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