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         NET_EPOCH_ASSERT();
 602         if (LIST_EMPTY(&lc->lro_active))
 603                 return;
 604
 605         /* get timeout time and current time in ns */
 606         binuptime(&bt);
 607         now = bintime2ns(&bt);
 608         tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000));
 609         LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) {
 610                 if (now >= (bintime2ns(&le->alloc_time) + tov)) {
 611                         tcp_lro_active_remove(le);
 612                         tcp_lro_flush(lc, le);
 613                 }
 614         }
 615 }
 616
 617 #ifdef INET
 618 static int
 619 tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4)
 620 {
 621         uint16_t csum;
 622
 623         /* Legacy IP has a header checksum that needs to be correct. */
 624         if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) {
 625                 if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) {
 626                         lc->lro_bad_csum++;
 627                         return (TCP_LRO_CANNOT);
 628                 }
 629         } else {
 630                 csum = in_cksum_hdr(ip4);
 631                 if (__predict_false(csum != 0)) {
 632                         lc->lro_bad_csum++;
 633                         return (TCP_LRO_CANNOT);
 634                 }
 635         }
 636         return (0);
 637 }
 638 #endif
 639
 640 #ifdef TCPHPTS
 641 static void
 642 tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc,
 643     const struct lro_entry *le, const struct mbuf *m,
 644     int frm, int32_t tcp_data_len, uint32_t th_seq,
 645     uint32_t th_ack, uint16_t th_win)
 646 {
 647         if (tp->t_logstate != TCP_LOG_STATE_OFF) {
 648                 union tcp_log_stackspecific log;
 649                 struct timeval tv, btv;
 650                 uint32_t cts;
 651
 652                 cts = tcp_get_usecs(&tv);
 653                 memset(&log, 0, sizeof(union tcp_log_stackspecific));
 654                 log.u_bbr.flex8 = frm;
 655                 log.u_bbr.flex1 = tcp_data_len;
 656                 if (m)
 657                         log.u_bbr.flex2 = m->m_pkthdr.len;
 658                 else
 659                         log.u_bbr.flex2 = 0;
 660                 log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs;
 661                 log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len;
 662                 if (le->m_head) {
 663                         log.u_bbr.flex5 = le->m_head->m_pkthdr.len;
 664                         log.u_bbr.delRate = le->m_head->m_flags;
 665                         log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp;
 666                 }
 667                 log.u_bbr.inflight = th_seq;
 668                 log.u_bbr.delivered = th_ack;
 669                 log.u_bbr.timeStamp = cts;
 670                 log.u_bbr.epoch = le->next_seq;
 671                 log.u_bbr.lt_epoch = le->ack_seq;
 672                 log.u_bbr.pacing_gain = th_win;
 673                 log.u_bbr.cwnd_gain = le->window;
 674                 log.u_bbr.lost = curcpu;
 675                 log.u_bbr.cur_del_rate = (uintptr_t)m;
 676                 log.u_bbr.bw_inuse = (uintptr_t)le->m_head;
 677                 bintime2timeval(&lc->lro_last_queue_time, &btv);
 678                 log.u_bbr.flex6 = tcp_tv_to_usectick(&btv);
 679                 log.u_bbr.flex7 = le->compressed;
 680                 log.u_bbr.pacing_gain = le->uncompressed;
 681                 if (in_epoch(net_epoch_preempt))
 682                         log.u_bbr.inhpts = 1;
 683                 else
 684                         log.u_bbr.inhpts = 0;
 685                 TCP_LOG_EVENTP(tp, NULL,
 686                                &tp->t_inpcb->inp_socket->so_rcv,
 687                                &tp->t_inpcb->inp_socket->so_snd,
 688                                TCP_LOG_LRO, 0,
 689                                0, &log, false, &tv);
 690         }
 691 }
 692 #endif
 693
 694 static inline void
 695 tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum)
 696 {
 697         uint32_t csum;
 698
 699         csum = 0xffff - *ptr + value;
 700         while (csum > 0xffff)
 701                 csum = (csum >> 16) + (csum & 0xffff);
 702         *ptr = value;
 703         *psum = csum;
 704 }
 705
 706 static uint16_t
 707 tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le,
 708     uint16_t payload_len, uint16_t delta_sum)
 709 {
 710         uint32_t csum;
 711         uint16_t tlen;
 712         uint16_t temp[5] = {};
 713
 714         switch (pa->data.lro_type) {
 715         case LRO_TYPE_IPV4_TCP:
 716                 /* Compute new IPv4 length. */
 717                 tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len;
 718                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
 719
 720                 /* Subtract delta from current IPv4 checksum. */
 721                 csum = pa->ip4->ip_sum + 0xffff - temp[0];
 722                 while (csum > 0xffff)
 723                         csum = (csum >> 16) + (csum & 0xffff);
 724                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
 725                 goto update_tcp_header;
 726
 727         case LRO_TYPE_IPV6_TCP:
 728                 /* Compute new IPv6 length. */
 729                 tlen = (pa->tcp->th_off << 2) + payload_len;
 730                 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
 731                 goto update_tcp_header;
 732
 733         case LRO_TYPE_IPV4_UDP:
 734                 /* Compute new IPv4 length. */
 735                 tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len;
 736                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]);
 737
 738                 /* Subtract delta from current IPv4 checksum. */
 739                 csum = pa->ip4->ip_sum + 0xffff - temp[0];
 740                 while (csum > 0xffff)
 741                         csum = (csum >> 16) + (csum & 0xffff);
 742                 tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]);
 743                 goto update_udp_header;
 744
 745         case LRO_TYPE_IPV6_UDP:
 746                 /* Compute new IPv6 length. */
 747                 tlen = sizeof(*pa->udp) + payload_len;
 748                 tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]);
 749                 goto update_udp_header;
 750
 751         default:
 752                 return (0);
 753         }
 754
 755 update_tcp_header:
 756         /* Compute current TCP header checksum. */
 757         temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp);
 758
 759         /* Incorporate the latest ACK into the TCP header. */
 760         pa->tcp->th_ack = le->ack_seq;
 761         pa->tcp->th_win = le->window;
 762
 763         /* Incorporate latest timestamp into the TCP header. */
 764         if (le->timestamp != 0) {
 765                 uint32_t *ts_ptr;
 766
 767                 ts_ptr = (uint32_t *)(pa->tcp + 1);
 768                 ts_ptr[1] = htonl(le->tsval);
 769                 ts_ptr[2] = le->tsecr;
 770         }
 771
 772         /* Compute new TCP header checksum. */
