bsd/netinet/tcp_lro.c

   1 /*
   2  * Copyright (c) 2011-2013 Apple Inc. All rights reserved.
   3  *
   4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
   5  *
   6  * This file contains Original Code and/or Modifications of Original Code
   7  * as defined in and that are subject to the Apple Public Source License
   8  * Version 2.0 (the 'License'). You may not use this file except in
   9  * compliance with the License. The rights granted to you under the License
  10  * may not be used to create, or enable the creation or redistribution of,
  11  * unlawful or unlicensed copies of an Apple operating system, or to
  12  * circumvent, violate, or enable the circumvention or violation of, any
  13  * terms of an Apple operating system software license agreement.
  14  *
  15  * Please obtain a copy of the License at
  16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
  17  *
  18  * The Original Code and all software distributed under the License are
  19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
  23  * Please see the License for the specific language governing rights and
  24  * limitations under the License.
  25  *
  26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
  27  */
  28
  29 #include <sys/param.h>
  30 #include <sys/systm.h>
  31 #include <sys/sysctl.h>
  32 #include <sys/mbuf.h>
  33 #include <sys/mcache.h>
  34 #include <sys/socket.h>
  35 #include <sys/socketvar.h>
  36 #include <net/if_types.h>
  37 #include <net/route.h>
  38 #include <netinet/in.h>
  39 #include <netinet/in_systm.h>
  40 #include <net/if.h>
  41 #include <net/dlil.h>
  42 #include <netinet/ip.h>
  43 #include <netinet/ip_var.h>
  44 #include <netinet/in_var.h>
  45 #include <netinet/tcp.h>
  46 #include <netinet/tcp_seq.h>
  47 #include <netinet/tcpip.h>
  48 #include <netinet/tcp_var.h>
  49 #include <netinet/tcp_lro.h>
  50 #include <netinet/lro_ext.h>
  51 #include <kern/locks.h>
  52
  53 unsigned int lrocount = 0; /* A counter used for debugging only */
  54 unsigned int lro_seq_outoforder = 0; /* Counter for debugging */
  55 unsigned int lro_seq_mismatch = 0; /* Counter for debugging */
  56 unsigned int lro_flushes = 0; /* Counter for tracking number of flushes */
  57 unsigned int lro_single_flushes = 0;
  58 unsigned int lro_double_flushes = 0;
  59 unsigned int lro_good_flushes = 0;
  60
  61 unsigned int coalesc_sz = LRO_MX_COALESCE_PKTS;
  62 SYSCTL_INT(_net_inet_tcp, OID_AUTO, lro_sz, CTLFLAG_RW | CTLFLAG_LOCKED,
  63                 &coalesc_sz, 0, "Max coalescing size");
  64
  65 unsigned int coalesc_time = LRO_MX_TIME_TO_BUFFER;
  66 SYSCTL_INT(_net_inet_tcp, OID_AUTO, lro_time, CTLFLAG_RW | CTLFLAG_LOCKED,
  67                 &coalesc_time, 0, "Max coalescing time");
  68
  69 struct lro_flow lro_flow_list[TCP_LRO_NUM_FLOWS];
  70
  71 char lro_flow_map[TCP_LRO_FLOW_MAP];
  72
  73 static lck_attr_t *tcp_lro_mtx_attr = NULL;             /* mutex attributes */
  74 static lck_grp_t *tcp_lro_mtx_grp = NULL;               /* mutex group */
  75 static lck_grp_attr_t *tcp_lro_mtx_grp_attr = NULL;     /* mutex group attrs */
  76 decl_lck_mtx_data( ,tcp_lro_lock);      /* Used to synchronize updates */
  77
  78 unsigned int lro_byte_count = 0;
  79
  80 uint64_t lro_deadline = 0; /* LRO's sense of time - protected by tcp_lro_lock */
  81 uint32_t lro_timer_set = 0;
  82
  83 /* Some LRO stats */
  84 u_int32_t lro_pkt_count = 0; /* Number of packets encountered in an LRO period */
  85 thread_call_t tcp_lro_timer;
  86
  87 extern u_int32_t kipf_count;
  88
  89 static void     tcp_lro_timer_proc(void*, void*);
  90 static void     lro_update_stats(struct mbuf*);
  91 static void     lro_update_flush_stats(struct mbuf *);
  92 static void     tcp_lro_flush_flows(void);
  93 static void     tcp_lro_sched_timer(uint64_t);
  94 static void     lro_proto_input(struct mbuf *);
  95
  96 static struct mbuf *lro_tcp_xsum_validate(struct mbuf*,  struct ip *,
  97                                 struct tcphdr*);
  98 static struct mbuf *tcp_lro_process_pkt(struct mbuf*, struct ip*, struct tcphdr*,
  99                                 int);
 100
 101 void
 102 tcp_lro_init(void)
 103 {
 104         int i;
 105
 106         bzero(lro_flow_list, sizeof (struct lro_flow) * TCP_LRO_NUM_FLOWS);
 107         for (i = 0; i < TCP_LRO_FLOW_MAP; i++) {
 108                 lro_flow_map[i] = TCP_LRO_FLOW_UNINIT;
 109         }
 110
 111         /*
 112          * allocate lock group attribute, group and attribute for tcp_lro_lock
 113          */
 114         tcp_lro_mtx_grp_attr = lck_grp_attr_alloc_init();
 115         tcp_lro_mtx_grp = lck_grp_alloc_init("tcplro", tcp_lro_mtx_grp_attr);
 116         tcp_lro_mtx_attr = lck_attr_alloc_init();
 117         lck_mtx_init(&tcp_lro_lock, tcp_lro_mtx_grp, tcp_lro_mtx_attr);
 118
 119         tcp_lro_timer = thread_call_allocate(tcp_lro_timer_proc, NULL);
