sys/netinet/tcp_lro.c

   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 #include <net/if_dl.h>
  57 #include <net/if_media.h>
  58 #include <net/if_types.h>
  59 #include <net/infiniband.h>
  60 #include <net/if_lagg.h>
  61
  62 #include <netinet/in_systm.h>
  63 #include <netinet/in.h>
  64 #include <netinet/ip6.h>
  65 #include <netinet/ip.h>
  66 #include <netinet/ip_var.h>
  67 #include <netinet/in_pcb.h>
  68 #include <netinet6/in6_pcb.h>
  69 #include <netinet/tcp.h>
  70 #include <netinet/tcp_seq.h>
  71 #include <netinet/tcp_lro.h>
  72 #include <netinet/tcp_var.h>
  73 #include <netinet/tcpip.h>
  74 #include <netinet/tcp_hpts.h>
  75 #include <netinet/tcp_log_buf.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 inpcb *, struct lro_ctrl *,
  93                 struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **,
  94                 bool *, bool, bool, struct ifnet *);
  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 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD,
 123     &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport");
 124 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD,
 125     &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts");
 126 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD,
 127     &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport");
 128 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD,
 129     &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken");
 130 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD,
 131     &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp");
 132 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD,
 133     &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed");
 134 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD,
 135     &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set");
 136 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD,
 137     &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP");
 138 SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD,
 139     &tcp_bad_csums, "Number of packets that the common code saw with bad csums");
 140
 141 void
 142 tcp_lro_reg_mbufq(void)
 143 {
 144         atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1);
 145 }
 146
 147 void
 148 tcp_lro_dereg_mbufq(void)
 149 {
 150         atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1);
 151 }
 152
 153 static __inline void
 154 tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket,
 155     struct lro_entry *le)
 156 {
 157
 158         LIST_INSERT_HEAD(&lc->lro_active, le, next);
 159         LIST_INSERT_HEAD(bucket, le, hash_next);
 160 }
 161
 162 static __inline void
 163 tcp_lro_active_remove(struct lro_entry *le)
 164 {
 165
 166         LIST_REMOVE(le, next);          /* active list */
 167         LIST_REMOVE(le, hash_next);     /* hash bucket */
 168 }
 169
 170 int
 171 tcp_lro_init(struct lro_ctrl *lc)
 172 {
 173         return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0));
 174 }
 175
 176 int
 177 tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp,
 178     unsigned lro_entries, unsigned lro_mbufs)
 179 {
 180         struct lro_entry *le;
 181         size_t size;
 182         unsigned i, elements;
 183
 184         lc->lro_bad_csum = 0;
 185         lc->lro_queued = 0;
 186         lc->lro_flushed = 0;
 187         lc->lro_mbuf_count = 0;
 188         lc->lro_mbuf_max = lro_mbufs;
 189         lc->lro_cnt = lro_entries;
 190         lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX;
 191         lc->lro_length_lim = TCP_LRO_LENGTH_MAX;
 192         lc->ifp = ifp;
 193         LIST_INIT(&lc->lro_free);
 194         LIST_INIT(&lc->lro_active);
 195
 196         /* create hash table to accelerate entry lookup */
 197         if (lro_entries > lro_mbufs)
 198                 elements = lro_entries;
 199         else
 200                 elements = lro_mbufs;
 201         lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz,
 202             HASH_NOWAIT);
 203         if (lc->lro_hash == NULL) {
 204                 memset(lc, 0, sizeof(*lc));
 205                 return (ENOMEM);
 206         }
 207
 208         /* compute size to allocate */
 209         size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) +
 210             (lro_entries * sizeof(*le));
 211         lc->lro_mbuf_data = (struct lro_mbuf_sort *)
 212             malloc(size, M_LRO, M_NOWAIT | M_ZERO);
 213
 214         /* check for out of memory */
 215         if (lc->lro_mbuf_data == NULL) {
 216                 free(lc->lro_hash, M_LRO);
 217                 memset(lc, 0, sizeof(*lc));
 218                 return (ENOMEM);
 219         }
 220         /* compute offset for LRO entries */
 221         le = (struct lro_entry *)
 222             (lc->lro_mbuf_data + lro_mbufs);
 223
 224         /* setup linked list */
 225         for (i = 0; i != lro_entries; i++)
 226                 LIST_INSERT_HEAD(&lc->lro_free, le + i, next);
 227
 228         return (0);
 229 }
 230
 231 struct vxlan_header {
 232         uint32_t        vxlh_flags;
 233         uint32_t        vxlh_vni;
 234 };
 235
 236 static inline void *
 237 tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen)
 238 {
 239         const struct ether_vlan_header *eh;
 240         void *old;
 241         uint16_t eth_type;
 242
 243         if (update_data)
 244                 memset(parser, 0, sizeof(*parser));
 245
 246         old = ptr;
 247
 248         if (is_vxlan) {
 249                 const struct vxlan_header *vxh;
 250                 vxh = ptr;
 251                 ptr = (uint8_t *)ptr + sizeof(*vxh);
 252                 if (update_data) {
 253                         parser->data.vxlan_vni =
 254                             vxh->vxlh_vni & htonl(0xffffff00);
 255                 }
 256         }
 257
 258         eh = ptr;
 259         if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) {
 260                 eth_type = eh->evl_proto;
 261                 if (update_data) {
 262                         /* strip priority and keep VLAN ID only */
 263                         parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK);
 264                 }
 265                 /* advance to next header */
 266                 ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
 267                 mlen -= (ETHER_HDR_LEN  + ETHER_VLAN_ENCAP_LEN);
 268         } else {
 269                 eth_type = eh->evl_encap_proto;
 270                 /* advance to next header */
 271                 mlen -= ETHER_HDR_LEN;
 272                 ptr = (uint8_t *)ptr + ETHER_HDR_LEN;
 273         }
 274         if (__predict_false(mlen <= 0))
 275                 return (NULL);
 276         switch (eth_type) {
 277 #ifdef INET
 278         case htons(ETHERTYPE_IP):
 279                 parser->ip4 = ptr;
 280                 if (__predict_false(mlen < sizeof(struct ip)))
 281                         return (NULL);
 282                 /* Ensure there are no IPv4 options. */
 283                 if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4))
 284                         break;
 285                 /* .. and the packet is not fragmented. */
 286                 if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK))
 287                         break;
 288                 ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2);
 289                 mlen -= sizeof(struct ip);
 290                 if (update_data) {
 291                         parser->data.s_addr.v4 = parser->ip4->ip_src;
 292                         parser->data.d_addr.v4 = parser->ip4->ip_dst;
 293                 }
 294                 switch (parser->ip4->ip_p) {
 295                 case IPPROTO_UDP:
 296                         if (__predict_false(mlen < sizeof(struct udphdr)))
 297                                 return (NULL);
 298                         parser->udp = ptr;
 299                         if (update_data) {
 300                                 parser->data.lro_type = LRO_TYPE_IPV4_UDP;
 301                                 parser->data.s_port = parser->udp->uh_sport;
 302                                 parser->data.d_port = parser->udp->uh_dport;
 303                         } else {
 304                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP);
 305                         }
 306                         ptr = ((uint8_t *)ptr + sizeof(*parser->udp));
 307                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
 308                         return (ptr);
 309                 case IPPROTO_TCP:
 310                         parser->tcp = ptr;
 311                         if (__predict_false(mlen < sizeof(struct tcphdr)))
 312                                 return (NULL);
 313                         if (update_data) {
 314                                 parser->data.lro_type = LRO_TYPE_IPV4_TCP;
 315                                 parser->data.s_port = parser->tcp->th_sport;
 316                                 parser->data.d_port = parser->tcp->th_dport;
 317                         } else {
 318                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP);
 319                         }
 320                         if (__predict_false(mlen < (parser->tcp->th_off << 2)))
 321                                 return (NULL);
 322                         ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
 323                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
 324                         return (ptr);
 325                 default:
 326                         break;
 327                 }
