkt = msave;
1004 }
1005
1006 static void
1007 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
1008 {
1009         struct mbuf *m;
1010         uint32_t csum;
1011
1012         m = le->m_head;
1013         if (m->m_pkthdr.lro_nsegs == 1) {
1014                 /* Compute relative checksum. */
1015                 csum = p->m_pkthdr.lro_tcp_d_csum;
1016         } else {
1017                 /* Merge TCP data checksums. */
1018                 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
1019                     (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
1020                 while (csum > 0xffff)
1021                         csum = (csum >> 16) + (csum & 0xffff);
1022         }
1023
1024         /* Update various counters. */
1025         m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
1026         m->m_pkthdr.lro_tcp_d_csum = csum;
1027         m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
1028         m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
1029         le->needs_merge = 1;
1030 }
1031
1032 static void
1033 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
1034 {
1035         /*
1036          * Walk through the mbuf chain we
1037          * have on tap and compress/condense
1038          * as required.
1039          */
1040         uint32_t *ts_ptr;
1041         struct mbuf *m;
1042         struct tcphdr *th;
1043         uint32_t tcp_data_len_total;
1044         uint32_t tcp_data_seg_total;
1045         uint16_t tcp_data_len;
1046         uint16_t tcp_opt_len;
1047
1048         /*
1049          * First we must check the lead (m_head)
1050          * we must make sure that it is *not*
1051          * something that should be sent up
1052          * right away (sack etc).
1053          */
1054 again:
1055         m = le->m_head->m_nextpkt;
1056         if (m == NULL) {
1057                 /* Just one left. */
1058                 return;
1059         }
1060
1061         th = tcp_lro_get_th(m);
1062         tcp_opt_len = (th->th_off << 2);
1063         tcp_opt_len -= sizeof(*th);
1064         ts_ptr = (uint32_t *)(th + 1);
1065
1066         if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1067             *ts_ptr != TCP_LRO_TS_OPTION)) {
1068                 /*
1069                  * Its not the timestamp. We can't
1070                  * use this guy as the head.
1071                  */
1072                 le->m_head->m_nextpkt = m->m_nextpkt;
1073                 tcp_push_and_replace(lc, le, m);
1074                 goto again;
1075         }
1076         if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1077                 /*
1078                  * Make sure that previously seen segments/ACKs are delivered
1079                  * before this segment, e.g. FIN.
1080                  */
1081                 le->m_head->m_nextpkt = m->m_nextpkt;
1082                 tcp_push_and_replace(lc, le, m);
1083                 goto again;
1084         }
1085         while((m = le->m_head->m_nextpkt) != NULL) {
1086                 /*
1087                  * condense m into le, first
1088                  * pull m out of the list.
1089                  */
1090                 le->m_head->m_nextpkt = m->m_nextpkt;
1091                 m->m_nextpkt = NULL;
1092                 /* Setup my data */
1093                 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1094                 th = tcp_lro_get_th(m);
1095                 ts_ptr = (uint32_t *)(th + 1);
1096                 tcp_opt_len = (th->th_off << 2);
1097                 tcp_opt_len -= sizeof(*th);
1098                 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1099                 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1100
1101                 if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1102                     tcp_data_len_total >= lc->lro_length_lim) {
1103                         /* Flush now if appending will result in overflow. */
1104                         tcp_push_and_replace(lc, le, m);
1105                         goto again;
1106                 }
1107                 if (tcp_opt_len != 0 &&
1108                     __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1109                     *ts_ptr != TCP_LRO_TS_OPTION)) {
1110                         /*
1111                          * Maybe a sack in the new one? We need to
1112                          * start all over after flushing the
1113                          * current le. We will go up to the beginning
1114                          * and flush it (calling the replace again possibly
1115                          * or just returning).