 773         temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp);
 774
 775         /* Compute new TCP checksum. */
 776         csum = pa->tcp->th_sum + 0xffff - delta_sum +
 777             0xffff - temp[0] + 0xffff - temp[3] + temp[2];
 778         while (csum > 0xffff)
 779                 csum = (csum >> 16) + (csum & 0xffff);
 780
 781         /* Assign new TCP checksum. */
 782         tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]);
 783
 784         /* Compute all modififications affecting next checksum. */
 785         csum = temp[0] + temp[1] + 0xffff - temp[2] +
 786             temp[3] + temp[4] + delta_sum;
 787         while (csum > 0xffff)
 788                 csum = (csum >> 16) + (csum & 0xffff);
 789
 790         /* Return delta checksum to next stage, if any. */
 791         return (csum);
 792
 793 update_udp_header:
 794         tlen = sizeof(*pa->udp) + payload_len;
 795         /* Assign new UDP length and compute checksum delta. */
 796         tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]);
 797
 798         /* Check if there is a UDP checksum. */
 799         if (__predict_false(pa->udp->uh_sum != 0)) {
 800                 /* Compute new UDP checksum. */
 801                 csum = pa->udp->uh_sum + 0xffff - delta_sum +
 802                     0xffff - temp[0] + 0xffff - temp[2];
 803                 while (csum > 0xffff)
 804                         csum = (csum >> 16) + (csum & 0xffff);
 805                 /* Assign new UDP checksum. */
 806                 tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]);
 807         }
 808
 809         /* Compute all modififications affecting next checksum. */
 810         csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum;
 811         while (csum > 0xffff)
 812                 csum = (csum >> 16) + (csum & 0xffff);
 813
 814         /* Return delta checksum to next stage, if any. */
 815         return (csum);
 816 }
 817
 818 static void
 819 tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le)
 820 {
 821         /* Check if we need to recompute any checksums. */
 822         if (le->m_head->m_pkthdr.lro_nsegs > 1) {
 823                 uint16_t csum;
 824
 825                 switch (le->inner.data.lro_type) {
 826                 case LRO_TYPE_IPV4_TCP:
 827                         csum = tcp_lro_update_checksum(&le->inner, le,
 828                             le->m_head->m_pkthdr.lro_tcp_d_len,
 829                             le->m_head->m_pkthdr.lro_tcp_d_csum);
 830                         csum = tcp_lro_update_checksum(&le->outer, NULL,
 831                             le->m_head->m_pkthdr.lro_tcp_d_len +
 832                             le->inner.total_hdr_len, csum);
 833                         le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 834                             CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
 835                         le->m_head->m_pkthdr.csum_data = 0xffff;
 836                         break;
 837                 case LRO_TYPE_IPV6_TCP:
 838                         csum = tcp_lro_update_checksum(&le->inner, le,
 839                             le->m_head->m_pkthdr.lro_tcp_d_len,
 840                             le->m_head->m_pkthdr.lro_tcp_d_csum);
 841                         csum = tcp_lro_update_checksum(&le->outer, NULL,
 842                             le->m_head->m_pkthdr.lro_tcp_d_len +
 843                             le->inner.total_hdr_len, csum);
 844                         le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 845                             CSUM_PSEUDO_HDR;
 846                         le->m_head->m_pkthdr.csum_data = 0xffff;
 847                         break;
 848                 case LRO_TYPE_NONE:
 849                         switch (le->outer.data.lro_type) {
 850                         case LRO_TYPE_IPV4_TCP:
 851                                 csum = tcp_lro_update_checksum(&le->outer, le,
 852                                     le->m_head->m_pkthdr.lro_tcp_d_len,
 853                                     le->m_head->m_pkthdr.lro_tcp_d_csum);
 854                                 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 855                                     CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID;
 856                                 le->m_head->m_pkthdr.csum_data = 0xffff;
 857                                 break;
 858                         case LRO_TYPE_IPV6_TCP:
 859                                 csum = tcp_lro_update_checksum(&le->outer, le,
 860                                     le->m_head->m_pkthdr.lro_tcp_d_len,
 861                                     le->m_head->m_pkthdr.lro_tcp_d_csum);
 862                                 le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID |
 863                                     CSUM_PSEUDO_HDR;
 864                                 le->m_head->m_pkthdr.csum_data = 0xffff;
 865                                 break;
 866                         default:
 867                                 break;
 868                         }
 869                         break;
 870                 default:
 871                         break;
 872                 }
 873         }
 874
 875         /*
 876          * Break any chain, this is not set to NULL on the singleton
 877          * case m_nextpkt points to m_head. Other case set them
 878          * m_nextpkt to NULL in push_and_replace.
 879          */
 880         le->m_head->m_nextpkt = NULL;
 881         lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs;
 882         (*lc->ifp->if_input)(lc->ifp, le->m_head);
 883 }
 884
 885 static void
 886 tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le,
 887     struct mbuf *m, struct tcphdr *th)
 888 {
 889         uint32_t *ts_ptr;
 890         uint16_t tcp_data_len;
 891         uint16_t tcp_opt_len;
 892
 893         ts_ptr = (uint32_t *)(th + 1);
 894         tcp_opt_len = (th->th_off << 2);
 895         tcp_opt_len -= sizeof(*th);
 896
 897         /* Check if there is a timestamp option. */
 898         if (tcp_opt_len == 0 ||
 899             __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
 900             *ts_ptr != TCP_LRO_TS_OPTION)) {
 901                 /* We failed to find the timestamp option. */
 902                 le->timestamp = 0;
 903         } else {
 904                 le->timestamp = 1;
 905                 le->tsval = ntohl(*(ts_ptr + 1));
 906                 le->tsecr = *(ts_ptr + 2);
 907         }
 908
 909         tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
 910
 911         /* Pull out TCP sequence numbers and window size. */
 912         le->next_seq = ntohl(th->th_seq) + tcp_data_len;
 913         le->ack_seq = th->th_ack;
 914         le->window = th->th_win;
 915
 916         /* Setup new data pointers. */
 917         le->m_head = m;
 918         le->m_tail = m_last(m);
 919 }
 920
 921 static void
 922 tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m)
 923 {
 924         struct lro_parser *pa;
 925
 926         /*
 927          * Push up the stack of the current entry
 928          * and replace it with "m".