 328                 break;
 329 #endif
 330 #ifdef INET6
 331         case htons(ETHERTYPE_IPV6):
 332                 parser->ip6 = ptr;
 333                 if (__predict_false(mlen < sizeof(struct ip6_hdr)))
 334                         return (NULL);
 335                 ptr = (uint8_t *)ptr + sizeof(*parser->ip6);
 336                 if (update_data) {
 337                         parser->data.s_addr.v6 = parser->ip6->ip6_src;
 338                         parser->data.d_addr.v6 = parser->ip6->ip6_dst;
 339                 }
 340                 mlen -= sizeof(struct ip6_hdr);
 341                 switch (parser->ip6->ip6_nxt) {
 342                 case IPPROTO_UDP:
 343                         if (__predict_false(mlen < sizeof(struct udphdr)))
 344                                 return (NULL);
 345                         parser->udp = ptr;
 346                         if (update_data) {
 347                                 parser->data.lro_type = LRO_TYPE_IPV6_UDP;
 348                                 parser->data.s_port = parser->udp->uh_sport;
 349                                 parser->data.d_port = parser->udp->uh_dport;
 350                         } else {
 351                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP);
 352                         }
 353                         ptr = (uint8_t *)ptr + sizeof(*parser->udp);
 354                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
 355                         return (ptr);
 356                 case IPPROTO_TCP:
 357                         if (__predict_false(mlen < sizeof(struct tcphdr)))
 358                                 return (NULL);
 359                         parser->tcp = ptr;
 360                         if (update_data) {
 361                                 parser->data.lro_type = LRO_TYPE_IPV6_TCP;
 362                                 parser->data.s_port = parser->tcp->th_sport;
 363                                 parser->data.d_port = parser->tcp->th_dport;
 364                         } else {
 365                                 MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP);
 366                         }
 367                         if (__predict_false(mlen < (parser->tcp->th_off << 2)))
 368                                 return (NULL);
 369                         ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2);
 370                         parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old;
 371                         return (ptr);
 372                 default:
 373                         break;
 374                 }
 375                 break;
 376 #endif
 377         default:
 378                 break;
 379         }
 380         /* Invalid packet - cannot parse */
 381         return (NULL);
 382 }
 383
 384 static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID |
 385     CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID;
 386
 387 static inline struct lro_parser *
 388 tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data)
 389 {
 390         void *data_ptr;
 391
 392         /* Try to parse outer headers first. */
 393         data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len);
 394         if (data_ptr == NULL || po->total_hdr_len > m->m_len)
 395                 return (NULL);
 396
 397         if (update_data) {
 398                 /* Store VLAN ID, if any. */
 399                 if (__predict_false(m->m_flags & M_VLANTAG)) {
 400                         po->data.vlan_id =
 401                             htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK);
 402                 }
 403                 /* Store decrypted flag, if any. */
 404                 if (__predict_false((m->m_pkthdr.csum_flags &
 405                     CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED))
 406                         po->data.lro_flags |= LRO_FLAG_DECRYPTED;
 407         }
 408
 409         switch (po->data.lro_type) {
 410         case LRO_TYPE_IPV4_UDP:
 411         case LRO_TYPE_IPV6_UDP:
 412                 /* Check for VXLAN headers. */
 413                 if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum)
 414                         break;
 415
 416                 /* Try to parse inner headers. */
 417                 data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true,
 418                                                     (m->m_len - ((caddr_t)data_ptr - m->m_data)));
 419                 if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len)
 420                         break;
 421
 422                 /* Verify supported header types. */
 423                 switch (pi->data.lro_type) {
 424                 case LRO_TYPE_IPV4_TCP:
 425                 case LRO_TYPE_IPV6_TCP:
 426                         return (pi);
 427                 default:
 428                         break;
 429                 }
 430                 break;
 431         case LRO_TYPE_IPV4_TCP:
 432         case LRO_TYPE_IPV6_TCP:
 433                 if (update_data)
 434                         memset(pi, 0, sizeof(*pi));
 435                 return (po);
 436         default:
 437                 break;
 438         }
 439         return (NULL);
 440 }
 441
 442 static inline int
 443 tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po)
 444 {
 445         int len;
 446
 447         switch (po->data.lro_type) {
 448 #ifdef INET
 449         case LRO_TYPE_IPV4_TCP:
 450                 len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) +
 451                     ntohs(po->ip4->ip_len);
 452                 break;
 453 #endif
 454 #ifdef INET6
 455         case LRO_TYPE_IPV6_TCP:
 456                 len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) +
 457                     ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6);
 458                 break;
 459 #endif
 460         default:
 461                 return (TCP_LRO_CANNOT);
 462         }
 463
 464         /*
 465          * If the frame is padded beyond the end of the IP packet,
 466          * then trim the extra bytes off:
 467          */
 468         if (__predict_true(m->m_pkthdr.len == len)) {
 469                 return (0);
 470         } else if (m->m_pkthdr.len > len) {
 471                 m_adj(m, len - m->m_pkthdr.len);
 472                 return (0);
 473         }
 474         return (TCP_LRO_CANNOT);
 475 }
 476
 477 static struct tcphdr *
 478 tcp_lro_get_th(struct mbuf *m)
 479 {
 480         return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off));
 481 }
 482
 483 static void
 484 lro_free_mbuf_chain(struct mbuf *m)
 485 {
 486         struct mbuf *save;
 487
 488         while (m) {
 489                 save = m->m_nextpkt;
 490                 m->m_nextpkt = NULL;
 491                 m_freem(m);
 492                 m = save;
 493         }
 494 }
 495
 496 void
 497 tcp_lro_free(struct lro_ctrl *lc)
 498 {
 499         struct lro_entry *le;
 500         unsigned x;
 501
 502         /* reset LRO free list */
 503         LIST_INIT(&lc->lro_free);
 504
 505         /* free active mbufs, if any */
 506         while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
 507                 tcp_lro_active_remove(le);
 508                 lro_free_mbuf_chain(le->m_head);
 509         }
 510
 511         /* free hash table */
 512         free(lc->lro_hash, M_LRO);
 513         lc->lro_hash = NULL;
 514         lc->lro_hashsz = 0;
 515
 516         /* free mbuf array, if any */
 517         for (x = 0; x != lc->lro_mbuf_count; x++)
 518                 m_freem(lc->lro_mbuf_data[x].mb);
 519         lc->lro_mbuf_count = 0;
 520
 521         /* free allocated memory, if any */
 522         free(lc->lro_mbuf_data, M_LRO);
 523         lc->lro_mbuf_data = NULL;
 524 }
 525
 526 static uint16_t
 527 tcp_lro_rx_csum_tcphdr(const struct tcphdr *th)
 528 {
 529         const uint16_t *ptr;
 530         uint32_t csum;
 531         uint16_t len;
 532
 533         csum = -th->th_sum;     /* exclude checksum field */
 534         len = th->th_off;
 535         ptr = (const uint16_t *)th;
 536         while (len--) {
 537                 csum += *ptr;
 538                 ptr++;
 539                 csum += *ptr;
 540                 ptr++;
 541         }
 542         while (csum > 0xffff)
 543                 csum = (csum >> 16) + (csum & 0xffff);
 544
 545         return (csum);
 546 }
 547
 548 static uint16_t
 549 tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum)
 550 {
 551         uint32_t c;
 552         uint16_t cs;
 553
 554         c = tcp_csum;
 555
 556         switch (pa->data.lro_type) {
 557 #ifdef INET6
 558         case LRO_TYPE_IPV6_TCP:
 559                 /* Compute full pseudo IPv6 header checksum. */
 560                 cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0);
 561                 break;
 562 #endif
 563 #ifdef INET
 564         case LRO_TYPE_IPV4_TCP:
 565                 /* Compute full pseudo IPv4 header checsum. */
 566                 cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP);
 567                 cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs));
 568                 break;
 569 #endif
 570         default:
 571                 cs = 0;         /* Keep compiler happy. */
 572                 break;
 573         }
 574
 575         /* Complement checksum. */
 576         cs = ~cs;
 577         c += cs;
 578
 579         /* Remove TCP header checksum. */
 580         cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp);
 581         c += cs;
 582
 583         /* Compute checksum remainder. */
 584         while (c > 0xffff)