1116                          */
1117                         tcp_push_and_replace(lc, le, m);
1118                         goto again;
1119                 }
1120                 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1121                         tcp_push_and_replace(lc, le, m);
1122                         goto again;
1123                 }
1124                 if (tcp_opt_len != 0) {
1125                         uint32_t tsval = ntohl(*(ts_ptr + 1));
1126                         /* Make sure timestamp values are increasing. */
1127                         if (TSTMP_GT(le->tsval, tsval))  {
1128                                 tcp_push_and_replace(lc, le, m);
1129                                 goto again;
1130                         }
1131                         le->tsval = tsval;
1132                         le->tsecr = *(ts_ptr + 2);
1133                 }
1134                 /* Try to append the new segment. */
1135                 if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1136                                     ((tcp_get_flags(th) & TH_ACK) !=
1137                                       (le->flags & TH_ACK)) ||
1138                                     (tcp_data_len == 0 &&
1139                                      le->ack_seq == th->th_ack &&
1140                                      le->window == th->th_win))) {
1141                         /* Out of order packet, non-ACK + ACK or dup ACK. */
1142                         tcp_push_and_replace(lc, le, m);
1143                         goto again;
1144                 }
1145                 if (tcp_data_len != 0 ||
1146                     SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1147                         le->next_seq += tcp_data_len;
1148                         le->ack_seq = th->th_ack;
1149                         le->window = th->th_win;
1150                         le->needs_merge = 1;
1151                 } else if (th->th_ack == le->ack_seq) {
1152                         if (WIN_GT(th->th_win, le->window)) {
1153                                 le->window = th->th_win;
1154                                 le->needs_merge = 1;
1155                         }
1156                 }
1157
1158                 if (tcp_data_len == 0) {
1159                         m_freem(m);
1160                         continue;
1161                 }
1162
1163                 /* Merge TCP data checksum and length to head mbuf. */
1164                 tcp_lro_mbuf_append_pkthdr(le, m);
1165
1166                 /*
1167                  * Adjust the mbuf so that m_data points to the first byte of
1168                  * the ULP payload.  Adjust the mbuf to avoid complications and
1169                  * append new segment to existing mbuf chain.
1170                  */
1171                 m_adj(m, m->m_pkthdr.len - tcp_data_len);
1172                 m_demote_pkthdr(m);
1173                 le->m_tail->m_next = m;
1174                 le->m_tail = m_last(m);
1175         }
1176 }
1177
1178 #ifdef TCPHPTS
1179 static void
1180 tcp_queue_pkts(struct tcpcb *tp, struct lro_entry *le)
1181 {
1182
1183         INP_WLOCK_ASSERT(tptoinpcb(tp));
1184
1185         STAILQ_HEAD(, mbuf) q = { le->m_head,
1186             &STAILQ_NEXT(le->m_last_mbuf, m_stailqpkt) };
1187         STAILQ_CONCAT(&tp->t_inqueue, &q);
1188         le->m_head = NULL;
1189         le->m_last_mbuf = NULL;
1190 }
1191
1192 static bool
1193 tcp_lro_check_wake_status(struct tcpcb *tp)
1194 {
1195
1196         if (tp->t_fb->tfb_early_wake_check != NULL)
1197                 return ((tp->t_fb->tfb_early_wake_check)(tp));
1198         return (false);
1199 }
1200
1201 static struct mbuf *
1202 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le,
1203     struct tcpcb *tp, int32_t *new_m, bool can_append_old_cmp)
1204 {
1205         struct mbuf *m;
1206
1207         /* Look at the last mbuf if any in queue */
1208         if (can_append_old_cmp) {
1209                 m = STAILQ_LAST(&tp->t_inqueue, mbuf, m_stailqpkt);
1210                 if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
1211                         if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
1212                                 tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
1213                                 *new_m = 0;
1214                                 counter_u64_add(tcp_extra_mbuf, 1);
1215                                 return (m);
1216                         } else {
1217                                 /* Mark we ran out of space */
1218                                 tp->t_flags2 |= TF2_MBUF_L_ACKS;
1219                         }
1220                 }
1221         }
1222         /* Decide mbuf size. */
1223         tcp_lro_log(tp, lc, le, NULL, 21, 0, 0, 0, 0);
1224         if (tp->t_flags2 & TF2_MBUF_L_ACKS)
1225                 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR);
1226         else
1227                 m = m_gethdr(M_NOWAIT, MT_DATA);
1228
1229         if (__predict_false(m == NULL)) {
1230                 counter_u64_add(tcp_would_have_but, 1);
1231                 return (NULL);
1232         }
1233         counter_u64_add(tcp_comp_total, 1);
1234         m->m_pkthdr.rcvif = lc->ifp;
1235         m->m_flags |= M_ACKCMP;
1236         *new_m = 1;
1237         return (m);
1238 }
1239
1240 static struct tcpcb *
1241 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa)
1242 {
1243         struct inpcb *inp;
1244
1245         switch (pa->data.lro_type) {
1246 #ifdef INET6
1247         case LRO_TYPE_IPV6_TCP:
1248                 inp = in6_pcblookup(&V_tcbinfo,
1249                     &pa->data.s_addr.v6,
1250                     pa->data.s_port,
1251                     &pa->data.d_addr.v6,
1252                     pa->data.d_port,
1253                     INPLOOKUP_WLOCKPCB,
1254                     ifp);
1255                 break;
1256 #endif
1257 #ifdef INET
1258         case LRO_TYPE_IPV4_TCP:
1259                 inp = in_pcblookup(&V_tcbinfo,
1260                     pa->data.s_addr.v4,
1261                     pa->data.s_port,
1262                     pa->data.d_addr.v4,
1263                     pa->data.d_port,
1264                     INPLOOKUP_WLOCKPCB,
1265                     ifp);
1266                 break;
1267 #endif
1268         default:
1269                 return (NULL);
1270         }
1271
1272         return (intotcpcb(inp));
1273 }
1274
1275 static inline bool
1276 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts)
1277 {
1278         /*
1279          * This function returns two bits of valuable information.