 929          */
 930         struct mbuf *msave;
 931
 932         /* Grab off the next and save it */
 933         msave = le->m_head->m_nextpkt;
 934         le->m_head->m_nextpkt = NULL;
 935
 936         /* Now push out the old entry */
 937         tcp_flush_out_entry(lc, le);
 938
 939         /* Re-parse new header, should not fail. */
 940         pa = tcp_lro_parser(m, &le->outer, &le->inner, false);
 941         KASSERT(pa != NULL,
 942             ("tcp_push_and_replace: LRO parser failed on m=%p\n", m));
 943
 944         /*
 945          * Now to replace the data properly in the entry
 946          * we have to reset the TCP header and
 947          * other fields.
 948          */
 949         tcp_set_entry_to_mbuf(lc, le, m, pa->tcp);
 950
 951         /* Restore the next list */
 952         m->m_nextpkt = msave;
 953 }
 954
 955 static void
 956 tcp_lro_mbuf_append_pkthdr(struct mbuf *m, const struct mbuf *p)
 957 {
 958         uint32_t csum;
 959
 960         if (m->m_pkthdr.lro_nsegs == 1) {
 961                 /* Compute relative checksum. */
 962                 csum = p->m_pkthdr.lro_tcp_d_csum;
 963         } else {
 964                 /* Merge TCP data checksums. */
 965                 csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum +
 966                     (uint32_t)p->m_pkthdr.lro_tcp_d_csum;
 967                 while (csum > 0xffff)
 968                         csum = (csum >> 16) + (csum & 0xffff);
 969         }
 970
 971         /* Update various counters. */
 972         m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len;
 973         m->m_pkthdr.lro_tcp_d_csum = csum;
 974         m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len;
 975         m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs;
 976 }
 977
 978 static void
 979 tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le)
 980 {
 981         /*
 982          * Walk through the mbuf chain we
 983          * have on tap and compress/condense
 984          * as required.
 985          */
 986         uint32_t *ts_ptr;
 987         struct mbuf *m;
 988         struct tcphdr *th;
 989         uint32_t tcp_data_len_total;
 990         uint32_t tcp_data_seg_total;
 991         uint16_t tcp_data_len;
 992         uint16_t tcp_opt_len;
 993
 994         /*
 995          * First we must check the lead (m_head)
 996          * we must make sure that it is *not*
 997          * something that should be sent up
 998          * right away (sack etc).
 999          */
1000 again:
1001         m = le->m_head->m_nextpkt;
1002         if (m == NULL) {
1003                 /* Just one left. */
1004                 return;
1005         }
1006
1007         th = tcp_lro_get_th(m);
1008         tcp_opt_len = (th->th_off << 2);
1009         tcp_opt_len -= sizeof(*th);
1010         ts_ptr = (uint32_t *)(th + 1);
1011
1012         if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1013             *ts_ptr != TCP_LRO_TS_OPTION)) {
1014                 /*
1015                  * Its not the timestamp. We can't
1016                  * use this guy as the head.
1017                  */
1018                 le->m_head->m_nextpkt = m->m_nextpkt;
1019                 tcp_push_and_replace(lc, le, m);
1020                 goto again;
1021         }
1022         if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
1023                 /*
1024                  * Make sure that previously seen segements/ACKs are delivered
1025                  * before this segment, e.g. FIN.
1026                  */
1027                 le->m_head->m_nextpkt = m->m_nextpkt;
1028                 tcp_push_and_replace(lc, le, m);
1029                 goto again;
1030         }
1031         while((m = le->m_head->m_nextpkt) != NULL) {
1032                 /*
1033                  * condense m into le, first
1034                  * pull m out of the list.
1035                  */
1036                 le->m_head->m_nextpkt = m->m_nextpkt;
1037                 m->m_nextpkt = NULL;
1038                 /* Setup my data */
1039                 tcp_data_len = m->m_pkthdr.lro_tcp_d_len;
1040                 th = tcp_lro_get_th(m);
1041                 ts_ptr = (uint32_t *)(th + 1);
1042                 tcp_opt_len = (th->th_off << 2);
1043                 tcp_opt_len -= sizeof(*th);
1044                 tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len;
1045                 tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs;
1046
1047                 if (tcp_data_seg_total >= lc->lro_ackcnt_lim ||
1048                     tcp_data_len_total >= lc->lro_length_lim) {
1049                         /* Flush now if appending will result in overflow. */
1050                         tcp_push_and_replace(lc, le, m);
1051                         goto again;
1052                 }
1053                 if (tcp_opt_len != 0 &&
1054                     __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA ||
1055                     *ts_ptr != TCP_LRO_TS_OPTION)) {
1056                         /*
1057                          * Maybe a sack in the new one? We need to
1058                          * start all over after flushing the
1059                          * current le. We will go up to the beginning
1060                          * and flush it (calling the replace again possibly
1061                          * or just returning).
