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