1280          * a) Is what is present capable of being ack-compressed,
1281          *    we can ack-compress if there is no options or just
1282          *    a timestamp option, and of course the th_flags must
1283          *    be correct as well.
1284          * b) Our other options present such as SACK. This is
1285          *    used to determine if we want to wakeup or not.
1286          */
1287         bool ret = true;
1288
1289         switch (th->th_off << 2) {
1290         case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
1291                 *ppts = (uint32_t *)(th + 1);
1292                 /* Check if we have only one timestamp option. */
1293                 if (**ppts == TCP_LRO_TS_OPTION)
1294                         *other_opts = false;
1295                 else {
1296                         *other_opts = true;
1297                         ret = false;
1298                 }
1299                 break;
1300         case (sizeof(*th)):
1301                 /* No options. */
1302                 *ppts = NULL;
1303                 *other_opts = false;
1304                 break;
1305         default:
1306                 *ppts = NULL;
1307                 *other_opts = true;
1308                 ret = false;
1309                 break;
1310         }
1311         /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
1312         if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0)
1313                 ret = false;
1314         /* If it has data on it we cannot compress it */
1315         if (m->m_pkthdr.lro_tcp_d_len)
1316                 ret = false;
1317
1318         /* ACK flag must be set. */
1319         if (!(tcp_get_flags(th) & TH_ACK))
1320                 ret = false;
1321         return (ret);
1322 }
1323
1324 static int
1325 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le)
1326 {
1327         struct tcpcb *tp;
1328         struct mbuf **pp, *cmp, *mv_to;
1329         struct ifnet *lagg_ifp;
1330         bool bpf_req, lagg_bpf_req, should_wake, can_append_old_cmp;
1331
1332         /* Check if packet doesn't belongs to our network interface. */
1333         if ((tcplro_stacks_wanting_mbufq == 0) ||
1334             (le->outer.data.vlan_id != 0) ||
1335             (le->inner.data.lro_type != LRO_TYPE_NONE))
1336                 return (TCP_LRO_CANNOT);
1337
1338 #ifdef INET6
1339         /*
1340          * Be proactive about unspecified IPv6 address in source. As
1341          * we use all-zero to indicate unbounded/unconnected pcb,
1342          * unspecified IPv6 address can be used to confuse us.
1343          *
1344          * Note that packets with unspecified IPv6 destination is
1345          * already dropped in ip6_input.
1346          */
1347         if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
1348             IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
1349                 return (TCP_LRO_CANNOT);
1350
1351         if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
1352             IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
1353                 return (TCP_LRO_CANNOT);
1354 #endif
1355         /* Lookup inp, if any.  Returns locked TCP inpcb. */
1356         tp = tcp_lro_lookup(lc->ifp,
1357             (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
1358         if (tp == NULL)
1359                 return (TCP_LRO_CANNOT);
1360
1361         counter_u64_add(tcp_inp_lro_locks_taken, 1);
1362
1363         /* Check if the inp is dead, Jim. */
1364         if (tp->t_state == TCPS_TIME_WAIT) {
1365                 INP_WUNLOCK(tptoinpcb(tp));
1366                 return (TCP_LRO_CANNOT);
1367         }
1368         if (tp->t_lro_cpu == HPTS_CPU_NONE && lc->lro_cpu_is_set == 1)
1369                 tp->t_lro_cpu = lc->lro_last_cpu;
1370         /* Check if the transport doesn't support the needed optimizations. */
1371         if ((tp->t_flags2 & (TF2_SUPPORTS_MBUFQ | TF2_MBUF_ACKCMP)) == 0) {
1372                 INP_WUNLOCK(tptoinpcb(tp));
1373                 return (TCP_LRO_CANNOT);
1374         }
1375
1376         if (tp->t_flags2 & TF2_MBUF_QUEUE_READY)
1377                 should_wake = false;
1378         else
1379                 should_wake = true;
1380         /* Check if packets should be tapped to BPF. */
1381         bpf_req = bpf_peers_present(lc->ifp->if_bpf);
1382         lagg_bpf_req = false;
1383         lagg_ifp = NULL;
1384         if (lc->ifp->if_type == IFT_IEEE8023ADLAG ||
1385             lc->ifp->if_type == IFT_INFINIBANDLAG) {
1386                 struct lagg_port *lp = lc->ifp->if_lagg;
1387                 struct lagg_softc *sc = lp->lp_softc;
1388
1389                 lagg_ifp = sc->sc_ifp;
1390                 if (lagg_ifp != NULL)
1391                         lagg_bpf_req = bpf_peers_present(lagg_ifp->if_bpf);
1392         }
1393
1394         /* Strip and compress all the incoming packets. */
1395         can_append_old_cmp = true;
1396         cmp = NULL;
1397         for (pp = &le->m_head; *pp != NULL; ) {
1398                 mv_to = NULL;
1399                 if (do_bpf_strip_and_compress(tp, lc, le, pp,
1400                         &cmp, &mv_to, &should_wake, bpf_req,
1401                         lagg_bpf_req, lagg_ifp, can_append_old_cmp) == false) {
1402                         /* Advance to next mbuf. */
1403                         pp = &(*pp)->m_nextpkt;
1404                         /*
1405                          * Once we have appended we can't look in the pending
1406                          * inbound packets for a compressed ack to append to.