1062                          */
1063                         tcp_push_and_replace(lc, le, m);
1064                         goto again;
1065                 }
1066                 if ((th->th_flags & ~(TH_ACK | TH_PUSH)) != 0) {
1067                         tcp_push_and_replace(lc, le, m);
1068                         goto again;
1069                 }
1070                 if (tcp_opt_len != 0) {
1071                         uint32_t tsval = ntohl(*(ts_ptr + 1));
1072                         /* Make sure timestamp values are increasing. */
1073                         if (TSTMP_GT(le->tsval, tsval))  {
1074                                 tcp_push_and_replace(lc, le, m);
1075                                 goto again;
1076                         }
1077                         le->tsval = tsval;
1078                         le->tsecr = *(ts_ptr + 2);
1079                 }
1080                 /* Try to append the new segment. */
1081                 if (__predict_false(ntohl(th->th_seq) != le->next_seq ||
1082                                     (tcp_data_len == 0 &&
1083                                      le->ack_seq == th->th_ack &&
1084                                      le->window == th->th_win))) {
1085                         /* Out of order packet or duplicate ACK. */
1086                         tcp_push_and_replace(lc, le, m);
1087                         goto again;
1088                 }
1089                 if (tcp_data_len != 0 ||
1090                     SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1091                         le->next_seq += tcp_data_len;
1092                         le->ack_seq = th->th_ack;
1093                         le->window = th->th_win;
1094                 } else if (th->th_ack == le->ack_seq) {
1095                         le->window = WIN_MAX(le->window, th->th_win);
1096                 }
1097
1098                 if (tcp_data_len == 0) {
1099                         m_freem(m);
1100                         continue;
1101                 }
1102
1103                 /* Merge TCP data checksum and length to head mbuf. */
1104                 tcp_lro_mbuf_append_pkthdr(le->m_head, m);
1105
1106                 /*
1107                  * Adjust the mbuf so that m_data points to the first byte of
1108                  * the ULP payload.  Adjust the mbuf to avoid complications and
1109                  * append new segment to existing mbuf chain.
1110                  */
1111                 m_adj(m, m->m_pkthdr.len - tcp_data_len);
1112                 m_demote_pkthdr(m);
1113                 le->m_tail->m_next = m;
1114                 le->m_tail = m_last(m);
1115         }
1116 }
1117
1118 #ifdef TCPHPTS
1119 static void
1120 tcp_queue_pkts(struct inpcb *inp, struct tcpcb *tp, struct lro_entry *le)
1121 {
1122         INP_WLOCK_ASSERT(inp);
1123         if (tp->t_in_pkt == NULL) {
1124                 /* Nothing yet there */
1125                 tp->t_in_pkt = le->m_head;
1126                 tp->t_tail_pkt = le->m_last_mbuf;
1127         } else {
1128                 /* Already some there */
1129                 tp->t_tail_pkt->m_nextpkt = le->m_head;
1130                 tp->t_tail_pkt = le->m_last_mbuf;
1131         }
1132         le->m_head = NULL;
1133         le->m_last_mbuf = NULL;
1134 }
1135
1136 static struct mbuf *
1137 tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le,
1138     struct inpcb *inp, int32_t *new_m)
1139 {
1140         struct tcpcb *tp;
1141         struct mbuf *m;
1142
1143         tp = intotcpcb(inp);
1144         if (__predict_false(tp == NULL))
1145                 return (NULL);
1146
1147         /* Look at the last mbuf if any in queue */
1148         m = tp->t_tail_pkt;
1149         if (m != NULL && (m->m_flags & M_ACKCMP) != 0) {
1150                 if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) {
1151                         tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0);
1152                         *new_m = 0;
1153                         counter_u64_add(tcp_extra_mbuf, 1);
1154                         return (m);
1155                 } else {
1156                         /* Mark we ran out of space */
1157                         inp->inp_flags2 |= INP_MBUF_L_ACKS;
1158                 }
1159         }
1160         /* Decide mbuf size. */
1161         if (inp->inp_flags2 & INP_MBUF_L_ACKS)
1162                 m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR);
1163         else
1164                 m = m_gethdr(M_NOWAIT, MT_DATA);
1165
1166         if (__predict_false(m == NULL)) {
1167                 counter_u64_add(tcp_would_have_but, 1);
1168                 return (NULL);
1169         }
1170         counter_u64_add(tcp_comp_total, 1);
1171         m->m_flags |= M_ACKCMP;
1172         *new_m = 1;
1173         return (m);
1174 }
1175
1176 static struct inpcb *
1177 tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa)
1178 {
1179         struct inpcb *inp;
1180
1181         switch (pa->data.lro_type) {
1182 #ifdef INET6
1183         case LRO_TYPE_IPV6_TCP:
1184                 inp = in6_pcblookup(&V_tcbinfo,
1185                     &pa->data.s_addr.v6,
1186                     pa->data.s_port,
1187                     &pa->data.d_addr.v6,
1188                     pa->data.d_port,
1189                     INPLOOKUP_WLOCKPCB,
1190                     ifp);
1191                 break;
1192 #endif
1193 #ifdef INET
1194         case LRO_TYPE_IPV4_TCP:
1195                 inp = in_pcblookup(&V_tcbinfo,
1196                     pa->data.s_addr.v4,
1197                     pa->data.s_port,
1198                     pa->data.d_addr.v4,
1199                     pa->data.d_port,
1200                     INPLOOKUP_WLOCKPCB,
1201                     ifp);
1202                 break;
1203 #endif
1204         default:
1205                 inp = NULL;
1206                 break;
1207         }
1208         return (inp);
1209 }
1210
1211 static inline bool
1212 tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts)
1213 {
1214         /*
1215          * This function returns two bits of valuable information.
1216          * a) Is what is present capable of being ack-compressed,
1217          *    we can ack-compress if there is no options or just
1218          *    a timestamp option, and of course the th_flags must
1219          *    be correct as well.
1220          * b) Our other options present such as SACK. This is
1221          *    used to determine if we want to wakeup or not.
1222          */
1223         bool ret = true;
1224
1225         switch (th->th_off << 2) {
1226         case (sizeof(*th) + TCPOLEN_TSTAMP_APPA):
1227                 *ppts = (uint32_t *)(th + 1);
1228                 /* Check if we have only one timestamp option. */
1229                 if (**ppts == TCP_LRO_TS_OPTION)
1230                         *other_opts = false;
1231                 else {
1232                         *other_opts = true;
1233                         ret = false;
1234                 }
1235                 break;
1236         case (sizeof(*th)):
1237                 /* No options. */
1238                 *ppts = NULL;
1239                 *other_opts = false;
1240                 break;
1241         default:
1242                 *ppts = NULL;
1243                 *other_opts = true;
1244                 ret = false;
1245                 break;
1246         }
1247         /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */
1248         if ((th->th_flags & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0)
1249                 ret = false;
1250         /* If it has data on it we cannot compress it */
1251         if (m->m_pkthdr.lro_tcp_d_len)
1252                 ret = false;
1253
1254         /* ACK flag must be set. */
1255         if (!(th->th_flags & TH_ACK))
1256                 ret = false;
1257         return (ret);
1258 }
1259
1260 static int
1261 tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le)
1262 {
1263         struct inpcb *inp;
1264         struct tcpcb *tp;
1265         struct mbuf **pp, *cmp, *mv_to;
1266         bool bpf_req, should_wake;
1267
1268         /* Check if packet doesn't belongs to our network interface. */
1269         if ((tcplro_stacks_wanting_mbufq == 0) ||
1270             (le->outer.data.vlan_id != 0) ||
1271             (le->inner.data.lro_type != LRO_TYPE_NONE))
1272                 return (TCP_LRO_CANNOT);
1273
1274 #ifdef INET6
1275         /*
1276          * Be proactive about unspecified IPv6 address in source. As
1277          * we use all-zero to indicate unbounded/unconnected pcb,
1278          * unspecified IPv6 address can be used to confuse us.