 585                 c = (c >> 16) + (c & 0xffff);
 586
 587         return (c);
 588 }
 589
 590 static void
 591 tcp_lro_rx_done(struct lro_ctrl *lc)
 592 {
 593         struct lro_entry *le;
 594
 595         while ((le = LIST_FIRST(&lc->lro_active)) != NULL) {
 596                 tcp_lro_active_remove(le);
 597                 tcp_lro_flush(lc, le);
 598         }
 599 }
 600
 601 void
 602 tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout)
 603 {
 604         struct lro_entry *le, *le_tmp;
 605         uint64_t now, tov;
 606         struct bintime bt;
 607
 608         NET_EPOCH_ASSERT();
 609         if (LIST_EMPTY(&lc->lro_active))
 610                 return;
 611
 612         /* get timeout time and current time in ns */
 613         binuptime(&bt);
 614         now = bintime2ns(&bt);
 615         tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
 616         LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
 617                 if (now >= (bintime2ns(&le->alloc_time) + tov)) {
 618                         tcp_lro_active_remove(le);
 619                         tcp_lro_flush(lc, le);
 620                 }
 621         }
 622 }
 623
 624 #ifdef INET
 625 static int
 626 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
 627 {
 628         uint16_t csum;
 629
 630         /* Legacy IP has a header checksum that needs to be correct. */
 631         if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
 632                 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
 633                         lc->lro_bad_csum++;
 634                         return (TCP_LRO_CANNOT);
 635                 }
 636         } else {
 637                 csum = in_cksum_hdr(ip4);
 638                 if (__predict_false(csum != 0)) {
 639                         lc->lro_bad_csum++;
 640                         return (TCP_LRO_CANNOT);
 641                 }
 642         }
 643         return (0);
 644 }
 645 #endif
 646
 647 #ifdef TCPHPTS
 648 static void
 649 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc,
 650     const struct lro_entry *le, const struct mbuf *m,
 651     int frm, int32_t tcp_data_len, uint32_t th_seq,
 652     uint32_t th_ack, uint16_t th_win)
 653 {
 654         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
 655                 union tcp_log_stackspecific log;
 656                 struct timeval tv, btv;
 657                 uint32_t cts;
 658
 659                 cts = tcp_get_usecs(&tv);
 660                 memset(&log, 0, sizeof(union tcp_log_stackspecific));
 661                 log.u_bbr.flex8 = frm;
 662                 log.u_bbr.flex1 = tcp_data_len;
 663                 if (m)
 664                         log.u_bbr.flex2 = m->m_pkthdr.len;
 665                 else
 666                         log.u_bbr.flex2 = 0;
 667                 log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
 668                 log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
 669                 if (le->m_head) {
 670                         log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
 671                         log.u_bbr.delRate = le->m_head->m_flags;
 672                         log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
 673                 }
 674                 log.u_bbr.inflight = th_seq;
 675                 log.u_bbr.delivered = th_ack;
 676                 log.u_bbr.timeStamp = cts;
 677                 log.u_bbr.epoch = le->next_seq;
 678                 log.u_bbr.lt_epoch = le->ack_seq;
 679                 log.u_bbr.pacing_gain = th_win;
 680                 log.u_bbr.cwnd_gain = le->window;
 681                 log.u_bbr.lost = curcpu;
 682                 log.u_bbr.cur_del_rate = (uintptr_t)m;
 683                 log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
 684                 bintime2timeval(&lc->lro_last_queue_time, &btv);
 685                 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv);
 686                 log.u_bbr.flex7 = le->compressed;
 687                 log.u_bbr.pacing_gain = le->uncompressed;
 688                 if (in_epoch(net_epoch_preempt))
 689                         log.u_bbr.inhpts = 1;
 690                 else
 691                         log.u_bbr.inhpts = 0;
 692                 TCP_LOG_EVENTP(tp, NULL,
 693                                &tp->t_inpcb->inp_socket->so_rcv,
 694                                &tp->t_inpcb->inp_socket->so_snd,
 695                                TCP_LOG_LRO, 0,
 696                                0, &log, false, &tv);
 697         }
 698 }
 699 #endif
 700
 701 static inline void
 702 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
 703 {
 704         uint32_t csum;
 705
 706         csum = 0xffff - *ptr + value;
 707         while (csum > 0xffff)
 708                 csum = (csum >> 16) + (csum & 0xffff);
 709         *ptr = value;
 710         *psum = csum;
 711 }
 712
 713 static uint16_t
 714 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
 715     uint16_t payload_len, uint16_t delta_sum)
 716 {
 717         uint32_t csum;
 718         uint16_t tlen;
 719         uint16_t temp[5] = {};
 720
 721         switch (pa->data.lro_type) {
 722         case LRO_TYPE_IPV4_TCP:
 723                 /* Compute new IPv4 length. */
 724                 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
 725                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
 726
 727                 /* Subtract delta from current IPv4 checksum. */
 728                 csum = pa->ip4->ip_sum + 0xffff - temp[0];
 729                 while (csum > 0xffff)
 730                         csum = (csum >> 16) + (csum & 0xffff);
 731                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
 732                 goto update_tcp_header;
 733
 734         case LRO_TYPE_IPV6_TCP:
 735                 /* Compute new IPv6 length. */
 736                 tlen = (pa->tcp->th_off << 2) + payload_len;
 737                 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
 738                 goto update_tcp_header;
 739
 740         case LRO_TYPE_IPV4_UDP:
 741                 /* Compute new IPv4 length. */
 742                 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
 743                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
 744
 745                 /* Subtract delta from current IPv4 checksum. */
 746                 csum = pa->ip4->ip_sum + 0xffff - temp[0];
 747                 while (csum > 0xffff)
 748                         csum = (csum >> 16) + (csum & 0xffff);
 749                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
 750                 goto update_udp_header;
 751
 752         case LRO_TYPE_IPV6_UDP:
 753                 /* Compute new IPv6 length. */
 754                 tlen = sizeof(*pa->udp) + payload_len;
 755                 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
 756                 goto update_udp_header;
 757
 758         default:
 759                 return (0);
 760         }
 761
 762 update_tcp_header:
 763         /* Compute current TCP header checksum. */
 764         temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
 765
 766         /* Incorporate the latest ACK into the TCP header. */
 767         pa->tcp->th_ack = le->ack_seq;
 768         pa->tcp->th_win = le->window;
 769
 770         /* Incorporate latest timestamp into the TCP header. */
 771         if (le->timestamp != 0) {
 772                 uint32_t *ts_ptr;
 773
 774                 ts_ptr = (uint32_t *)(pa->tcp + 1);
 775                 ts_ptr[1] = htonl(le->tsval);
 776                 ts_ptr[2] = le->tsecr;
 777         }
 778
 779         /* Compute new TCP header checksum. */
 780         temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
 781
 782         /* Compute new TCP checksum. */
 783         csum = pa->tcp->th_sum + 0xffff - delta_sum +
 784             0xffff - temp[0] + 0xffff - temp[3] + temp[2];
 785         while (csum > 0xffff)
 786                 csum = (csum >> 16) + (csum & 0xffff);
 787
 788         /* Assign new TCP checksum. */
 789         tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
 790
 791         /* Compute all modififications affecting next checksum. */
 792         csum = temp[0] + temp[1] + 0xffff - temp[2] +
 793             temp[3] + temp[4] + delta_sum;
 794         while (csum > 0xffff)
 795                 csum = (csum >> 16) + (csum & 0xffff);
 796
 797         /* Return delta checksum to next stage, if any. */
 798         return (csum);
 799
 800 update_udp_header:
 801         tlen = sizeof(*pa->udp) + payload_len;
 802         /* Assign new UDP length and compute checksum delta. */
 803         tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
 804
 805         /* Check if there is a UDP checksum. */
 806         if (__predict_false(pa->udp->uh_sum != 0)) {
 807                 /* Compute new UDP checksum. */
 808                 csum = pa->udp->uh_sum + 0xffff - delta_sum +
 809                     0xffff - temp[0] + 0xffff - temp[2];
 810                 while (csum > 0xffff)
 811                         csum = (csum >> 16) + (csum & 0xffff);
 812                 /* Assign new UDP checksum. */
 813                 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
 814         }
 815
 816         /* Compute all modififications affecting next checksum. */
 817         csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
 818         while (csum > 0xffff)
 819                 csum = (csum >> 16) + (csum & 0xffff);
 820
 821         /* Return delta checksum to next stage, if any. */