1407                          */
1408                         can_append_old_cmp = false;
1409                         /*
1410                          * Once we append we also need to stop adding to any
1411                          * compressed ack we were remembering. A new cmp
1412                          * ack will be required.
1413                          */
1414                         cmp = NULL;
1415                         tcp_lro_log(tp, lc, le, NULL, 25, 0, 0, 0, 0);
1416                 } else if (mv_to != NULL) {
1417                         /* We are asked to move pp up */
1418                         pp = &mv_to->m_nextpkt;
1419                         tcp_lro_log(tp, lc, le, NULL, 24, 0, 0, 0, 0);
1420                 } else
1421                         tcp_lro_log(tp, lc, le, NULL, 26, 0, 0, 0, 0);
1422         }
1423         /* Update "m_last_mbuf", if any. */
1424         if (pp == &le->m_head)
1425                 le->m_last_mbuf = *pp;
1426         else
1427                 le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt);
1428
1429         /* Check if any data mbufs left. */
1430         if (le->m_head != NULL) {
1431                 counter_u64_add(tcp_inp_lro_direct_queue, 1);
1432                 tcp_lro_log(tp, lc, le, NULL, 22, 1, tp->t_flags2, 0, 1);
1433                 tcp_queue_pkts(tp, le);
1434         }
1435         if (should_wake) {
1436                 /* Wakeup */
1437                 counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
1438                 if ((*tp->t_fb->tfb_do_queued_segments)(tp, 0))
1439                         /* TCP cb gone and unlocked. */
1440                         return (0);
1441         }
1442         INP_WUNLOCK(tptoinpcb(tp));
1443
1444         return (0);     /* Success. */
1445 }
1446 #endif
1447
1448 void
1449 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1450 {
1451         /* Only optimise if there are multiple packets waiting. */
1452 #ifdef TCPHPTS
1453         int error;
1454 #endif
1455
1456         NET_EPOCH_ASSERT();
1457 #ifdef TCPHPTS
1458         CURVNET_SET(lc->ifp->if_vnet);
1459         error = tcp_lro_flush_tcphpts(lc, le);
1460         CURVNET_RESTORE();
1461         if (error != 0) {
1462 #endif
1463                 tcp_lro_condense(lc, le);
1464                 tcp_flush_out_entry(lc, le);
1465 #ifdef TCPHPTS
1466         }
1467 #endif
1468         lc->lro_flushed++;
1469         bzero(le, sizeof(*le));
1470         LIST_INSERT_HEAD(&lc->lro_free, le, next);
1471 }
1472
1473 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1474
1475 /*
1476  * The tcp_lro_sort() routine is comparable to qsort(), except it has
1477  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1478  * number of elements to sort and 64 is the number of sequence bits
1479  * available. The algorithm is bit-slicing the 64-bit sequence number,
1480  * sorting one bit at a time from the most significant bit until the
1481  * least significant one, skipping the constant bits. This is
1482  * typically called a radix sort.