1279          *
1280          * Note that packets with unspecified IPv6 destination is
1281          * already dropped in ip6_input.
1282          */
1283         if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP &&
1284             IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6)))
1285                 return (TCP_LRO_CANNOT);
1286
1287         if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP &&
1288             IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6)))
1289                 return (TCP_LRO_CANNOT);
1290 #endif
1291         /* Lookup inp, if any. */
1292         inp = tcp_lro_lookup(lc->ifp,
1293             (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner);
1294         if (inp == NULL)
1295                 return (TCP_LRO_CANNOT);
1296
1297         counter_u64_add(tcp_inp_lro_locks_taken, 1);
1298
1299         /* Get TCP control structure. */
1300         tp = intotcpcb(inp);
1301
1302         /* Check if the inp is dead, Jim. */
1303         if (tp == NULL ||
1304             (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) ||
1305             (inp->inp_flags2 & INP_FREED)) {
1306                 INP_WUNLOCK(inp);
1307                 return (TCP_LRO_CANNOT);
1308         }
1309         if ((inp->inp_irq_cpu_set == 0)  && (lc->lro_cpu_is_set == 1)) {
1310                 inp->inp_irq_cpu = lc->lro_last_cpu;
1311                 inp->inp_irq_cpu_set = 1;
1312         }
1313         /* Check if the transport doesn't support the needed optimizations. */
1314         if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) {
1315                 INP_WUNLOCK(inp);
1316                 return (TCP_LRO_CANNOT);
1317         }
1318
1319         if (inp->inp_flags2 & INP_MBUF_QUEUE_READY)
1320                 should_wake = false;
1321         else
1322                 should_wake = true;
1323         /* Check if packets should be tapped to BPF. */
1324         bpf_req = bpf_peers_present(lc->ifp->if_bpf);
1325
1326         /* Strip and compress all the incoming packets. */
1327         cmp = NULL;
1328         for (pp = &le->m_head; *pp != NULL; ) {
1329                 mv_to = NULL;
1330                 if (do_bpf_strip_and_compress(inp, lc, le, pp,
1331                          &cmp, &mv_to, &should_wake, bpf_req ) == false) {
1332                         /* Advance to next mbuf. */
1333                         pp = &(*pp)->m_nextpkt;
1334                 } else if (mv_to != NULL) {
1335                         /* We are asked to move pp up */
1336                         pp = &mv_to->m_nextpkt;
1337                 }
1338         }
1339         /* Update "m_last_mbuf", if any. */
1340         if (pp == &le->m_head)
1341                 le->m_last_mbuf = *pp;
1342         else
1343                 le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt);
1344
1345         /* Check if any data mbufs left. */
1346         if (le->m_head != NULL) {
1347                 counter_u64_add(tcp_inp_lro_direct_queue, 1);
1348                 tcp_lro_log(tp, lc, le, NULL, 22, 1,
1349                             inp->inp_flags2, inp->inp_in_input, 1);
1350                 tcp_queue_pkts(inp, tp, le);
1351         }
1352         if (should_wake) {
1353                 /* Wakeup */
1354                 counter_u64_add(tcp_inp_lro_wokeup_queue, 1);
1355                 if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0))
1356                         inp = NULL;
1357         }
1358         if (inp != NULL)
1359                 INP_WUNLOCK(inp);
1360         return (0);     /* Success. */
1361 }
1362 #endif
1363
1364 void
1365 tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
1366 {
1367         /* Only optimise if there are multiple packets waiting. */
1368 #ifdef TCPHPTS
1369         int error;
1370 #endif
1371
1372         NET_EPOCH_ASSERT();
1373 #ifdef TCPHPTS
1374         CURVNET_SET(lc->ifp->if_vnet);
1375         error = tcp_lro_flush_tcphpts(lc, le);
1376         CURVNET_RESTORE();
1377         if (error != 0) {
1378 #endif
1379                 tcp_lro_condense(lc, le);
1380                 tcp_flush_out_entry(lc, le);
1381 #ifdef TCPHPTS
1382         }
1383 #endif
1384         lc->lro_flushed++;
1385         bzero(le, sizeof(*le));
1386         LIST_INSERT_HEAD(&lc->lro_free, le, next);
1387 }
1388
1389 #ifdef HAVE_INLINE_FLSLL
1390 #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1))
1391 #else
1392 static inline uint64_t
1393 tcp_lro_msb_64(uint64_t x)
1394 {
1395         x |= (x >> 1);
1396         x |= (x >> 2);
1397         x |= (x >> 4);
1398         x |= (x >> 8);
1399         x |= (x >> 16);
1400         x |= (x >> 32);
1401         return (x & ~(x >> 1));
1402 }
1403 #endif
1404
1405 /*
1406  * The tcp_lro_sort() routine is comparable to qsort(), except it has
1407  * a worst case complexity limit of O(MIN(N,64)*N), where N is the
1408  * number of elements to sort and 64 is the number of sequence bits
1409  * available. The algorithm is bit-slicing the 64-bit sequence number,
1410  * sorting one bit at a time from the most significant bit until the
1411  * least significant one, skipping the constant bits. This is
1412  * typically called a radix sort.