 822         return (csum);
 823 }
 824
 825 static void
 826 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
 827 {
 828         /* Check if we need to recompute any checksums. */
 829         if (le->needs_merge) {
 830                 uint16_t csum;
 831
 832                 switch (le->inner.data.lro_type) {
 833                 case LRO_TYPE_IPV4_TCP:
 834                         csum = tcp_lro_update_checksum(&le->inner, le,
 835                             le->m_head->m_pkthdr.lro_tcp_d_len,
 836                             le->m_head->m_pkthdr.lro_tcp_d_csum);
 837                         csum = tcp_lro_update_checksum(&le->outer, NULL,
 838                             le->m_head->m_pkthdr.lro_tcp_d_len +
 839                             le->inner.total_hdr_len, csum);
 840                         le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 841                             CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
 842                         le->m_head->m_pkthdr.csum_data = 0xffff;
 843                         if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
 844                                 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
 845                         break;
 846                 case LRO_TYPE_IPV6_TCP:
 847                         csum = tcp_lro_update_checksum(&le->inner, le,
 848                             le->m_head->m_pkthdr.lro_tcp_d_len,
 849                             le->m_head->m_pkthdr.lro_tcp_d_csum);
 850                         csum = tcp_lro_update_checksum(&le->outer, NULL,
 851                             le->m_head->m_pkthdr.lro_tcp_d_len +
 852                             le->inner.total_hdr_len, csum);
 853                         le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 854                             CSUM_PSEUDO_HDR;
 855                         le->m_head->m_pkthdr.csum_data = 0xffff;
 856                         if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
 857                                 le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
 858                         break;
 859                 case LRO_TYPE_NONE:
 860                         switch (le->outer.data.lro_type) {
 861                         case LRO_TYPE_IPV4_TCP:
 862                                 csum = tcp_lro_update_checksum(&le->outer, le,
 863                                     le->m_head->m_pkthdr.lro_tcp_d_len,
 864                                     le->m_head->m_pkthdr.lro_tcp_d_csum);
 865                                 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 866                                     CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
 867                                 le->m_head->m_pkthdr.csum_data = 0xffff;
 868                                 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
 869                                         le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
 870                                 break;
 871                         case LRO_TYPE_IPV6_TCP:
 872                                 csum = tcp_lro_update_checksum(&le->outer, le,
 873                                     le->m_head->m_pkthdr.lro_tcp_d_len,
 874                                     le->m_head->m_pkthdr.lro_tcp_d_csum);
 875                                 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 876                                     CSUM_PSEUDO_HDR;
 877                                 le->m_head->m_pkthdr.csum_data = 0xffff;
 878                                 if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED))
 879                                         le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED;
 880                                 break;
 881                         default:
 882                                 break;
 883                         }
 884                         break;
 885                 default:
 886                         break;
 887                 }
 888         }
 889
 890         /*
 891          * Break any chain, this is not set to NULL on the singleton
 892          * case m_nextpkt points to m_head. Other case set them
 893          * m_nextpkt to NULL in push_and_replace.
 894          */
 895         le->m_head->m_nextpkt = NULL;
 896         lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
 897         (*lc->ifp->if_input)(lc->ifp, le->m_head);
 898 }
 899
 900 static void
 901 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
 902     struct mbuf *m, struct tcphdr *th)
 903 {
 904         uint32_t *ts_ptr;
 905         uint16_t tcp_data_len;
 906         uint16_t tcp_opt_len;
 907
 908         ts_ptr = (uint32_t *)(th + 1);
 909         tcp_opt_len = (th->th_off << 2);
 910         tcp_opt_len -= sizeof(*th);
 911
 912         /* Check if there is a timestamp option. */
 913         if (tcp_opt_len == 0 ||
 914             __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
 915             *ts_ptr != TCP_LRO_TS_OPTION)) {
 916                 /* We failed to find the timestamp option. */
 917                 le->timestamp = 0;
 918         } else {
 919                 le->timestamp = 1;
 920                 le->tsval = ntohl(*(ts_ptr + 1));
 921                 le->tsecr = *(ts_ptr + 2);
 922         }
 923
 924         tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
 925
 926         /* Pull out TCP sequence numbers and window size. */
 927         le->next_seq = ntohl(th->th_seq) + tcp_data_len;
 928         le->ack_seq = th->th_ack;
 929         le->window = th->th_win;
 930         le->flags = tcp_get_flags(th);
 931         le->needs_merge = 0;
 932
 933         /* Setup new data pointers. */
 934         le->m_head = m;
 935         le->m_tail = m_last(m);
 936 }
 937
 938 static void
 939 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
 940 {
 941         struct lro_parser *pa;
 942
 943         /*
 944          * Push up the stack of the current entry
 945          * and replace it with "m".
 946          */
 947         struct mbuf *msave;
 948
 949         /* Grab off the next and save it */
 950         msave = le->m_head->m_nextpkt;
 951         le->m_head->m_nextpkt = NULL;
 952
 953         /* Now push out the old entry */
 954         tcp_flush_out_entry(lc, le);
 955
 956         /* Re-parse new header, should not fail. */
 957         pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
 958         KASSERT(pa != NULL,
 959             ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
 960
 961         /*
 962          * Now to replace the data properly in the entry
 963          * we have to reset the TCP header and
 964          * other fields.
 965          */
 966         tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
 967
 968         /* Restore the next list */
 969         m->m_nextpkt = msave;
 970 }
 971
 972 static void
 973 tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p)
 974 {
 975         struct mbuf *m;
 976         uint32_t csum;
 977
 978         m = le->m_head;
 979         if (m->m_pkthdr.lro_nsegs == 1) {
 980                 /* Compute relative checksum. */
 981                 csum = p->m_pkthdr.lro_tcp_d_csum;
 982         } else {
 983                 /* Merge TCP data checksums. */
 984                 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
 985                     (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
 986                 while (csum > 0xffff)
 987                         csum = (csum >> 16) + (csum & 0xffff);
 988         }
 989
 990         /* Update various counters. */
 991         m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
 992         m->m_pkthdr.lro_tcp_d_csum = csum;
 993         m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
 994         m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
 995         le->needs_merge = 1;
 996 }
 997
 998 static void
 999 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
1000 {
1001         /*
1002          * Walk through the mbuf chain we
1003          * have on tap and compress/condense
1004          * as required.
1005          */
1006         uint32_t *ts_ptr;
1007         struct mbuf *m;
1008         struct tcphdr *th;
1009         uint32_t tcp_data_len_total;
1010         uint32_t tcp_data_seg_total;
1011         uint16_t tcp_data_len;
1012         uint16_t tcp_opt_len;
1013
1014         /*
1015          * First we must check the lead (m_head)
1016          * we must make sure that it is *not*
1017          * something that should be sent up
1018          * right away (sack etc).
1019          */
1020 again:
1021         m = le->m_head->m_nextpkt;
1022         if (m == NULL) {
1023                 /* Just one left. */
1024                 return;
1025         }
1026
1027         th = tcp_lro_get_th(m);
1028         tcp_opt_len = (th->th_off << 2);
1029         tcp_opt_len -= sizeof(*th);
1030         ts_ptr = (uint32_t *)(th + 1);
1031
1032         if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1033             *ts_ptr != TCP_LRO_TS_OPTION)) {
1034                 /*
1035                  * Its not the timestamp. We can't
1036                  * use this guy as the head.
1037                  */
1038                 le->m_head->m_nextpkt = m->m_nextpkt;
1039                 tcp_push_and_replace(lc, le, m);
1040                 goto again;
1041         }
1042         if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1043                 /*
1044                  * Make sure that previously seen segments/ACKs are delivered
1045                  * before this segment, e.g. FIN.