1483  */
1484 static void
1485 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1486 {
1487         struct lro_mbuf_sort temp;
1488         uint64_t ones;
1489         uint64_t zeros;
1490         uint32_t x;
1491         uint32_t y;
1492
1493 repeat:
1494         /* for small arrays insertion sort is faster */
1495         if (size <= 12) {
1496                 for (x = 1; x < size; x++) {
1497                         temp = parray[x];
1498                         for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1499                                 parray[y] = parray[y - 1];
1500                         parray[y] = temp;
1501                 }
1502                 return;
1503         }
1504
1505         /* compute sequence bits which are constant */
1506         ones = 0;
1507         zeros = 0;
1508         for (x = 0; x != size; x++) {
1509                 ones |= parray[x].seq;
1510                 zeros |= ~parray[x].seq;
1511         }
1512
1513         /* compute bits which are not constant into "ones" */
1514         ones &= zeros;
1515         if (ones == 0)
1516                 return;
1517
1518         /* pick the most significant bit which is not constant */
1519         ones = tcp_lro_msb_64(ones);
1520
1521         /*
1522          * Move entries having cleared sequence bits to the beginning
1523          * of the array:
1524          */
1525         for (x = y = 0; y != size; y++) {
1526                 /* skip set bits */
1527                 if (parray[y].seq & ones)
1528                         continue;
1529                 /* swap entries */
1530                 temp = parray[x];
1531                 parray[x] = parray[y];
1532                 parray[y] = temp;
1533                 x++;
1534         }
1535
1536         KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1537
1538         /* sort zeros */
1539         tcp_lro_sort(parray, x);
1540
1541         /* sort ones */
1542         parray += x;
1543         size -= x;
1544         goto repeat;
1545 }
1546
1547 void
1548 tcp_lro_flush_all(struct lro_ctrl *lc)
1549 {
1550         uint64_t seq;
1551         uint64_t nseq;
1552         unsigned x;
1553
1554         NET_EPOCH_ASSERT();
1555         /* check if no mbufs to flush */
1556         if (lc->lro_mbuf_count == 0)
1557                 goto done;
1558         if (lc->lro_cpu_is_set == 0) {
1559                 if (lc->lro_last_cpu == curcpu) {
1560                         lc->lro_cnt_of_same_cpu++;
1561                         /* Have we reached the threshold to declare a cpu? */
1562                         if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1563                                 lc->lro_cpu_is_set = 1;
1564                 } else {
1565                         lc->lro_last_cpu = curcpu;
1566                         lc->lro_cnt_of_same_cpu = 0;
1567                 }
1568         }
1569         CURVNET_SET(lc->ifp->if_vnet);
1570
1571         /* get current time */
1572         binuptime(&lc->lro_last_queue_time);
1573
1574         /* sort all mbufs according to stream */
1575         tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1576
1577         /* input data into LRO engine, stream by stream */
1578         seq = 0;
1579         for (x = 0; x != lc->lro_mbuf_count; x++) {
1580                 struct mbuf *mb;
1581
1582                 /* get mbuf */
1583                 mb = lc->lro_mbuf_data[x].mb;
1584
1585                 /* get sequence number, masking away the packet index */
1586                 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1587
1588                 /* check for new stream */
1589                 if (seq != nseq) {
1590                         seq = nseq;
1591
1592                         /* flush active streams */
1593                         tcp_lro_rx_done(lc);
1594                 }
1595
1596                 /* add packet to LRO engine */
1597                 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1598                         /* Flush anything we have acummulated */
1599                         tcp_lro_flush_active(lc);
1600                         /* input packet to network layer */
1601                         (*lc->ifp->if_input)(lc->ifp, mb);
1602                         lc->lro_queued++;
1603                         lc->lro_flushed++;
1604                 }
1605         }
1606         CURVNET_RESTORE();
1607 done:
1608         /* flush active streams */
1609         tcp_lro_rx_done(lc);
1610
1611 #ifdef TCPHPTS
1612         tcp_run_hpts();
1613 #endif
1614         lc->lro_mbuf_count = 0;
1615 }
1616
1617 #ifdef TCPHPTS
1618 static void
1619 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m,
1620     uint32_t *ts_ptr, uint16_t iptos)
1621 {
1622         /*
1623          * Given a TCP ACK, summarize it down into the small TCP ACK
1624          * entry.
1625          */
1626         ae->timestamp = m->m_pkthdr.rcv_tstmp;
1627         ae->flags = 0;
1628         if (m->m_flags & M_TSTMP_LRO)
1629                 ae->flags |= TSTMP_LRO;
1630         else if (m->m_flags & M_TSTMP)
1631                 ae->flags |= TSTMP_HDWR;
1632         ae->seq = ntohl(th->th_seq);
1633         ae->ack = ntohl(th->th_ack);
1634         ae->flags |= tcp_get_flags(th);
1635         if (ts_ptr != NULL) {
1636                 ae->ts_value = ntohl(ts_ptr[1]);
1637                 ae->ts_echo = ntohl(ts_ptr[2]);
1638                 ae->flags |= HAS_TSTMP;
1639         }
1640         ae->win = ntohs(th->th_win);
1641         ae->codepoint = iptos;
1642 }
1643
1644 /*
1645  * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
1646  * and strip all, but the IPv4/IPv6 header.