1413  */
1414 static void
1415 tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size)
1416 {
1417         struct lro_mbuf_sort temp;
1418         uint64_t ones;
1419         uint64_t zeros;
1420         uint32_t x;
1421         uint32_t y;
1422
1423 repeat:
1424         /* for small arrays insertion sort is faster */
1425         if (size <= 12) {
1426                 for (x = 1; x < size; x++) {
1427                         temp = parray[x];
1428                         for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--)
1429                                 parray[y] = parray[y - 1];
1430                         parray[y] = temp;
1431                 }
1432                 return;
1433         }
1434
1435         /* compute sequence bits which are constant */
1436         ones = 0;
1437         zeros = 0;
1438         for (x = 0; x != size; x++) {
1439                 ones |= parray[x].seq;
1440                 zeros |= ~parray[x].seq;
1441         }
1442
1443         /* compute bits which are not constant into "ones" */
1444         ones &= zeros;
1445         if (ones == 0)
1446                 return;
1447
1448         /* pick the most significant bit which is not constant */
1449         ones = tcp_lro_msb_64(ones);
1450
1451         /*
1452          * Move entries having cleared sequence bits to the beginning
1453          * of the array:
1454          */
1455         for (x = y = 0; y != size; y++) {
1456                 /* skip set bits */
1457                 if (parray[y].seq & ones)
1458                         continue;
1459                 /* swap entries */
1460                 temp = parray[x];
1461                 parray[x] = parray[y];
1462                 parray[y] = temp;
1463                 x++;
1464         }
1465
1466         KASSERT(x != 0 && x != size, ("Memory is corrupted\n"));
1467
1468         /* sort zeros */
1469         tcp_lro_sort(parray, x);
1470
1471         /* sort ones */
1472         parray += x;
1473         size -= x;
1474         goto repeat;
1475 }
1476
1477 void
1478 tcp_lro_flush_all(struct lro_ctrl *lc)
1479 {
1480         uint64_t seq;
1481         uint64_t nseq;
1482         unsigned x;
1483
1484         NET_EPOCH_ASSERT();
1485         /* check if no mbufs to flush */
1486         if (lc->lro_mbuf_count == 0)
1487                 goto done;
1488         if (lc->lro_cpu_is_set == 0) {
1489                 if (lc->lro_last_cpu == curcpu) {
1490                         lc->lro_cnt_of_same_cpu++;
1491                         /* Have we reached the threshold to declare a cpu? */
1492                         if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh)
1493                                 lc->lro_cpu_is_set = 1;
1494                 } else {
1495                         lc->lro_last_cpu = curcpu;
1496                         lc->lro_cnt_of_same_cpu = 0;
1497                 }
1498         }
1499         CURVNET_SET(lc->ifp->if_vnet);
1500
1501         /* get current time */
1502         binuptime(&lc->lro_last_queue_time);
1503
1504         /* sort all mbufs according to stream */
1505         tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count);
1506
1507         /* input data into LRO engine, stream by stream */
1508         seq = 0;
1509         for (x = 0; x != lc->lro_mbuf_count; x++) {
1510                 struct mbuf *mb;
1511
1512                 /* get mbuf */
1513                 mb = lc->lro_mbuf_data[x].mb;
1514
1515                 /* get sequence number, masking away the packet index */
1516                 nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24);
1517
1518                 /* check for new stream */
1519                 if (seq != nseq) {
1520                         seq = nseq;
1521
1522                         /* flush active streams */
1523                         tcp_lro_rx_done(lc);
1524                 }
1525
1526                 /* add packet to LRO engine */
1527                 if (tcp_lro_rx_common(lc, mb, 0, false) != 0) {
1528                         /* input packet to network layer */
1529                         (*lc->ifp->if_input)(lc->ifp, mb);
1530                         lc->lro_queued++;
1531                         lc->lro_flushed++;
1532                 }
1533         }
1534         CURVNET_RESTORE();
1535 done:
1536         /* flush active streams */
1537         tcp_lro_rx_done(lc);
1538
1539 #ifdef TCPHPTS
1540         tcp_run_hpts();
1541 #endif
1542         lc->lro_mbuf_count = 0;
1543 }
1544
1545 #ifdef TCPHPTS
1546 static void
1547 build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m,
1548     uint32_t *ts_ptr, uint16_t iptos)
1549 {
1550         /*
1551          * Given a TCP ACK, summarize it down into the small TCP ACK
1552          * entry.
1553          */
1554         ae->timestamp = m->m_pkthdr.rcv_tstmp;
1555         if (m->m_flags & M_TSTMP_LRO)
1556                 ae->flags = TSTMP_LRO;
1557         else if (m->m_flags & M_TSTMP)
1558                 ae->flags = TSTMP_HDWR;
1559         ae->seq = ntohl(th->th_seq);
1560         ae->ack = ntohl(th->th_ack);
1561         ae->flags |= th->th_flags;
1562         if (ts_ptr != NULL) {
1563                 ae->ts_value = ntohl(ts_ptr[1]);
1564                 ae->ts_echo = ntohl(ts_ptr[2]);
1565                 ae->flags |= HAS_TSTMP;
1566         }
1567         ae->win = ntohs(th->th_win);
1568         ae->codepoint = iptos;
1569 }
1570
1571 /*
1572  * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets
1573  * and strip all, but the IPv4/IPv6 header.