1046                  */
1047                 le->m_head->m_nextpkt = m->m_nextpkt;
1048                 tcp_push_and_replace(lc, le, m);
1049                 goto again;
1050         }
1051         while((m = le->m_head->m_nextpkt) != NULL) {
1052                 /*
1053                  * condense m into le, first
1054                  * pull m out of the list.
1055                  */
1056                 le->m_head->m_nextpkt = m->m_nextpkt;
1057                 m->m_nextpkt = NULL;
1058                 /* Setup my data */
1059                 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1060                 th = tcp_lro_get_th(m);
1061                 ts_ptr = (uint32_t *)(th + 1);
1062                 tcp_opt_len = (th->th_off << 2);
1063                 tcp_opt_len -= sizeof(*th);
1064                 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1065                 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1066
1067                 if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1068                     tcp_data_len_total >= lc->lro_length_lim) {
1069                         /* Flush now if appending will result in overflow. */
1070                         tcp_push_and_replace(lc, le, m);
1071                         goto again;
1072                 }
1073                 if (tcp_opt_len != 0 &&
1074                     __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1075                     *ts_ptr != TCP_LRO_TS_OPTION)) {
1076                         /*
1077                          * Maybe a sack in the new one? We need to
1078                          * start all over after flushing the
1079                          * current le. We will go up to the beginning
1080                          * and flush it (calling the replace again possibly
1081                          * or just returning).
1082                          */
1083                         tcp_push_and_replace(lc, le, m);
1084                         goto again;
1085                 }
1086                 if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) {
1087                         tcp_push_and_replace(lc, le, m);
1088                         goto again;
1089                 }
1090                 if (tcp_opt_len != 0) {
1091                         uint32_t tsval = ntohl(*(ts_ptr + 1));
1092                         /* Make sure timestamp values are increasing. */
1093                         if (TSTMP_GT(le->tsval, tsval))  {
1094                                 tcp_push_and_replace(lc, le, m);
1095                                 goto again;
1096                         }
1097                         le->tsval = tsval;
1098                         le->tsecr = *(ts_ptr + 2);
1099                 }
1100                 /* Try to append the new segment. */
1101                 if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1102                                     ((tcp_get_flags(th) & TH_ACK) !=
1103                                       (le->flags & TH_ACK)) ||
1104                                     (tcp_data_len == 0 &&
1105                                      le->ack_seq == th->th_ack &&
1106                                      le->window == th->th_win))) {
1107                         /* Out of order packet, non-ACK + ACK or dup ACK. */
1108                         tcp_push_and_replace(lc, le, m);
1109                         goto again;
1110                 }
1111                 if (tcp_data_len != 0 ||
1112                     SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1113                         le->next_seq += tcp_data_len;
1114                         le->ack_seq = th->th_ack;
1115                         le->window = th->th_win;
1116                         le->needs_merge = 1;
1117                 } else if (th->th_ack == le->ack_seq) {
1118                         if (WIN_GT(th->th_win, le->window)) {
1119                                 le->window = th->th_win;
1120                                 le->needs_merge = 1;
1121                         }
1122                 }
1123
1124                 if (tcp_data_len == 0) {
1125                         m_freem(m);
1126                         continue;
1127                 }
1128
1129                 /* Merge TCP data checksum and length to head mbuf. */
1130                 tcp_lro_mbuf_append_pkthdr(le, m);
1131
1132                 /*
1133                  * Adjust the mbuf so that m_data points to the first byte of
1134                  * the ULP payload.  Adjust the mbuf to avoid complications and
1135                  * append new segment to existing mbuf chain.
1136                  */
1137                 m_adj(m, m->m_pkthdr.len - tcp_data_len);
1138                 m_demote_pkthdr(m);
1139                 le->m_tail->m_next = m;
1140                 le->m_tail = m_last(m);
1141         }
1142 }
1143
1144 #ifdef TCPHPTS
1145 static void
1146 tcp_queue_pkts(struct inpcb *inp, struct tcpcb *tp, struct lro_entry *le)
1147 {
1148         INP_WLOCK_ASSERT(inp);
1149         if (tp->t_in_pkt == NULL) {
1150                 /* Nothing yet there */
1151                 tp->t_in_pkt = le->m_head;
1152                 tp->t_tail_pkt = le->m_last_mbuf;
1153         } else {
1154                 /* Already some there */
1155                 tp->t_tail_pkt->m_nextpkt = le->m_head;
1156                 tp->t_tail_pkt = le->m_last_mbuf;
1157         }
1158         le->m_head = NULL;
1159         le->m_last_mbuf = NULL;
1160 }
1161
1162 static struct mbuf *
1163 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le,
1164     struct inpcb *inp, int32_t *new_m)
1165 {
1166         struct tcpcb *tp;
1167         struct mbuf *m;
1168
1169         tp = intotcpcb(inp);
1170         if (__predict_false(tp == NULL))
1171                 return (NULL);
1172
1173         /* Look at the last mbuf if any in queue */
1174         m = tp->t_tail_pkt;
1175         if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
1176                 if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
1177                         tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
1178                         *new_m = 0;
1179                         counter_u64_add(tcp_extra_mbuf, 1);
1180                         return (m);
1181                 } else {
1182                         /* Mark we ran out of space */
1183                         inp->inp_flags2 |= INP_MBUF_L_ACKS;
1184                 }
1185         }
1186         /* Decide mbuf size. */
1187         if (inp->inp_flags2 & INP_MBUF_L_ACKS)
1188                 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR);
1189         else
1190                 m = m_gethdr(M_NOWAIT, MT_DATA);
1191
1192         if (__predict_false(m == NULL)) {
1193                 counter_u64_add(tcp_would_have_but, 1);
1194                 return (NULL);
1195         }
1196         counter_u64_add(tcp_comp_total, 1);
1197         m->m_flags |= M_ACKCMP;
1198         *new_m = 1;
1199         return (m);
1200 }
1201
1202 static struct inpcb *
1203 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa)
1204 {
1205         struct inpcb *inp;
1206
1207         switch (pa->data.lro_type) {
1208 #ifdef INET6
1209         case LRO_TYPE_IPV6_TCP:
1210                 inp = in6_pcblookup(&V_tcbinfo,
1211                     &pa->data.s_addr.v6,
1212                     pa->data.s_port,
1213                     &pa->data.d_addr.v6,
1214                     pa->data.d_port,
1215                     INPLOOKUP_WLOCKPCB,
1216                     ifp);
1217                 break;
1218 #endif
1219 #ifdef INET
1220         case LRO_TYPE_IPV4_TCP:
1221                 inp = in_pcblookup(&V_tcbinfo,
1222                     pa->data.s_addr.v4,
1223                     pa->data.s_port,
1224                     pa->data.d_addr.v4,
1225                     pa->data.d_port,
1226                     INPLOOKUP_WLOCKPCB,
1227                     ifp);
1228                 break;
1229 #endif
1230         default:
1231                 inp = NULL;
1232                 break;
1233         }
1234         return (inp);
1235 }
1236
1237 static inline bool
1238 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts)
1239 {
1240         /*
1241          * This function returns two bits of valuable information.
1242          * a) Is what is present capable of being ack-compressed,
1243          *    we can ack-compress if there is no options or just
1244          *    a timestamp option, and of course the th_flags must
1245          *    be correct as well.
1246          * b) Our other options present such as SACK. This is
1247          *    used to determine if we want to wakeup or not.