1647  */
1648 static bool
1649 do_bpf_strip_and_compress(struct tcpcb *tp, struct lro_ctrl *lc,
1650     struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to,
1651     bool *should_wake, bool bpf_req, bool lagg_bpf_req, struct ifnet *lagg_ifp, bool can_append_old_cmp)
1652 {
1653         union {
1654                 void *ptr;
1655                 struct ip *ip4;
1656                 struct ip6_hdr *ip6;
1657         } l3;
1658         struct mbuf *m;
1659         struct mbuf *nm;
1660         struct tcphdr *th;
1661         struct tcp_ackent *ack_ent;
1662         uint32_t *ts_ptr;
1663         int32_t n_mbuf;
1664         bool other_opts, can_compress;
1665         uint8_t lro_type;
1666         uint16_t iptos;
1667         int tcp_hdr_offset;
1668         int idx;
1669
1670         /* Get current mbuf. */
1671         m = *pp;
1672
1673         /* Let the BPF see the packet */
1674         if (__predict_false(bpf_req))
1675                 ETHER_BPF_MTAP(lc->ifp, m);
1676
1677         if (__predict_false(lagg_bpf_req))
1678                 ETHER_BPF_MTAP(lagg_ifp, m);
1679
1680         tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
1681         lro_type = le->inner.data.lro_type;
1682         switch (lro_type) {
1683         case LRO_TYPE_NONE:
1684                 lro_type = le->outer.data.lro_type;
1685                 switch (lro_type) {
1686                 case LRO_TYPE_IPV4_TCP:
1687                         tcp_hdr_offset -= sizeof(*le->outer.ip4);
1688                         m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1689                         break;
1690                 case LRO_TYPE_IPV6_TCP:
1691                         tcp_hdr_offset -= sizeof(*le->outer.ip6);
1692                         m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1693                         break;
1694                 default:
1695                         goto compressed;
1696                 }
1697                 break;
1698         case LRO_TYPE_IPV4_TCP:
1699                 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1700                 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1701                 break;
1702         case LRO_TYPE_IPV6_TCP:
1703                 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1704                 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1705                 break;
1706         default:
1707                 goto compressed;
1708         }
1709
1710         MPASS(tcp_hdr_offset >= 0);
1711
1712         m_adj(m, tcp_hdr_offset);
1713         m->m_flags |= M_LRO_EHDRSTRP;
1714         m->m_flags &= ~M_ACKCMP;
1715         m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;
1716
1717         th = tcp_lro_get_th(m);
1718
1719         th->th_sum = 0;         /* TCP checksum is valid. */
1720
1721         /* Check if ACK can be compressed */
1722         can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts);
1723
1724         /* Now lets look at the should wake states */
1725         if ((other_opts == true) &&
1726             ((tp->t_flags2 & TF2_DONT_SACK_QUEUE) == 0)) {
1727                 /*
1728                  * If there are other options (SACK?) and the
1729                  * tcp endpoint has not expressly told us it does
1730                  * not care about SACKS, then we should wake up.
1731                  */
1732                 *should_wake = true;
1733         } else if (*should_wake == false) {
1734                 /* Wakeup override check if we are false here  */
1735                 *should_wake = tcp_lro_check_wake_status(tp);
1736         }
1737         /* Is the ack compressable? */
1738         if (can_compress == false)
1739                 goto done;
1740         /* Does the TCP endpoint support ACK compression? */
1741         if ((tp->t_flags2 & TF2_MBUF_ACKCMP) == 0)
1742                 goto done;
1743
1744         /* Lets get the TOS/traffic class field */
1745         l3.ptr = mtod(m, void *);
1746         switch (lro_type) {
1747         case LRO_TYPE_IPV4_TCP:
1748                 iptos = l3.ip4->ip_tos;
1749                 break;
1750         case LRO_TYPE_IPV6_TCP:
1751                 iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
1752                 break;
1753         default:
1754                 iptos = 0;      /* Keep compiler happy. */
1755                 break;
1756         }
1757         /* Now lets get space if we don't have some already */
1758         if (*cmp == NULL) {
1759 new_one:
1760                 nm = tcp_lro_get_last_if_ackcmp(lc, le, tp, &n_mbuf,
1761                     can_append_old_cmp);
1762                 if (__predict_false(nm == NULL))
1763                         goto done;
1764                 *cmp = nm;
1765                 if (n_mbuf) {
1766                         /*
1767                          *  Link in the new cmp ack to our in-order place,
1768                          * first set our cmp ack's next to where we are.
1769                          */
1770                         nm->m_nextpkt = m;
1771                         (*pp) = nm;
1772                         /*
1773                          * Set it up so mv_to is advanced to our
1774                          * compressed ack. This way the caller can
1775                          * advance pp to the right place.