1574  */
1575 static bool
1576 do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc,
1577     struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to,
1578     bool *should_wake, bool bpf_req)
1579 {
1580         union {
1581                 void *ptr;
1582                 struct ip *ip4;
1583                 struct ip6_hdr *ip6;
1584         } l3;
1585         struct mbuf *m;
1586         struct mbuf *nm;
1587         struct tcphdr *th;
1588         struct tcp_ackent *ack_ent;
1589         uint32_t *ts_ptr;
1590         int32_t n_mbuf;
1591         bool other_opts, can_compress;
1592         uint8_t lro_type;
1593         uint16_t iptos;
1594         int tcp_hdr_offset;
1595         int idx;
1596
1597         /* Get current mbuf. */
1598         m = *pp;
1599
1600         /* Let the BPF see the packet */
1601         if (__predict_false(bpf_req))
1602                 ETHER_BPF_MTAP(lc->ifp, m);
1603
1604         tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off;
1605         lro_type = le->inner.data.lro_type;
1606         switch (lro_type) {
1607         case LRO_TYPE_NONE:
1608                 lro_type = le->outer.data.lro_type;
1609                 switch (lro_type) {
1610                 case LRO_TYPE_IPV4_TCP:
1611                         tcp_hdr_offset -= sizeof(*le->outer.ip4);
1612                         m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1613                         break;
1614                 case LRO_TYPE_IPV6_TCP:
1615                         tcp_hdr_offset -= sizeof(*le->outer.ip6);
1616                         m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1617                         break;
1618                 default:
1619                         goto compressed;
1620                 }
1621                 break;
1622         case LRO_TYPE_IPV4_TCP:
1623                 tcp_hdr_offset -= sizeof(*le->outer.ip4);
1624                 m->m_pkthdr.lro_etype = ETHERTYPE_IP;
1625                 break;
1626         case LRO_TYPE_IPV6_TCP:
1627                 tcp_hdr_offset -= sizeof(*le->outer.ip6);
1628                 m->m_pkthdr.lro_etype = ETHERTYPE_IPV6;
1629                 break;
1630         default:
1631                 goto compressed;
1632         }
1633
1634         MPASS(tcp_hdr_offset >= 0);
1635
1636         m_adj(m, tcp_hdr_offset);
1637         m->m_flags |= M_LRO_EHDRSTRP;
1638         m->m_flags &= ~M_ACKCMP;
1639         m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset;
1640
1641         th = tcp_lro_get_th(m);
1642
1643         th->th_sum = 0;         /* TCP checksum is valid. */
1644
1645         /* Check if ACK can be compressed */
1646         can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts);
1647
1648         /* Now lets look at the should wake states */
1649         if ((other_opts == true) &&
1650             ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) {
1651                 /*
1652                  * If there are other options (SACK?) and the
1653                  * tcp endpoint has not expressly told us it does
1654                  * not care about SACKS, then we should wake up.
1655                  */
1656                 *should_wake = true;
1657         }
1658         /* Is the ack compressable? */
1659         if (can_compress == false)
1660                 goto done;
1661         /* Does the TCP endpoint support ACK compression? */
1662         if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0)
1663                 goto done;
1664
1665         /* Lets get the TOS/traffic class field */
1666         l3.ptr = mtod(m, void *);
1667         switch (lro_type) {
1668         case LRO_TYPE_IPV4_TCP:
1669                 iptos = l3.ip4->ip_tos;
1670                 break;
1671         case LRO_TYPE_IPV6_TCP:
1672                 iptos = IPV6_TRAFFIC_CLASS(l3.ip6);
1673                 break;
1674         default:
1675                 iptos = 0;      /* Keep compiler happy. */
1676                 break;
1677         }
1678         /* Now lets get space if we don't have some already */
1679         if (*cmp == NULL) {
1680 new_one:
1681                 nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf);
1682                 if (__predict_false(nm == NULL))
1683                         goto done;
1684                 *cmp = nm;
1685                 if (n_mbuf) {
1686                         /*
1687                          *  Link in the new cmp ack to our in-order place,
1688                          * first set our cmp ack's next to where we are.
1689                          */
1690                         nm->m_nextpkt = m;
1691                         (*pp) = nm;
1692                         /*
1693                          * Set it up so mv_to is advanced to our
1694                          * compressed ack. This way the caller can
1695                          * advance pp to the right place.
1696                          */
1697                         *mv_to = nm;
1698                         /*
1699                          * Advance it here locally as well.
1700                          */
1701                         pp = &nm->m_nextpkt;
1702                 }
1703         } else {
1704                 /* We have one already we are working on */
1705                 nm = *cmp;
1706                 if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) {
1707                         /* We ran out of space */
1708                         inp->inp_flags2 |= INP_MBUF_L_ACKS;
1709                         goto new_one;
1710                 }
1711         }
1712         MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent));
1713         counter_u64_add(tcp_inp_lro_compressed, 1);
1714         le->compressed++;
1715         /* We can add in to the one on the tail */
1716         ack_ent = mtod(nm, struct tcp_ackent *);
1717         idx = (nm->m_len / sizeof(struct tcp_ackent));
1718         build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos);
1719
1720         /* Bump the size of both pkt-hdr and len */
1721         nm->m_len += sizeof(struct tcp_ackent);
1722         nm->m_pkthdr.len += sizeof(struct tcp_ackent);
1723 compressed:
1724         /* Advance to next mbuf before freeing. */
1725         *pp = m->m_nextpkt;
1726         m->m_nextpkt = NULL;
1727         m_freem(m);
1728         return (true);
1729 done:
1730         counter_u64_add(tcp_uncomp_total, 1);
1731         le->uncompressed++;
1732         return (false);
1733 }
1734 #endif
1735
1736 static struct lro_head *
1737 tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser)
1738 {
1739         u_long hash;
1740
1741         if (M_HASHTYPE_ISHASH(m)) {
1742                 hash = m->m_pkthdr.flowid;
1743         } else {
1744                 for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++)
1745                         hash += parser->data.raw[i];
1746         }
1747         return (&lc->lro_hash[hash % lc->lro_hashsz]);
1748 }
1749
1750 static int
1751 tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash)
1752 {
1753         struct lro_parser pi;   /* inner address data */
1754         struct lro_parser po;   /* outer address data */
1755         struct lro_parser *pa;  /* current parser for TCP stream */
1756         struct lro_entry *le;
1757         struct lro_head *bucket;
1758         struct tcphdr *th;
1759         int tcp_data_len;
1760         int tcp_opt_len;
1761         int error;
1762         uint16_t tcp_data_sum;
1763
1764 #ifdef INET
1765         /* Quickly decide if packet cannot be LRO'ed */
1766         if (__predict_false(V_ipforwarding != 0))
1767                 return (TCP_LRO_CANNOT);
1768 #endif
1769 #ifdef INET6
1770         /* Quickly decide if packet cannot be LRO'ed */
1771         if (__predict_false(V_ip6_forwarding != 0))
1772                 return (TCP_LRO_CANNOT);
1773 #endif
1774         if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1775              ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1776             (m->m_pkthdr.csum_data != 0xffff)) {
1777                 /*
1778                  * The checksum either did not have hardware offload
1779                  * or it was a bad checksum. We can't LRO such
1780                  * a packet.