1248          */
1249         bool ret = true;
1250
1251         switch (th->th_off << 2) {
1252         case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
1253                 *ppts = (uint32_t *)(th + 1);
1254                 /* Check if we have only one timestamp option. */
1255                 if (**ppts == TCP_LRO_TS_OPTION)
1256                         *other_opts = false;
1257                 else {
1258                         *other_opts = true;
1259                         ret = false;
1260                 }
1261                 break;
1262         case (sizeof(*th)):
1263                 /* No options. */
1264                 *ppts = NULL;
1265                 *other_opts = false;
1266                 break;
1267         default:
1268                 *ppts = NULL;
1269                 *other_opts = true;
1270                 ret = false;
1271                 break;
1272         }
1273         /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
1274         if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0)
1275                 ret = false;
1276         /* If it has data on it we cannot compress it */
1277         if (m->m_pkthdr.lro_tcp_d_len)
1278                 ret = false;
1279
1280         /* ACK flag must be set. */
1281         if (!(tcp_get_flags(th) & TH_ACK))
1282                 ret = false;
1283         return (ret);
1284 }
1285
1286 static int
1287 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le)
1288 {
1289         struct inpcb *inp;
1290         struct tcpcb *tp;
1291         struct mbuf **pp, *cmp, *mv_to;
1292         struct ifnet *lagg_ifp;
1293         bool bpf_req, lagg_bpf_req, should_wake;
1294
1295         /* Check if packet doesn't belongs to our network interface. */
1296         if ((tcplro_stacks_wanting_mbufq == 0) ||
1297             (le->outer.data.vlan_id != 0) ||
1298             (le->inner.data.lro_type != LRO_TYPE_NONE))
1299                 return (TCP_LRO_CANNOT);
1300
1301 #ifdef INET6
1302         /*
1303          * Be proactive about unspecified IPv6 address in source. As
1304          * we use all-zero to indicate unbounded/unconnected pcb,
1305          * unspecified IPv6 address can be used to confuse us.
1306          *
1307          * Note that packets with unspecified IPv6 destination is
1308          * already dropped in ip6_input.
1309          */
1310         if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
1311             IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
1312                 return (TCP_LRO_CANNOT);
1313
1314         if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
1315             IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
1316                 return (TCP_LRO_CANNOT);
1317 #endif
1318         /* Lookup inp, if any. */
1319         inp = tcp_lro_lookup(lc->ifp,
1320             (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
1321         if (inp == NULL)
1322                 return (TCP_LRO_CANNOT);
1323
1324         counter_u64_add(tcp_inp_lro_locks_taken, 1);
1325
1326         /* Get TCP control structure. */
1327         tp = intotcpcb(inp);
1328
1329         /* Check if the inp is dead, Jim. */
1330         if (tp == NULL ||
1331             (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
1332                 INP_WUNLOCK(inp);
1333                 return (TCP_LRO_CANNOT);
1334         }
1335         if ((inp->inp_irq_cpu_set == 0)  && (lc->lro_cpu_is_set == 1)) {
1336                 inp->inp_irq_cpu = lc->lro_last_cpu;
1337                 inp->inp_irq_cpu_set = 1;
1338         }
1339         /* Check if the transport doesn't support the needed optimizations. */
1340         if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) {
1341                 INP_WUNLOCK(inp);
1342                 return (TCP_LRO_CANNOT);
1343         }
1344
1345         if (inp->inp_flags2 & INP_MBUF_QUEUE_READY)
1346                 should_wake = false;
1347         else
1348                 should_wake = true;
1349         /* Check if packets should be tapped to BPF. */
1350         bpf_req = bpf_peers_present(lc->ifp->if_bpf);
1351         lagg_bpf_req = false;
1352         lagg_ifp = NULL;
1353         if (lc->ifp->if_type == IFT_IEEE8023ADLAG ||
1354             lc->ifp->if_type == IFT_INFINIBANDLAG) {
1355                 struct lagg_port *lp = lc->ifp->if_lagg;
1356                 struct lagg_softc *sc = lp->lp_softc;
1357
1358                 lagg_ifp = sc->sc_ifp;
1359                 if (lagg_ifp != NULL)
1360                         lagg_bpf_req = bpf_peers_present(lagg_ifp->if_bpf);
1361         }
1362
1363         /* Strip and compress all the incoming packets. */
1364         cmp = NULL;
1365         for (pp = &le->m_head; *pp != NULL; ) {
1366                 mv_to = NULL;
1367                 if (do_bpf_strip_and_compress(inp, lc, le, pp,
1368                         &cmp, &mv_to, &should_wake, bpf_req,
1369                         lagg_bpf_req, lagg_ifp) == false) {
1370                         /* Advance to next mbuf. */
1371                         pp = &(*pp)->m_nextpkt;
1372                 } else if (mv_to != NULL) {
1373                         /* We are asked to move pp up */
1374                         pp = &mv_to->m_nextpkt;
1375                 }
1376         }
1377         /* Update "m_last_mbuf", if any. */
1378         if (pp == &le->m_head)
1379                 le->m_last_mbuf = *pp;
1380         else
1381                 le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt);
1382
1383         /* Check if any data mbufs left. */
1384         if (le->m_head != NULL) {
1385                 counter_u64_add(tcp_inp_lro_direct_queue, 1);
1386                 tcp_lro_log(tp, lc, le, NULL, 22, 1, inp->inp_flags2, 0, 1);
1387                 tcp_queue_pkts(inp, tp, le);
1388         }
1389         if (should_wake) {
1390                 /* Wakeup */
1391                 counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
1392                 if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0))
1393                         inp = NULL;
1394         }
1395         if (inp != NULL)
1396                 INP_WUNLOCK(inp);
1397         return (0);     /* Success. */
1398 }
1399 #endif
1400
1401 void
1402 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1403 {
1404         /* Only optimise if there are multiple packets waiting. */
1405 #ifdef TCPHPTS
1406         int error;
1407 #endif
1408
1409         NET_EPOCH_ASSERT();
1410 #ifdef TCPHPTS
1411         CURVNET_SET(lc->ifp->if_vnet);
1412         error = tcp_lro_flush_tcphpts(lc, le);
1413         CURVNET_RESTORE();
1414         if (error != 0) {
1415 #endif
1416                 tcp_lro_condense(lc, le);
1417                 tcp_flush_out_entry(lc, le);
1418 #ifdef TCPHPTS
1419         }
1420 #endif
1421         lc->lro_flushed++;
1422         bzero(le, sizeof(*le));
1423         LIST_INSERT_HEAD(&lc->lro_free, le, next);
1424 }
1425
1426 #ifdef HAVE_INLINE_FLSLL
1427 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1428 #else
1429 static inline uint64_t
1430 tcp_lro_msb_64(uint64_t x)
1431 {
1432         x |= (x >> 1);
1433         x |= (x >> 2);
1434         x |= (x >> 4);
1435         x |= (x >> 8);
1436         x |= (x >> 16);
1437         x |= (x >> 32);
1438         return (x & ~(x >> 1));
1439 }
1440 #endif
1441
1442 /*
1443  * The tcp_lro_sort() routine is comparable to qsort(), except it has
1444  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1445  * number of elements to sort and 64 is the number of sequence bits
1446  * available. The algorithm is bit-slicing the 64-bit sequence number,
1447  * sorting one bit at a time from the most significant bit until the
1448  * least significant one, skipping the constant bits. This is
1449  * typically called a radix sort.