 120         if (tcp_lro_timer == NULL) {
 121                 panic_plain("%s: unable to allocate lro timer", __func__);
 122         }
 123
 124         return;
 125 }
 126
 127 static int
 128 tcp_lro_matching_tuple(struct ip* ip_hdr, struct tcphdr *tcp_hdr, int *hash,
 129                         int *flow_id )
 130 {
 131         struct lro_flow *flow;
 132         tcp_seq seqnum;
 133         unsigned int off = 0;
 134         int payload_len = 0;
 135
 136         *hash = LRO_HASH(ip_hdr->ip_src.s_addr, ip_hdr->ip_dst.s_addr,
 137                 tcp_hdr->th_sport, tcp_hdr->th_dport, (TCP_LRO_FLOW_MAP - 1));
 138
 139         *flow_id = lro_flow_map[*hash];
 140         if (*flow_id == TCP_LRO_FLOW_NOTFOUND) {
 141                 return TCP_LRO_NAN;
 142         }
 143
 144         seqnum = tcp_hdr->th_seq;
 145         off = tcp_hdr->th_off << 2;
 146         payload_len = ip_hdr->ip_len - off;
 147
 148         flow = &lro_flow_list[*flow_id];
 149
 150         if ((flow->lr_faddr.s_addr == ip_hdr->ip_src.s_addr) &&
 151                         (flow->lr_laddr.s_addr == ip_hdr->ip_dst.s_addr) &&
 152                         (flow->lr_fport == tcp_hdr->th_sport) &&
 153                         (flow->lr_lport == tcp_hdr->th_dport)) {
 154                 if (flow->lr_tcphdr == NULL) {
 155                         if (ntohl(seqnum) == flow->lr_seq) {
 156                                 return TCP_LRO_COALESCE;
 157                         }
 158                         if (lrodebug >= 4) {
 159                                 printf("%s: seqnum = %x, lr_seq = %x\n",
 160                                         __func__, ntohl(seqnum), flow->lr_seq);
 161                         }
 162                         lro_seq_mismatch++;
 163                         if (SEQ_GT(ntohl(seqnum), flow->lr_seq)) {
 164                                 lro_seq_outoforder++;
 165                                 /*
 166                                  * Whenever we receive out of order packets it
 167                                  * signals loss and recovery and LRO doesn't
 168                                  * let flows recover quickly. So eject.
 169                                  */
 170                                  flow->lr_flags |= LRO_EJECT_REQ;
 171
 172                         }
 173                         return TCP_LRO_NAN;
 174                 }
 175
 176                 if (flow->lr_flags & LRO_EJECT_REQ) {
 177                         if (lrodebug)
 178                                 printf("%s: eject. \n", __func__);
 179                         return TCP_LRO_EJECT_FLOW;
 180                 }
 181                 if (SEQ_GT(tcp_hdr->th_ack, flow->lr_tcphdr->th_ack)) {
 182                         if (lrodebug) {
 183                                 printf("%s: th_ack = %x flow_ack = %x \n",
 184                                         __func__, tcp_hdr->th_ack,
 185                                         flow->lr_tcphdr->th_ack);
 186                         }
 187                         return TCP_LRO_EJECT_FLOW;
 188                 }
 189
 190                 if (ntohl(seqnum) == (ntohl(lro_flow_list[*flow_id].lr_tcphdr->th_seq) + lro_flow_list[*flow_id].lr_len)) {
 191                         return TCP_LRO_COALESCE;
 192                 } else {
 193                         /* LRO does not handle loss recovery well, eject */
 194                         flow->lr_flags |= LRO_EJECT_REQ;
 195                         return TCP_LRO_EJECT_FLOW;
 196                 }
 197         }
 198         if (lrodebug) printf("tcp_lro_matching_tuple: collision \n");
 199         return TCP_LRO_COLLISION;
 200 }
 201
 202 static void
 203 tcp_lro_init_flow(int flow_id, struct ip* ip_hdr, struct tcphdr *tcp_hdr,
 204                         int hash, u_int32_t timestamp, int payload_len)
 205 {
 206         struct lro_flow *flow = NULL;
 207
 208         flow = &lro_flow_list[flow_id];
 209
 210         flow->lr_hash_map = hash;
 211         flow->lr_faddr.s_addr = ip_hdr->ip_src.s_addr;
 212         flow->lr_laddr.s_addr = ip_hdr->ip_dst.s_addr;
 213         flow->lr_fport = tcp_hdr->th_sport;
 214         flow->lr_lport = tcp_hdr->th_dport;
 215         lro_flow_map[hash] = flow_id;
 216         flow->lr_timestamp = timestamp;
 217         flow->lr_seq = ntohl(tcp_hdr->th_seq) + payload_len;
 218         flow->lr_flags = 0;
 219         return;
 220 }
 221
 222 static void
 223 tcp_lro_coalesce(int flow_id, struct mbuf *lro_mb, struct tcphdr *tcphdr,
 224                         int payload_len, int drop_hdrlen, struct tcpopt *topt,
 225                         u_int32_t* tsval, u_int32_t* tsecr, int thflags)
 226 {
 227         struct lro_flow *flow = NULL;
 228         struct mbuf *last;
 229         struct ip *ip = NULL;
 230
 231         flow =  &lro_flow_list[flow_id];