1776                          */
1777                         *mv_to = nm;
1778                         /*
1779                          * Advance it here locally as well.
1780                          */
1781                         pp = &nm->m_nextpkt;
1782                 }
1783         } else {
1784                 /* We have one already we are working on */
1785                 nm = *cmp;
1786                 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
1787                         /* We ran out of space */
1788                         tp->t_flags2 |= TF2_MBUF_L_ACKS;
1789                         goto new_one;
1790                 }
1791         }
1792         MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
1793         counter_u64_add(tcp_inp_lro_compressed, 1);
1794         le->compressed++;
1795         /* We can add in to the one on the tail */
1796         ack_ent = mtod(nm, struct tcp_ackent *);
1797         idx = (nm->m_len / sizeof(struct tcp_ackent));
1798         build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
1799
1800         /* Bump the size of both pkt-hdr and len */
1801         nm->m_len += sizeof(struct tcp_ackent);
1802         nm->m_pkthdr.len += sizeof(struct tcp_ackent);
1803 compressed:
1804         /* Advance to next mbuf before freeing. */
1805         *pp = m->m_nextpkt;
1806         m->m_nextpkt = NULL;
1807         m_freem(m);
1808         return (true);
1809 done:
1810         counter_u64_add(tcp_uncomp_total, 1);
1811         le->uncompressed++;
1812         return (false);
1813 }
1814 #endif
1815
1816 static struct lro_head *
1817 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1818 {
1819         u_long hash;
1820
1821         if (M_HASHTYPE_ISHASH(m)) {
1822                 hash = m->m_pkthdr.flowid;
1823         } else {
1824                 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1825                         hash += parser->data.raw[i];
1826         }
1827         return (&lc->lro_hash[hash % lc->lro_hashsz]);
1828 }
1829
1830 static int
1831 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1832 {
1833         struct lro_parser pi;   /* inner address data */
1834         struct lro_parser po;   /* outer address data */
1835         struct lro_parser *pa;  /* current parser for TCP stream */
1836         struct lro_entry *le;
1837         struct lro_head *bucket;
1838         struct tcphdr *th;
1839         int tcp_data_len;
1840         int tcp_opt_len;
1841         int error;
1842         uint16_t tcp_data_sum;
1843
1844 #ifdef INET
1845         /* Quickly decide if packet cannot be LRO'ed */
1846         if (__predict_false(V_ipforwarding != 0))
1847                 return (TCP_LRO_CANNOT);
1848 #endif
1849 #ifdef INET6
1850         /* Quickly decide if packet cannot be LRO'ed */
1851         if (__predict_false(V_ip6_forwarding != 0))
1852                 return (TCP_LRO_CANNOT);
1853 #endif
1854         if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1855              ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1856             (m->m_pkthdr.csum_data != 0xffff)) {
1857                 /*
1858                  * The checksum either did not have hardware offload
1859                  * or it was a bad checksum. We can't LRO such
1860                  * a packet.
1861                  */
1862                 counter_u64_add(tcp_bad_csums, 1);
1863                 return (TCP_LRO_CANNOT);
1864         }
1865         /* We expect a contiguous header [eh, ip, tcp]. */
1866         pa = tcp_lro_parser(m, &po, &pi, true);
1867         if (__predict_false(pa == NULL))
1868                 return (TCP_LRO_NOT_SUPPORTED);
1869
1870         /* We don't expect any padding. */
1871         error = tcp_lro_trim_mbuf_chain(m, pa);
1872         if (__predict_false(error != 0))
1873                 return (error);
1874
1875 #ifdef INET
1876         switch (pa->data.lro_type) {
1877         case LRO_TYPE_IPV4_TCP:
1878                 error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1879                 if (__predict_false(error != 0))
1880                         return (error);
1881                 break;
1882         default:
1883                 break;
1884         }
1885 #endif
1886         /* If no hardware or arrival stamp on the packet add timestamp */
1887         if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1888                 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1889                 m->m_flags |= M_TSTMP_LRO;
1890         }
1891
1892         /* Get pointer to TCP header. */
1893         th = pa->tcp;
1894
1895         /* Don't process SYN packets. */
1896         if (__predict_false(tcp_get_flags(th) & TH_SYN))
1897                 return (TCP_LRO_CANNOT);
1898
1899         /* Get total TCP header length and compute payload length. */
1900         tcp_opt_len = (th->th_off << 2);
1901         tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1902             (uint8_t *)m->m_data) - tcp_opt_len;
1903         tcp_opt_len -= sizeof(*th);
1904
1905         /* Don't process invalid TCP headers. */
1906         if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1907                 return (TCP_LRO_CANNOT);
1908
1909         /* Compute TCP data only checksum. */
1910         if (tcp_data_len == 0)
1911                 tcp_data_sum = 0;       /* no data, no checksum */
1912         else if (__predict_false(csum != 0))