1450  */
1451 static void
1452 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1453 {
1454         struct lro_mbuf_sort temp;
1455         uint64_t ones;
1456         uint64_t zeros;
1457         uint32_t x;
1458         uint32_t y;
1459
1460 repeat:
1461         /* for small arrays insertion sort is faster */
1462         if (size <= 12) {
1463                 for (x = 1; x < size; x++) {
1464                         temp = parray[x];
1465                         for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1466                                 parray[y] = parray[y - 1];
1467                         parray[y] = temp;
1468                 }
1469                 return;
1470         }
1471
1472         /* compute sequence bits which are constant */
1473         ones = 0;
1474         zeros = 0;
1475         for (x = 0; x != size; x++) {
1476                 ones |= parray[x].seq;
1477                 zeros |= ~parray[x].seq;
1478         }
1479
1480         /* compute bits which are not constant into "ones" */
1481         ones &= zeros;
1482         if (ones == 0)
1483                 return;
1484
1485         /* pick the most significant bit which is not constant */
1486         ones = tcp_lro_msb_64(ones);
1487
1488         /*
1489          * Move entries having cleared sequence bits to the beginning
1490          * of the array:
1491          */
1492         for (x = y = 0; y != size; y++) {
1493                 /* skip set bits */
1494                 if (parray[y].seq & ones)
1495                         continue;
1496                 /* swap entries */
1497                 temp = parray[x];
1498                 parray[x] = parray[y];
1499                 parray[y] = temp;
1500                 x++;
1501         }
1502
1503         KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1504
1505         /* sort zeros */
1506         tcp_lro_sort(parray, x);
1507
1508         /* sort ones */
1509         parray += x;
1510         size -= x;
1511         goto repeat;
1512 }
1513
1514 void
1515 tcp_lro_flush_all(struct lro_ctrl *lc)
1516 {
1517         uint64_t seq;
1518         uint64_t nseq;
1519         unsigned x;
1520
1521         NET_EPOCH_ASSERT();
1522         /* check if no mbufs to flush */
1523         if (lc->lro_mbuf_count == 0)
1524                 goto done;
1525         if (lc->lro_cpu_is_set == 0) {
1526                 if (lc->lro_last_cpu == curcpu) {
1527                         lc->lro_cnt_of_same_cpu++;
1528                         /* Have we reached the threshold to declare a cpu? */
1529                         if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1530                                 lc->lro_cpu_is_set = 1;
1531                 } else {
1532                         lc->lro_last_cpu = curcpu;
1533                         lc->lro_cnt_of_same_cpu = 0;
1534                 }
1535         }
1536         CURVNET_SET(lc->ifp->if_vnet);
1537
1538         /* get current time */
1539         binuptime(&lc->lro_last_queue_time);
1540
1541         /* sort all mbufs according to stream */
1542         tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1543
1544         /* input data into LRO engine, stream by stream */
1545         seq = 0;
1546         for (x = 0; x != lc->lro_mbuf_count; x++) {
1547                 struct mbuf *mb;
1548
1549                 /* get mbuf */
1550                 mb = lc->lro_mbuf_data[x].mb;
1551
1552                 /* get sequence number, masking away the packet index */
1553                 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1554
1555                 /* check for new stream */
1556                 if (seq != nseq) {
1557                         seq = nseq;
1558
1559                         /* flush active streams */
1560                         tcp_lro_rx_done(lc);
1561                 }
1562
1563                 /* add packet to LRO engine */
1564                 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1565                         /* input packet to network layer */
1566                         (*lc->ifp->if_input)(lc->ifp, mb);
1567                         lc->lro_queued++;
1568                         lc->lro_flushed++;
1569                 }
1570         }
1571         CURVNET_RESTORE();
1572 done:
1573         /* flush active streams */
1574         tcp_lro_rx_done(lc);
1575
1576 #ifdef TCPHPTS
1577         tcp_run_hpts();
1578 #endif
1579         lc->lro_mbuf_count = 0;
1580 }
1581
1582 #ifdef TCPHPTS
1583 static void
1584 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m,
1585     uint32_t *ts_ptr, uint16_t iptos)
1586 {
1587         /*
1588          * Given a TCP ACK, summarize it down into the small TCP ACK
1589          * entry.
1590          */
1591         ae->timestamp = m->m_pkthdr.rcv_tstmp;
1592         if (m->m_flags & M_TSTMP_LRO)
1593                 ae->flags = TSTMP_LRO;
1594         else if (m->m_flags & M_TSTMP)
1595                 ae->flags = TSTMP_HDWR;
1596         ae->seq = ntohl(th->th_seq);
1597         ae->ack = ntohl(th->th_ack);
1598         ae->flags |= tcp_get_flags(th);
1599         if (ts_ptr != NULL) {
1600                 ae->ts_value = ntohl(ts_ptr[1]);
1601                 ae->ts_echo = ntohl(ts_ptr[2]);
1602                 ae->flags |= HAS_TSTMP;
1603         }
1604         ae->win = ntohs(th->th_win);
1605         ae->codepoint = iptos;
1606 }
1607
1608 /*
1609  * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
1610  * and strip all, but the IPv4/IPv6 header.
1611  */
1612 static bool
1613 do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc,
1614     struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to,
1615     bool *should_wake, bool bpf_req, bool lagg_bpf_req, struct ifnet *lagg_ifp)
1616 {
1617         union {
1618                 void *ptr;
1619                 struct ip *ip4;
1620                 struct ip6_hdr *ip6;
1621         } l3;
1622         struct mbuf *m;
1623         struct mbuf *nm;
1624         struct tcphdr *th;
1625         struct tcp_ackent *ack_ent;
1626         uint32_t *ts_ptr;
1627         int32_t n_mbuf;
1628         bool other_opts, can_compress;
1629         uint8_t lro_type;
1630         uint16_t iptos;
1631         int tcp_hdr_offset;
1632         int idx;
1633
1634         /* Get current mbuf. */
1635         m = *pp;
1636
1637         /* Let the BPF see the packet */
1638         if (__predict_false(bpf_req))
1639                 ETHER_BPF_MTAP(lc->ifp, m);
1640
1641         if (__predict_false(lagg_bpf_req))
1642                 ETHER_BPF_MTAP(lagg_ifp, m);
1643
1644         tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
1645         lro_type = le->inner.data.lro_type;
1646         switch (lro_type) {
1647         case LRO_TYPE_NONE:
1648                 lro_type = le->outer.data.lro_type;
1649                 switch (lro_type) {
1650                 case LRO_TYPE_IPV4_TCP:
1651                         tcp_hdr_offset -= sizeof(*le->outer.ip4);
1652                         m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1653                         break;
1654                 case LRO_TYPE_IPV6_TCP:
1655                         tcp_hdr_offset -= sizeof(*le->outer.ip6);
1656                         m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1657                         break;
1658                 default:
1659                         goto compressed;
1660                 }
1661                 break;
1662         case LRO_TYPE_IPV4_TCP:
1663                 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1664                 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1665                 break;
1666         case LRO_TYPE_IPV6_TCP:
1667                 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1668                 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1669                 break;
1670         default:
1671                 goto compressed;
1672         }
1673
1674         MPASS(tcp_hdr_offset >= 0);
1675
1676         m_adj(m, tcp_hdr_offset);
1677         m->m_flags |= M_LRO_EHDRSTRP;
1678         m->m_flags &= ~M_ACKCMP;
1679         m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;
1680
1681         th = tcp_lro_get_th(m);
1682
1683         th->th_sum = 0;         /* TCP checksum is valid. */
1684
1685         /* Check if ACK can be compressed */
1686         can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts);
1687
1688         /* Now lets look at the should wake states */
1689         if ((other_opts == true) &&
1690             ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) {
1691                 /*
1692                  * If there are other options (SACK?) and the
1693                  * tcp endpoint has not expressly told us it does
1694                  * not care about SACKS, then we should wake up.
1695                  */
1696                 *should_wake = true;
1697         }
1698         /* Is the ack compressable? */
1699         if (can_compress == false)
1700                 goto done;
1701         /* Does the TCP endpoint support ACK compression? */
1702         if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0)
1703                 goto done;
1704
1705         /* Lets get the TOS/traffic class field */
1706         l3.ptr = mtod(m, void *);
1707         switch (lro_type) {
1708         case LRO_TYPE_IPV4_TCP:
1709                 iptos = l3.ip4->ip_tos;
1710                 break;
1711         case LRO_TYPE_IPV6_TCP:
1712                 iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
1713                 break;
1714         default:
1715                 iptos = 0;      /* Keep compiler happy. */
1716                 break;
1717         }
1718         /* Now lets get space if we don't have some already */
1719         if (*cmp == NULL) {
1720 new_one:
1721                 nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf);
1722                 if (__predict_false(nm == NULL))
1723                         goto done;
1724                 *cmp = nm;
1725                 if (n_mbuf) {
1726                         /*
1727                          *  Link in the new cmp ack to our in-order place,
1728                          * first set our cmp ack's next to where we are.
1729                          */
1730                         nm->m_nextpkt = m;
1731                         (*pp) = nm;
1732                         /*
1733                          * Set it up so mv_to is advanced to our
1734                          * compressed ack. This way the caller can
1735                          * advance pp to the right place.
1736                          */
1737                         *mv_to = nm;
1738                         /*
1739                          * Advance it here locally as well.