 232         if (flow->lr_mhead) {
 233                 if (lrodebug)
 234                         printf("%s: lr_mhead %x %d \n", __func__, flow->lr_seq,
 235                                 payload_len);
 236                 m_adj(lro_mb, drop_hdrlen);
 237
 238                 last = flow->lr_mtail;
 239                 while (last->m_next != NULL) {
 240                         last = last->m_next;
 241                 }
 242                 last->m_next = lro_mb;
 243
 244                 flow->lr_mtail = lro_mb;
 245
 246                 ip = mtod(flow->lr_mhead, struct ip *);
 247                 ip->ip_len += lro_mb->m_pkthdr.len;
 248                 flow->lr_mhead->m_pkthdr.len += lro_mb->m_pkthdr.len;
 249
 250                 if (flow->lr_len == 0) {
 251                         panic_plain("%s: Inconsistent LRO flow state", __func__);
 252                 }
 253                 flow->lr_len += payload_len;
 254                 flow->lr_seq += payload_len;
 255                 /*
 256                  * This bit is re-OR'd each time a packet is added to the
 257                  * large coalesced packet.
 258                  */
 259                 flow->lr_mhead->m_pkthdr.pkt_flags |= PKTF_SW_LRO_PKT;
 260                 flow->lr_mhead->m_pkthdr.lro_npkts++; /* for tcpstat.tcps_rcvpack */
 261                 if (flow->lr_mhead->m_pkthdr.lro_pktlen <
 262                                 lro_mb->m_pkthdr.lro_pktlen) {
 263                         /*
 264                          * For TCP Inter Arrival Jitter calculation, return max
 265                          * size encountered while coalescing a stream of pkts.
 266                          */
 267                         flow->lr_mhead->m_pkthdr.lro_pktlen =
 268                                                 lro_mb->m_pkthdr.lro_pktlen;
 269                 }
 270                 /* Update the timestamp value */
 271                 if (topt->to_flags & TOF_TS) {
 272                         if ((flow->lr_tsval) &&
 273                                 (TSTMP_GT(topt->to_tsval, ntohl(*(flow->lr_tsval))))) {
 274                                 *(flow->lr_tsval) = htonl(topt->to_tsval);
 275                         }
 276                         if ((flow->lr_tsecr) &&
 277                                 (topt->to_tsecr != 0) &&
 278                                 (TSTMP_GT(topt->to_tsecr, ntohl(*(flow->lr_tsecr))))) {
 279                                 if (lrodebug >= 2) {
 280                                         printf("%s: instantaneous RTT = %d \n", __func__,
 281                                                 topt->to_tsecr - ntohl(*(flow->lr_tsecr)));
 282                                 }
 283                                 *(flow->lr_tsecr) = htonl(topt->to_tsecr);
 284                         }
 285                 }
 286                 /* Coalesce the flags */
 287                 if (thflags) {
 288                         flow->lr_tcphdr->th_flags |= thflags;
 289                 }
 290                 /* Update receive window */
 291                 flow->lr_tcphdr->th_win = tcphdr->th_win;
 292         } else {
 293                 if (lro_mb) {
 294                         flow->lr_mhead = flow->lr_mtail = lro_mb;
 295                         flow->lr_mhead->m_pkthdr.pkt_flags |= PKTF_SW_LRO_PKT;
 296                         flow->lr_tcphdr = tcphdr;
 297                         if ((topt) && (topt->to_flags & TOF_TS)) {
 298                                 ASSERT(tsval != NULL);
 299                                 ASSERT(tsecr != NULL);
 300                                 flow->lr_tsval = tsval;
 301                                 flow->lr_tsecr = tsecr;
 302                         }
 303                         flow->lr_len = payload_len;
 304                         calculate_tcp_clock();
 305                         flow->lr_timestamp = tcp_now;
 306                         tcp_lro_sched_timer(0);
 307                 }
 308                 flow->lr_seq = ntohl(tcphdr->th_seq) + payload_len;
 309         }
 310         if (lro_mb) {
 311                 tcpstat.tcps_coalesced_pack++;
 312         }
 313         return;
 314 }
 315
 316 static struct mbuf *
 317 tcp_lro_eject_flow(int flow_id)
 318 {
 319         struct mbuf *mb = NULL;
 320
 321         mb = lro_flow_list[flow_id].lr_mhead;
 322         ASSERT(lro_flow_map[lro_flow_list[flow_id].lr_hash_map] == flow_id);
 323         lro_flow_map[lro_flow_list[flow_id].lr_hash_map] = TCP_LRO_FLOW_UNINIT;
 324         bzero(&lro_flow_list[flow_id], sizeof(struct lro_flow));
 325
 326         return mb;
 327 }
 328
 329 static struct mbuf*
 330 tcp_lro_eject_coalesced_pkt(int flow_id)
 331 {
 332         struct mbuf *mb = NULL;
 333         mb = lro_flow_list[flow_id].lr_mhead;
 334         lro_flow_list[flow_id].lr_mhead =