1913                 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1914         else
1915                 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1916
1917         /* Save TCP info in mbuf. */
1918         m->m_nextpkt = NULL;
1919         m->m_pkthdr.rcvif = lc->ifp;
1920         m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1921         m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1922         m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1923         m->m_pkthdr.lro_nsegs = 1;
1924
1925         /* Get hash bucket. */
1926         if (!use_hash) {
1927                 bucket = &lc->lro_hash[0];
1928         } else {
1929                 bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1930         }
1931
1932         /* Try to find a matching previous segment. */
1933         LIST_FOREACH(le, bucket, hash_next) {
1934                 /* Compare addresses and ports. */
1935                 if (lro_address_compare(&po.data, &le->outer.data) == false ||
1936                     lro_address_compare(&pi.data, &le->inner.data) == false)
1937                         continue;
1938
1939                 /* Check if no data and old ACK. */
1940                 if (tcp_data_len == 0 &&
1941                     SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1942                         m_freem(m);
1943                         return (0);
1944                 }
1945
1946                 /* Mark "m" in the last spot. */
1947                 le->m_last_mbuf->m_nextpkt = m;
1948                 /* Now set the tail to "m". */
1949                 le->m_last_mbuf = m;
1950                 return (0);
1951         }
1952
1953         /* Try to find an empty slot. */
1954         if (LIST_EMPTY(&lc->lro_free))
1955                 return (TCP_LRO_NO_ENTRIES);
1956
1957         /* Start a new segment chain. */
1958         le = LIST_FIRST(&lc->lro_free);
1959         LIST_REMOVE(le, next);
1960         tcp_lro_active_insert(lc, bucket, le);
1961
1962         /* Make sure the headers are set. */
1963         le->inner = pi;
1964         le->outer = po;
1965
1966         /* Store time this entry was allocated. */
1967         le->alloc_time = lc->lro_last_queue_time;
1968
1969         tcp_set_entry_to_mbuf(lc, le, m, th);
1970
1971         /* Now set the tail to "m". */
1972         le->m_last_mbuf = m;
1973
1974         return (0);
1975 }
1976
1977 int
1978 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1979 {
1980         int error;
1981
1982         if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1983              ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1984             (m->m_pkthdr.csum_data != 0xffff)) {
1985                 /*
1986                  * The checksum either did not have hardware offload
1987                  * or it was a bad checksum. We can't LRO such
1988                  * a packet.
1989                  */
1990                 counter_u64_add(tcp_bad_csums, 1);
1991                 return (TCP_LRO_CANNOT);
1992         }
1993         /* get current time */
1994         binuptime(&lc->lro_last_queue_time);
1995         CURVNET_SET(lc->ifp->if_vnet);
1996         error = tcp_lro_rx_common(lc, m, csum, true);
1997         if (__predict_false(error != 0)) {
1998                 /*
1999                  * Flush anything we have acummulated
2000                  * ahead of this packet that can't
2001                  * be LRO'd. This preserves order.
2002                  */
2003                 tcp_lro_flush_active(lc);
2004         }
2005         CURVNET_RESTORE();
2006
2007         return (error);
2008 }
2009
2010 void
2011 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
2012 {
2013         NET_EPOCH_ASSERT();
2014         /* sanity checks */
2015         if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
2016             lc->lro_mbuf_max == 0)) {
2017                 /* packet drop */
2018                 m_freem(mb);
2019                 return;
2020         }
2021
2022         /* check if packet is not LRO capable */
2023         if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
2024                 /* input packet to network layer */
2025                 (*lc->ifp->if_input) (lc->ifp, mb);
2026                 return;
2027         }
2028
2029         /* If no hardware or arrival stamp on the packet add timestamp */
2030         if ((tcplro_stacks_wanting_mbufq > 0) &&
2031             (tcp_less_accurate_lro_ts == 0) &&
2032             ((mb->m_flags & M_TSTMP) == 0)) {
2033                 /* Add in an LRO time since no hardware */
2034                 binuptime(&lc->lro_last_queue_time);
2035                 mb->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
2036                 mb->m_flags |= M_TSTMP_LRO;
2037         }
2038
2039         /* create sequence number */
2040         lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
2041             (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
2042             (((uint64_t)mb->m_pkthdr.flowid) << 24) |
2043             ((uint64_t)lc->lro_mbuf_count);
2044
2045         /* enter mbuf */
2046         lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
2047
2048         /* flush if array is full */
2049         if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
2050                 tcp_lro_flush_all(lc);
2051 }
2052
2053 /* end */