1781                  */
1782                 counter_u64_add(tcp_bad_csums, 1);
1783                 return (TCP_LRO_CANNOT);
1784         }
1785         /* We expect a contiguous header [eh, ip, tcp]. */
1786         pa = tcp_lro_parser(m, &po, &pi, true);
1787         if (__predict_false(pa == NULL))
1788                 return (TCP_LRO_NOT_SUPPORTED);
1789
1790         /* We don't expect any padding. */
1791         error = tcp_lro_trim_mbuf_chain(m, pa);
1792         if (__predict_false(error != 0))
1793                 return (error);
1794
1795 #ifdef INET
1796         switch (pa->data.lro_type) {
1797         case LRO_TYPE_IPV4_TCP:
1798                 error = tcp_lro_rx_ipv4(lc, m, pa->ip4);
1799                 if (__predict_false(error != 0))
1800                         return (error);
1801                 break;
1802         default:
1803                 break;
1804         }
1805 #endif
1806         /* If no hardware or arrival stamp on the packet add timestamp */
1807         if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) {
1808                 m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time);
1809                 m->m_flags |= M_TSTMP_LRO;
1810         }
1811
1812         /* Get pointer to TCP header. */
1813         th = pa->tcp;
1814
1815         /* Don't process SYN packets. */
1816         if (__predict_false(th->th_flags & TH_SYN))
1817                 return (TCP_LRO_CANNOT);
1818
1819         /* Get total TCP header length and compute payload length. */
1820         tcp_opt_len = (th->th_off << 2);
1821         tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th -
1822             (uint8_t *)m->m_data) - tcp_opt_len;
1823         tcp_opt_len -= sizeof(*th);
1824
1825         /* Don't process invalid TCP headers. */
1826         if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0))
1827                 return (TCP_LRO_CANNOT);
1828
1829         /* Compute TCP data only checksum. */
1830         if (tcp_data_len == 0)
1831                 tcp_data_sum = 0;       /* no data, no checksum */
1832         else if (__predict_false(csum != 0))
1833                 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum);
1834         else
1835                 tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum);
1836
1837         /* Save TCP info in mbuf. */
1838         m->m_nextpkt = NULL;
1839         m->m_pkthdr.rcvif = lc->ifp;
1840         m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum;
1841         m->m_pkthdr.lro_tcp_d_len = tcp_data_len;
1842         m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data);
1843         m->m_pkthdr.lro_nsegs = 1;
1844
1845         /* Get hash bucket. */
1846         if (!use_hash) {
1847                 bucket = &lc->lro_hash[0];
1848         } else {
1849                 bucket = tcp_lro_rx_get_bucket(lc, m, pa);
1850         }
1851
1852         /* Try to find a matching previous segment. */
1853         LIST_FOREACH(le, bucket, hash_next) {
1854                 /* Compare addresses and ports. */
1855                 if (lro_address_compare(&po.data, &le->outer.data) == false ||
1856                     lro_address_compare(&pi.data, &le->inner.data) == false)
1857                         continue;
1858
1859                 /* Check if no data and old ACK. */
1860                 if (tcp_data_len == 0 &&
1861                     SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) {
1862                         m_freem(m);
1863                         return (0);
1864                 }
1865
1866                 /* Mark "m" in the last spot. */
1867                 le->m_last_mbuf->m_nextpkt = m;
1868                 /* Now set the tail to "m". */
1869                 le->m_last_mbuf = m;
1870                 return (0);
1871         }
1872
1873         /* Try to find an empty slot. */
1874         if (LIST_EMPTY(&lc->lro_free))
1875                 return (TCP_LRO_NO_ENTRIES);
1876
1877         /* Start a new segment chain. */
1878         le = LIST_FIRST(&lc->lro_free);
1879         LIST_REMOVE(le, next);
1880         tcp_lro_active_insert(lc, bucket, le);
1881
1882         /* Make sure the headers are set. */
1883         le->inner = pi;
1884         le->outer = po;
1885
1886         /* Store time this entry was allocated. */
1887         le->alloc_time = lc->lro_last_queue_time;
1888
1889         tcp_set_entry_to_mbuf(lc, le, m, th);
1890
1891         /* Now set the tail to "m". */
1892         le->m_last_mbuf = m;
1893
1894         return (0);
1895 }
1896
1897 int
1898 tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum)
1899 {
1900         int error;
1901
1902         if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) !=
1903              ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) ||
1904             (m->m_pkthdr.csum_data != 0xffff)) {
1905                 /*
1906                  * The checksum either did not have hardware offload
1907                  * or it was a bad checksum. We can't LRO such
1908                  * a packet.
1909                  */
1910                 counter_u64_add(tcp_bad_csums, 1);
1911                 return (TCP_LRO_CANNOT);
1912         }
1913         /* get current time */
1914         binuptime(&lc->lro_last_queue_time);
1915         CURVNET_SET(lc->ifp->if_vnet);
1916         error = tcp_lro_rx_common(lc, m, csum, true);
1917         CURVNET_RESTORE();
1918
1919         return (error);
1920 }
1921
1922 void
1923 tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb)
1924 {
1925         NET_EPOCH_ASSERT();
1926         /* sanity checks */
1927         if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL ||
1928             lc->lro_mbuf_max == 0)) {
1929                 /* packet drop */
1930                 m_freem(mb);
1931                 return;
1932         }
1933
1934         /* check if packet is not LRO capable */
1935         if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) {
1936                 /* input packet to network layer */
1937                 (*lc->ifp->if_input) (lc->ifp, mb);
1938                 return;
1939         }
1940
1941         /* create sequence number */
1942         lc->lro_mbuf_data[lc->lro_mbuf_count].seq =
1943             (((uint64_t)M_HASHTYPE_GET(mb)) << 56) |
1944             (((uint64_t)mb->m_pkthdr.flowid) << 24) |
1945             ((uint64_t)lc->lro_mbuf_count);
1946
1947         /* enter mbuf */
1948         lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb;
1949
1950         /* flush if array is full */
1951         if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max))
1952                 tcp_lro_flush_all(lc);
1953 }
1954
1955 /* end */