1740                          */
1741                         pp = &nm->m_nextpkt;
1742                 }
1743         } else {
1744                 /* We have one already we are working on */
1745                 nm = *cmp;
1746                 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
1747                         /* We ran out of space */
1748                         inp->inp_flags2 |= INP_MBUF_L_ACKS;
1749                         goto new_one;
1750                 }
1751         }
1752         MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
1753         counter_u64_add(tcp_inp_lro_compressed, 1);
1754         le->compressed++;
1755         /* We can add in to the one on the tail */
1756         ack_ent = mtod(nm, struct tcp_ackent *);
1757         idx = (nm->m_len / sizeof(struct tcp_ackent));
1758         build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
1759
1760         /* Bump the size of both pkt-hdr and len */
1761         nm->m_len += sizeof(struct tcp_ackent);
1762         nm->m_pkthdr.len += sizeof(struct tcp_ackent);
1763 compressed:
1764         /* Advance to next mbuf before freeing. */
1765         *pp = m->m_nextpkt;
1766         m->m_nextpkt = NULL;
1767         m_freem(m);
1768         return (true);
1769 done:
1770         counter_u64_add(tcp_uncomp_total, 1);
1771         le->uncompressed++;
1772         return (false);
1773 }
1774 #endif
1775
1776 static struct lro_head *
1777 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1778 {
1779         u_long hash;
1780
1781         if (M_HASHTYPE_ISHASH(m)) {
1782                 hash = m->m_pkthdr.flowid;
1783         } else {
1784                 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1785                         hash += parser->data.raw[i];
1786         }
1787         return (&lc->lro_hash[hash % lc->lro_hashsz]);
1788 }
1789
1790 static int
1791 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1792 {
1793         struct lro_parser pi;   /* inner address data */
1794         struct lro_parser po;   /* outer address data */
1795         struct lro_parser *pa;  /* current parser for TCP stream */
1796         struct lro_entry *le;
1797         struct lro_head *bucket;
1798         struct tcphdr *th;
1799         int tcp_data_len;
1800         int tcp_opt_len;
1801         int error;
1802         uint16_t tcp_data_sum;
1803
1804 #ifdef INET
1805         /* Quickly decide if packet cannot be LRO'ed */
1806         if (__predict_false(V_ipforwarding != 0))
1807                 return (TCP_LRO_CANNOT);
1808 #endif
1809 #ifdef INET6
1810         /* Quickly decide if packet cannot be LRO'ed */
1811         if (__predict_false(V_ip6_forwarding != 0))
1812                 return (TCP_LRO_CANNOT);
1813 #endif
1814         if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1815              ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1816             (m->m_pkthdr.csum_data != 0xffff)) {
1817                 /*
1818                  * The checksum either did not have hardware offload
1819                  * or it was a bad checksum. We can't LRO such
1820                  * a packet.
1821                  */
1822                 counter_u64_add(tcp_bad_csums, 1);
1823                 return (TCP_LRO_CANNOT);
1824         }
1825         /* We expect a contiguous header [eh, ip, tcp]. */
1826         pa = tcp_lro_parser(m, &po, &pi, true);
1827         if (__predict_false(pa == NULL))
1828                 return (TCP_LRO_NOT_SUPPORTED);
1829
1830         /* We don't expect any padding. */
1831         error = tcp_lro_trim_mbuf_chain(m, pa);
1832         if (__predict_false(error != 0))
1833                 return (error);
1834
1835 #ifdef INET
1836         switch (pa->data.lro_type) {
1837         case LRO_TYPE_IPV4_TCP:
1838                 error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1839                 if (__predict_false(error != 0))
1840                         return (error);
1841                 break;
1842         default:
1843                 break;
1844         }
1845 #endif
1846         /* If no hardware or arrival stamp on the packet add timestamp */
1847         if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1848                 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1849                 m->m_flags |= M_TSTMP_LRO;
1850         }
1851
1852         /* Get pointer to TCP header. */
1853         th = pa->tcp;
1854
1855         /* Don't process SYN packets. */
1856         if (__predict_false(tcp_get_flags(th) & TH_SYN))
1857                 return (TCP_LRO_CANNOT);
1858
1859         /* Get total TCP header length and compute payload length. */
1860         tcp_opt_len = (th->th_off << 2);
1861         tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1862             (uint8_t *)m->m_data) - tcp_opt_len;
1863         tcp_opt_len -= sizeof(*th);
1864
1865         /* Don't process invalid TCP headers. */
1866         if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1867                 return (TCP_LRO_CANNOT);
1868
1869         /* Compute TCP data only checksum. */
1870         if (tcp_data_len == 0)
1871                 tcp_data_sum = 0;       /* no data, no checksum */
1872         else if (__predict_false(csum != 0))
1873                 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1874         else
1875                 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1876
1877         /* Save TCP info in mbuf. */
1878         m->m_nextpkt = NULL;
1879         m->m_pkthdr.rcvif = lc->ifp;
1880         m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1881         m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1882         m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1883         m->m_pkthdr.lro_nsegs = 1;
1884
1885         /* Get hash bucket. */
1886         if (!use_hash) {
1887                 bucket = &lc->lro_hash[0];
1888         } else {
1889                 bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1890         }
1891
1892         /* Try to find a matching previous segment. */
1893         LIST_FOREACH(le, bucket, hash_next) {
1894                 /* Compare addresses and ports. */
1895                 if (lro_address_compare(&po.data, &le->outer.data) == false ||
1896                     lro_address_compare(&pi.data, &le->inner.data) == false)
1897                         continue;
1898
1899                 /* Check if no data and old ACK. */
1900                 if (tcp_data_len == 0 &&
1901                     SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1902                         m_freem(m);
1903                         return (0);
1904                 }
1905
1906                 /* Mark "m" in the last spot. */
1907                 le->m_last_mbuf->m_nextpkt = m;
1908                 /* Now set the tail to "m". */
1909                 le->m_last_mbuf = m;
1910                 return (0);
1911         }
1912
1913         /* Try to find an empty slot. */
1914         if (LIST_EMPTY(&lc->lro_free))
1915                 return (TCP_LRO_NO_ENTRIES);
1916
1917         /* Start a new segment chain. */
1918         le = LIST_FIRST(&lc->lro_free);
1919         LIST_REMOVE(le, next);
1920         tcp_lro_active_insert(lc, bucket, le);
1921
1922         /* Make sure the headers are set. */
1923         le->inner = pi;
1924         le->outer = po;
1925
1926         /* Store time this entry was allocated. */
1927         le->alloc_time = lc->lro_last_queue_time;
1928
1929         tcp_set_entry_to_mbuf(lc, le, m, th);
1930
1931         /* Now set the tail to "m". */
1932         le->m_last_mbuf = m;
1933
1934         return (0);
1935 }
1936
1937 int
1938 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1939 {
1940         int error;
1941
1942         if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1943              ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1944             (m->m_pkthdr.csum_data != 0xffff)) {
1945                 /*
1946                  * The checksum either did not have hardware offload
1947                  * or it was a bad checksum. We can't LRO such
1948                  * a packet.
1949                  */
1950                 counter_u64_add(tcp_bad_csums, 1);
1951                 return (TCP_LRO_CANNOT);
1952         }
1953         /* get current time */
1954         binuptime(&lc->lro_last_queue_time);
1955         CURVNET_SET(lc->ifp->if_vnet);
1956         error = tcp_lro_rx_common(lc, m, csum, true);
1957         CURVNET_RESTORE();
1958
1959         return (error);
1960 }
1961
1962 void
1963 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
1964 {
1965         NET_EPOCH_ASSERT();
1966         /* sanity checks */
1967         if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
1968             lc->lro_mbuf_max == 0)) {
1969                 /* packet drop */
1970                 m_freem(mb);
1971                 return;
1972         }
1973
1974         /* check if packet is not LRO capable */
1975         if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
1976                 /* input packet to network layer */
1977                 (*lc->ifp->if_input) (lc->ifp, mb);
1978                 return;
1979         }
1980
1981         /* create sequence number */
1982         lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
1983             (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
1984             (((uint64_t)mb->m_pkthdr.flowid) << 24) |
1985             ((uint64_t)lc->lro_mbuf_count);
1986
1987         /* enter mbuf */
1988         lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
1989
1990         /* flush if array is full */
1991         if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
1992                 tcp_lro_flush_all(lc);
1993 }
1994
1995 /* end */