 335                 lro_flow_list[flow_id].lr_mtail = NULL;
 336         lro_flow_list[flow_id].lr_tcphdr = NULL;
 337         return mb;
 338 }
 339
 340 static struct mbuf*
 341 tcp_lro_insert_flow(struct mbuf *lro_mb, struct ip *ip_hdr,
 342                         struct tcphdr *tcp_hdr, int payload_len,
 343                         int drop_hdrlen, int hash, struct tcpopt *topt,
 344                         u_int32_t *tsval, u_int32_t *tsecr)
 345 {
 346         int i;
 347         int slot_available = 0;
 348         int candidate_flow = 0;
 349         u_int32_t oldest_timestamp;
 350         struct mbuf *mb = NULL;
 351         int collision = 0;
 352
 353         oldest_timestamp = tcp_now;
 354
 355         /* handle collision */
 356         if (lro_flow_map[hash] != TCP_LRO_FLOW_UNINIT) {
 357                 if (lrodebug) {
 358                         collision = 1;
 359                 }
 360                 candidate_flow = lro_flow_map[hash];
 361                 tcpstat.tcps_flowtbl_collision++;
 362                 goto kick_flow;
 363         }
 364
 365         for (i = 0; i < TCP_LRO_NUM_FLOWS; i++) {
 366                 if (lro_flow_list[i].lr_mhead == NULL) {
 367                         candidate_flow = i;
 368                         slot_available = 1;
 369                         break;
 370                 }
 371                 if (oldest_timestamp >= lro_flow_list[i].lr_timestamp) {
 372                         candidate_flow = i;
 373                         oldest_timestamp = lro_flow_list[i].lr_timestamp;
 374                 }
 375         }
 376
 377         if (!slot_available) {
 378                 tcpstat.tcps_flowtbl_full++;
 379 kick_flow:
 380                 /* kick the oldest flow */
 381                 mb = tcp_lro_eject_flow(candidate_flow);
 382
 383                 if (lrodebug) {
 384                         if (!slot_available) {
 385                                 printf("%s: slot unavailable.\n",__func__);
 386                         }
 387                         if (collision) {
 388                                 printf("%s: collision.\n",__func__);
 389                         }
 390                 }
 391         } else {
 392                 candidate_flow = i; /* this is now the flow to be used */
 393
 394         }
 395
 396         tcp_lro_init_flow(candidate_flow, ip_hdr, tcp_hdr, hash,
 397                                 tcp_now, payload_len);
 398         tcp_lro_coalesce(candidate_flow, lro_mb, tcp_hdr, payload_len,
 399                                 drop_hdrlen, topt, tsval, tsecr, 0);
 400         return mb;
 401 }
 402
 403 struct mbuf*
 404 tcp_lro_process_pkt(struct mbuf *lro_mb, struct ip *ip_hdr,
 405                                 struct tcphdr *tcp_hdr, int drop_hdrlen)
 406 {
 407         int flow_id = TCP_LRO_FLOW_UNINIT;
 408         int hash;
 409         unsigned int off = 0;
 410         int eject_flow = 0;
 411         int optlen;
 412         int retval = 0;
 413         struct mbuf *mb = NULL;
 414         int payload_len = 0;
 415         u_char *optp = NULL;
 416         int thflags = 0;
 417         struct tcpopt to;
 418         int ret_response = TCP_LRO_CONSUMED;
 419         int coalesced = 0, tcpflags = 0, unknown_tcpopts = 0;
 420         u_int8_t ecn;
 421
 422         if (lro_mb->m_len < (int32_t)sizeof (struct tcpiphdr)) {
 423                 if ((lro_mb = m_pullup(lro_mb, sizeof(struct tcpiphdr))) == 0) {
 424                         tcpstat.tcps_rcvshort++;
 425                         m_freem(lro_mb);
 426                         if (lrodebug) {
 427                                 printf("tcp_lro_process_pkt:mbuf too short.\n");
 428                         }
 429                         return NULL;
 430                 }
 431         }
 432
 433         /* Just in case */
 434         lro_mb->m_pkthdr.pkt_flags &= ~PKTF_SW_LRO_DID_CSUM;
 435
 436         if ((lro_mb = lro_tcp_xsum_validate(lro_mb, ip_hdr, tcp_hdr)) == NULL) {
 437                 if (lrodebug) {
 438                         printf("tcp_lro_process_pkt: TCP xsum failed.\n");
 439                 }
 440                 return NULL;
 441         }
 442
 443         /* Update stats */
 444         lro_pkt_count++;
 445
 446         /* Avoids checksumming in tcp_input */
 447         lro_mb->m_pkthdr.pkt_flags |= PKTF_SW_LRO_DID_CSUM;
 448
 449         off = tcp_hdr->th_off << 2;
 450         optlen = off - sizeof (struct tcphdr);
 451         payload_len = ip_hdr->ip_len - off;
 452         optp = (u_char *)(tcp_hdr + 1);
 453         /*
 454          * Do quick retrieval of timestamp options ("options
 455          * prediction?").  If timestamp is the only option and it's
 456          * formatted as recommended in RFC 1323 appendix A, we
 457          * quickly get the values now and not bother calling
 458          * tcp_dooptions(), etc.
 459          */
 460         if ((optlen == TCPOLEN_TSTAMP_APPA ||
 461                         (optlen > TCPOLEN_TSTAMP_APPA &&
 462                         optp[TCPOLEN_TSTAMP_APPA] == TCPOPT_EOL)) &&
 463                         *(u_int32_t *)optp == htonl(TCPOPT_TSTAMP_HDR) &&
 464                         (tcp_hdr->th_flags & TH_SYN) == 0) {
 465                         to.to_flags |= TOF_TS;
 466                         to.to_tsval = ntohl(*(u_int32_t *)(void *)(optp + 4));
 467                         to.to_tsecr = ntohl(*(u_int32_t *)(void *)(optp + 8));
 468         } else {
 469                 /*
 470                  * If TCP timestamps are not in use, or not the first option,
 471                  * skip LRO path since timestamps are used to avoid LRO
 472                  * from introducing additional latencies for retransmissions
 473                  * and other slow-paced transmissions.
 474                  */
 475                 to.to_flags = to.to_tsecr = 0;
 476                 eject_flow = 1;
 477         }
 478
 479         /* list all the conditions that can trigger a flow ejection here */
 480
 481         thflags = tcp_hdr->th_flags;
 482         if (thflags & (TH_SYN | TH_URG | TH_ECE | TH_CWR | TH_PUSH | TH_RST | TH_FIN)) {
 483                 eject_flow = tcpflags = 1;
 484         }
 485
 486         if (optlen && !((optlen == TCPOLEN_TSTAMP_APPA) &&
 487                         (to.to_flags & TOF_TS))) {
 488                 eject_flow = unknown_tcpopts = 1;
 489         }
 490
 491         if (payload_len <= LRO_MIN_COALESC_SZ) { /* zero payload ACK */
 492                 eject_flow = 1;
 493         }
 494
 495         /* Can't coalesce ECN marked packets. */
 496         ecn = ip_hdr->ip_tos & IPTOS_ECN_MASK;
 497         if (ecn == IPTOS_ECN_CE) {
 498                 /*
 499                  * ECN needs quick notification
 500                  */
 501                 if (lrodebug) {
 502                         printf("%s: ECE bits set.\n", __func__);
 503                 }
 504                 eject_flow = 1;
 505         }
 506
 507         lck_mtx_lock_spin(&tcp_lro_lock);
 508
 509         retval = tcp_lro_matching_tuple(ip_hdr, tcp_hdr, &hash, &flow_id);
 510
 511         switch (retval) {
 512         case TCP_LRO_NAN:
 513                 lck_mtx_unlock(&tcp_lro_lock);
 514                 ret_response = TCP_LRO_FLOW_NOTFOUND;
 515                 break;
 516
 517         case TCP_LRO_COALESCE:
 518                 if ((payload_len != 0) && (unknown_tcpopts == 0) &&
 519                         (tcpflags == 0) && (ecn != IPTOS_ECN_CE) && (to.to_flags & TOF_TS)) {
 520                         tcp_lro_coalesce(flow_id, lro_mb, tcp_hdr, payload_len,
 521                                 drop_hdrlen, &to,
 522                                 (to.to_flags & TOF_TS) ? (u_int32_t *)(void *)(optp + 4) : NULL,
 523                                 (to.to_flags & TOF_TS) ? (u_int32_t *)(void *)(optp + 8) : NULL,
 524                                 thflags);
 525                         if (lrodebug >= 2) {
 526                                 printf("tcp_lro_process_pkt: coalesce len = %d. flow_id = %d payload_len = %d drop_hdrlen = %d optlen = %d lport = %d seqnum = %x.\n",
 527                                         lro_flow_list[flow_id].lr_len, flow_id,
 528                                         payload_len, drop_hdrlen, optlen,
 529                                         ntohs(lro_flow_list[flow_id].lr_lport),
 530                                         ntohl(tcp_hdr->th_seq));
 531                         }
 532                         if (lro_flow_list[flow_id].lr_mhead->m_pkthdr.lro_npkts >= coalesc_sz) {
 533                                 eject_flow = 1;
 534                         }
 535                         coalesced = 1;
 536                 }
 537                 if (eject_flow) {
 538                         mb = tcp_lro_eject_coalesced_pkt(flow_id);
 539                         lro_flow_list[flow_id].lr_seq = ntohl(tcp_hdr->th_seq) +
 540                                                                 payload_len;
 541                         calculate_tcp_clock();
 542                         u_int8_t timestamp = tcp_now - lro_flow_list[flow_id].lr_timestamp;
 543                         lck_mtx_unlock(&tcp_lro_lock);
 544                         if (mb) {
 545                                 mb->m_pkthdr.lro_elapsed = timestamp;
 546                                 lro_proto_input(mb);
 547                         }
 548                         if (!coalesced) {
 549                                 if (lrodebug >= 2) {
 550                                         printf("%s: pkt payload_len = %d \n", __func__, payload_len);
 551                                 }
 552                                 lro_proto_input(lro_mb);
 553                         }
 554                 } else {
 555                         lck_mtx_unlock(&tcp_lro_lock);
 556                 }
 557                 break;
 558
 559         case TCP_LRO_EJECT_FLOW:
 560                 mb = tcp_lro_eject_coalesced_pkt(flow_id);
 561                 calculate_tcp_clock();
 562                 u_int8_t timestamp = tcp_now - lro_flow_list[flow_id].lr_timestamp;
 563                 lck_mtx_unlock(&tcp_lro_lock);
 564                 if (mb) {
 565                         if (lrodebug)
 566                                 printf("tcp_lro_process_pkt eject_flow, len = %d\n", mb->m_pkthdr.len);
 567                         mb->m_pkthdr.lro_elapsed = timestamp;
 568                         lro_proto_input(mb);
 569                 }
 570
 571                 lro_proto_input(lro_mb);
 572                 break;
 573
 574         case TCP_LRO_COLLISION:
 575                 lck_mtx_unlock(&tcp_lro_lock);
 576                 ret_response = TCP_LRO_FLOW_NOTFOUND;
 577                 break;
 578
 579         default:
 580                 lck_mtx_unlock(&tcp_lro_lock);
 581                 panic_plain("%s: unrecognized type %d", __func__, retval);
 582                 break;
 583         }
 584
 585         if (ret_response == TCP_LRO_FLOW_NOTFOUND) {
 586                 lro_proto_input(lro_mb);
 587         }
 588         return NULL;
 589 }
 590
 591 static void
 592 tcp_lro_timer_proc(void *arg1, void *arg2)
 593 {
 594 #pragma unused(arg1, arg2)
 595
 596         lck_mtx_lock_spin(&tcp_lro_lock);
 597         lro_timer_set = 0;
 598         lck_mtx_unlock(&tcp_lro_lock);
 599         tcp_lro_flush_flows();
 600 }
 601
 602 static void
 603 tcp_lro_flush_flows(void)
 604 {
 605         int i = 0;
 606         struct mbuf *mb;
 607         struct lro_flow *flow;
 608         int tcpclock_updated = 0;
 609
 610         lck_mtx_lock(&tcp_lro_lock);
 611
 612         while (i < TCP_LRO_NUM_FLOWS) {
 613                 flow = &lro_flow_list[i];
 614                 if (flow->lr_mhead != NULL) {
 615
 616                         if (!tcpclock_updated) {
 617                                 calculate_tcp_clock();
 618                                 tcpclock_updated = 1;
 619                         }
 620
 621                         if (lrodebug >= 2)
 622                                 printf("tcp_lro_flush_flows: len =%d n_pkts = %d %d %d \n",
 623                                         flow->lr_len,
 624                                         flow->lr_mhead->m_pkthdr.lro_npkts,
 625                                         flow->lr_timestamp, tcp_now);
 626
 627                         u_int8_t timestamp = tcp_now - flow->lr_timestamp;
 628
 629                         mb = tcp_lro_eject_flow(i);
 630
 631                         if (mb) {
 632                                 mb->m_pkthdr.lro_elapsed = timestamp;
 633                                 lck_mtx_unlock(&tcp_lro_lock);
 634                                 lro_update_flush_stats(mb);
 635                                 lro_proto_input(mb);
 636                                 lck_mtx_lock(&tcp_lro_lock);
 637                         }
 638                 }
 639                 i++;
 640         }
 641         lck_mtx_unlock(&tcp_lro_lock);
 642 }
 643
 644 /*
 645  * Must be called with tcp_lro_lock held.
 646  * The hint is non-zero for longer waits. The wait time dictated by coalesc_time
 647  * takes precedence, so lro_timer_set is not set for the hint case
 648  */
 649 static void
 650 tcp_lro_sched_timer(uint64_t hint)
 651 {
 652         if (lro_timer_set) {
 653                 return;
 654         }
 655
 656         lro_timer_set = 1;
 657         if (!hint) {
 658                 /* the intent is to wake up every coalesc_time msecs */
 659                 clock_interval_to_deadline(coalesc_time,
 660                         (NSEC_PER_SEC / TCP_RETRANSHZ), &lro_deadline);
 661         } else {
 662                 clock_interval_to_deadline(hint, NSEC_PER_SEC / TCP_RETRANSHZ,
 663                         &lro_deadline);
 664         }
 665         thread_call_enter_delayed(tcp_lro_timer, lro_deadline);
 666 }
 667
 668 struct mbuf*
 669 tcp_lro(struct mbuf *m, unsigned int hlen)
 670 {
 671         struct ip *ip_hdr;
 672         unsigned int tlen;
 673         struct tcphdr * tcp_hdr = NULL;
 674         unsigned int off = 0;
 675
 676         if (kipf_count != 0)
 677                 return m;
 678
 679         /*
 680          * Experiments on cellular show that the RTT is much higher
 681          * than the coalescing time of 5 msecs, causing lro to flush
 682          * 80% of the time on a single packet. Increasing
 683          * coalescing time for cellular does not show marked
 684          * improvement to throughput either. Loopback perf is hurt
 685          * by the 5 msec latency and it already sends large packets.
 686          */
 687         if (IFNET_IS_CELLULAR(m->m_pkthdr.rcvif) ||
 688                 (m->m_pkthdr.rcvif->if_type == IFT_LOOP)) {
 689                 return m;
 690         }
 691
 692         ip_hdr = mtod(m, struct ip*);
 693
 694         /* don't deal with IP options */
 695         if (hlen > sizeof (struct ip))
 696                 return (m);
 697
 698         /* only TCP is coalesced */
 699         if (ip_hdr->ip_p != IPPROTO_TCP) {
 700                 return m;
 701         }
 702
 703         if (m->m_len < (int32_t) sizeof (struct tcpiphdr)) {
 704                 if (lrodebug) printf("tcp_lro m_pullup \n");
 705                 if ((m = m_pullup(m, sizeof (struct tcpiphdr))) == 0) {
 706                         tcpstat.tcps_rcvshort++;
 707                         if (lrodebug) {
 708                                 printf("ip_lro: rcvshort.\n");
 709                         }
 710                         return NULL;
 711                 }
 712         }
 713
 714         tcp_hdr = (struct tcphdr *)((caddr_t)ip_hdr + hlen);
 715         tlen = ip_hdr->ip_len ; //ignore IP header bytes len
 716         m->m_pkthdr.lro_pktlen = tlen; /* Used to return max pkt encountered to tcp */
 717         m->m_pkthdr.lro_npkts = 1; /* Initialize a counter to hold num pkts coalesced */
 718         m->m_pkthdr.lro_elapsed = 0; /* Initialize the field to carry elapsed time */
 719         off = tcp_hdr->th_off << 2;
 720         if (off < sizeof (struct tcphdr) || off > tlen) {
 721                 tcpstat.tcps_rcvbadoff++;
 722                 if (lrodebug) {
 723                         printf("ip_lro: TCP off greater than TCP header.\n");
 724                 }
 725                 return m;
 726         }
 727
 728         return (tcp_lro_process_pkt(m, ip_hdr, tcp_hdr, hlen + off));
 729 }
 730
 731 static void
 732 lro_proto_input(struct mbuf *m)
 733 {
 734         struct ip* ip_hdr = mtod(m, struct ip*);
 735
 736         if (lrodebug >= 3) {
 737                 printf("lro_proto_input: ip_len = %d \n",
 738                         ip_hdr->ip_len);
 739         }
 740         lro_update_stats(m);
 741         ip_proto_dispatch_in_wrapper(m, ip_hdr->ip_hl << 2, ip_hdr->ip_p);
 742 }
 743
 744 static struct mbuf *
 745 lro_tcp_xsum_validate(struct mbuf *m, struct ip *ip, struct tcphdr * th)
 746 {
 747         /* Expect 32-bit aligned data pointer on strict-align platforms */
 748         MBUF_STRICT_DATA_ALIGNMENT_CHECK_32(m);
 749
 750         /* we shouldn't get here for IP with options; hence sizeof (ip) */
 751         if (tcp_input_checksum(AF_INET, m, th, sizeof (*ip), ip->ip_len)) {
 752                 if (lrodebug)
 753                         printf("%s: bad xsum and drop m = 0x%llx.\n", __func__,
 754                         (uint64_t)VM_KERNEL_ADDRPERM(m));
 755                 m_freem(m);
 756                 return (NULL);
 757         }
 758
 759         return (m);
 760 }
 761
 762 /*
 763  * When TCP detects a stable, steady flow without out of ordering,
 764  * with a sufficiently high cwnd, it invokes LRO.
 765  */
 766 int
 767 tcp_start_coalescing(struct ip *ip_hdr, struct tcphdr *tcp_hdr, int tlen)
 768 {
 769         int hash;
 770         int flow_id;
 771         struct mbuf *eject_mb;
 772         struct lro_flow *lf;
 773
 774         hash = LRO_HASH(ip_hdr->ip_src.s_addr, ip_hdr->ip_dst.s_addr,
 775                 tcp_hdr->th_sport, tcp_hdr->th_dport,
 776                 (TCP_LRO_FLOW_MAP - 1));
 777
 778
 779         lck_mtx_lock_spin(&tcp_lro_lock);
 780         flow_id = lro_flow_map[hash];
 781         if (flow_id != TCP_LRO_FLOW_NOTFOUND) {
 782                 lf = &lro_flow_list[flow_id];
 783                 if ((lf->lr_faddr.s_addr == ip_hdr->ip_src.s_addr) &&
 784                     (lf->lr_laddr.s_addr == ip_hdr->ip_dst.s_addr) &&
 785                     (lf->lr_fport == tcp_hdr->th_sport) &&
 786                     (lf->lr_lport == tcp_hdr->th_dport)) {
 787                         if ((lf->lr_tcphdr == NULL) &&
 788                                 (lf->lr_seq != (tcp_hdr->th_seq + tlen))) {
 789                                 lf->lr_seq = tcp_hdr->th_seq + tlen;
 790                         }
 791                         lf->lr_flags &= ~LRO_EJECT_REQ;
 792                 }
 793                 lck_mtx_unlock(&tcp_lro_lock);
 794                 return 0;
 795         }
 796
 797         HTONL(tcp_hdr->th_seq);
 798         HTONL(tcp_hdr->th_ack);
 799         eject_mb =
 800                 tcp_lro_insert_flow(NULL, ip_hdr, tcp_hdr, tlen, 0, hash,
 801                 NULL, NULL, NULL);
 802
 803         lck_mtx_unlock(&tcp_lro_lock);
 804
 805         NTOHL(tcp_hdr->th_seq);
 806         NTOHL(tcp_hdr->th_ack);
 807         if (lrodebug >= 3) {
 808                 printf("%s: src = %x dst = %x sport = %d dport = %d seq %x \n",
 809                         __func__, ip_hdr->ip_src.s_addr, ip_hdr->ip_dst.s_addr,
 810                         tcp_hdr->th_sport, tcp_hdr->th_dport, tcp_hdr->th_seq);
 811         }
 812         ASSERT(eject_mb == NULL);
 813         return 0;
 814 }
 815
 816 /*
 817  * When TCP detects loss or idle condition, it stops offloading
 818  * to LRO.
 819  */
 820 int
 821 tcp_lro_remove_state(struct in_addr saddr, struct in_addr daddr,
 822                 unsigned short sport, unsigned short dport)
 823 {
 824         int hash, flow_id;
 825         struct lro_flow *lf;
 826
 827         hash = LRO_HASH(daddr.s_addr, saddr.s_addr, dport, sport,
 828                 (TCP_LRO_FLOW_MAP - 1));
 829         lck_mtx_lock_spin(&tcp_lro_lock);
 830         flow_id = lro_flow_map[hash];
 831         if (flow_id == TCP_LRO_FLOW_UNINIT) {
 832                 lck_mtx_unlock(&tcp_lro_lock);
 833                 return 0;
 834         }
 835         lf = &lro_flow_list[flow_id];
 836         if ((lf->lr_faddr.s_addr == daddr.s_addr) &&
 837             (lf->lr_laddr.s_addr == saddr.s_addr) &&
 838             (lf->lr_fport == dport) &&
 839             (lf->lr_lport == sport)) {
 840                 if (lrodebug) {
 841                         printf("%s: %x %x\n", __func__,
 842                                 lf->lr_flags, lf->lr_seq);
 843                 }
 844                 lf->lr_flags |= LRO_EJECT_REQ;
 845         }
 846         lck_mtx_unlock(&tcp_lro_lock);
 847         return 0;
 848 }
 849
 850 void
 851 tcp_update_lro_seq(__uint32_t rcv_nxt, struct in_addr saddr, struct in_addr daddr,
 852                 unsigned short sport, unsigned short dport)
 853 {
 854         int hash, flow_id;
 855         struct lro_flow *lf;
 856
 857         hash = LRO_HASH(daddr.s_addr, saddr.s_addr, dport, sport,
 858                 (TCP_LRO_FLOW_MAP - 1));
 859         lck_mtx_lock_spin(&tcp_lro_lock);
 860         flow_id = lro_flow_map[hash];
 861         if (flow_id == TCP_LRO_FLOW_UNINIT) {
 862                 lck_mtx_unlock(&tcp_lro_lock);
 863                 return;
 864         }
 865         lf = &lro_flow_list[flow_id];
 866         if ((lf->lr_faddr.s_addr == daddr.s_addr) &&
 867             (lf->lr_laddr.s_addr == saddr.s_addr) &&
 868             (lf->lr_fport == dport) &&
 869             (lf->lr_lport == sport) &&
 870             (lf->lr_tcphdr == NULL)) {
 871                 lf->lr_seq = (tcp_seq)rcv_nxt;
 872         }
 873         lck_mtx_unlock(&tcp_lro_lock);
 874         return;
 875 }
 876
 877 static void
 878 lro_update_stats(struct mbuf *m)
 879 {
 880         switch(m->m_pkthdr.lro_npkts) {
 881         case 0: /* fall through */
 882         case 1:
 883                 break;
 884
 885         case 2:
 886                 tcpstat.tcps_lro_twopack++;
 887                 break;
 888
 889         case 3: /* fall through */
 890         case 4:
 891                 tcpstat.tcps_lro_multpack++;
 892                 break;
 893
 894         default:
 895                 tcpstat.tcps_lro_largepack++;
 896                 break;
 897         }
 898         return;
 899 }
 900
 901 static void
 902 lro_update_flush_stats(struct mbuf *m)
 903 {
 904         lro_flushes++;
 905         switch(m->m_pkthdr.lro_npkts) {
 906         case 0: ASSERT(0);
 907         case 1: lro_single_flushes++;
 908                 break;
 909         case 2: lro_double_flushes++;
 910                 break;
 911         default: lro_good_flushes++;
 912                 break;
 913         }
 914         return;
 915 }