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*, int);
  99
 100 void
 101 tcp_lro_init(void)
 102 {
 103         int i;
 104
 105         bzero(lro_flow_list, sizeof (struct lro_flow) * TCP_LRO_NUM_FLOWS);
 106         for (i = 0; i < TCP_LRO_FLOW_MAP; i++) {
 107                 lro_flow_map[i] = TCP_LRO_FLOW_UNINIT;
 108         }
 109
 110         /*
 111          * allocate lock group attribute, group and attribute for tcp_lro_lock
 112          */
 113         tcp_lro_mtx_grp_attr = lck_grp_attr_alloc_init();
 114         tcp_lro_mtx_grp = lck_grp_alloc_init("tcplro", tcp_lro_mtx_grp_attr);
 115         tcp_lro_mtx_attr = lck_attr_alloc_init();
 116         lck_mtx_init(&tcp_lro_lock, tcp_lro_mtx_grp, tcp_lro_mtx_attr);
 117
 118         tcp_lro_timer = thread_call_allocate(tcp_lro_timer_proc, NULL);
 119         if (tcp_lro_timer == NULL) {
 120                 panic_plain("%s: unable to allocate lro timer", __func__);
 121         }
 122
 123         return;
 124 }
 125
 126 static int
 127 tcp_lro_matching_tuple(struct ip* ip_hdr, struct tcphdr *tcp_hdr, int *hash,
 128                         int *flow_id )
 129 {
 130         struct lro_flow *flow;
 131         tcp_seq seqnum;
 132         unsigned int off = 0;
 133         int payload_len = 0;
 134
 135         *hash = LRO_HASH(ip_hdr->ip_src.s_addr, ip_hdr->ip_dst.s_addr,
 136                 tcp_hdr->th_sport, tcp_hdr->th_dport, (TCP_LRO_FLOW_MAP - 1));
 137
 138         *flow_id = lro_flow_map[*hash];
 139         if (*flow_id == TCP_LRO_FLOW_NOTFOUND) {
 140                 return TCP_LRO_NAN;
 141         }
 142
 143         seqnum = tcp_hdr->th_seq;
 144         off = tcp_hdr->th_off << 2;
 145         payload_len = ip_hdr->ip_len - off;
 146
 147         flow = &lro_flow_list[*flow_id];
 148
 149         if ((flow->lr_faddr.s_addr == ip_hdr->ip_src.s_addr) &&
 150                         (flow->lr_laddr.s_addr == ip_hdr->ip_dst.s_addr) &&
 151                         (flow->lr_fport == tcp_hdr->th_sport) &&
 152                         (flow->lr_lport == tcp_hdr->th_dport)) {
 153                 if (flow->lr_tcphdr == NULL) {
 154                         if (ntohl(seqnum) == flow->lr_seq) {
 155                                 return TCP_LRO_COALESCE;
 156                         }
 157                         if (lrodebug >= 4) {
 158                                 printf("%s: seqnum = %x, lr_seq = %x\n",
 159                                         __func__, ntohl(seqnum), flow->lr_seq);
 160                         }
 161                         lro_seq_mismatch++;
 162                         if (SEQ_GT(ntohl(seqnum), flow->lr_seq)) {
 163                                 lro_seq_outoforder++;
 164                                 /*
 165                                  * Whenever we receive out of order packets it
 166                                  * signals loss and recovery and LRO doesn't
 167                                  * let flows recover quickly. So eject.
 168                                  */
 169                                  flow->lr_flags |= LRO_EJECT_REQ;
 170
 171                         }
 172                         return TCP_LRO_NAN;
 173                 }
 174
 175                 if (flow->lr_flags & LRO_EJECT_REQ) {
 176                         if (lrodebug)
 177                                 printf("%s: eject. \n", __func__);
 178                         return TCP_LRO_EJECT_FLOW;
 179                 }
 180                 if (SEQ_GT(tcp_hdr->th_ack, flow->lr_tcphdr->th_ack)) {
 181                         if (lrodebug) {
 182                                 printf("%s: th_ack = %x flow_ack = %x \n",
 183                                         __func__, tcp_hdr->th_ack,
 184                                         flow->lr_tcphdr->th_ack);
 185                         }
 186                         return TCP_LRO_EJECT_FLOW;
 187                 }
 188
 189                 if (ntohl(seqnum) == (ntohl(lro_flow_list[*flow_id].lr_tcphdr->th_seq) + lro_flow_list[*flow_id].lr_len)) {
 190                         return TCP_LRO_COALESCE;
 191                 } else {
 192                         /* LRO does not handle loss recovery well, eject */
 193                         flow->lr_flags |= LRO_EJECT_REQ;
 194                         return TCP_LRO_EJECT_FLOW;
 195                 }
 196         }
 197         if (lrodebug) printf("tcp_lro_matching_tuple: collision \n");
 198         return TCP_LRO_COLLISION;
 199 }
 200
 201 static void
 202 tcp_lro_init_flow(int flow_id, struct ip* ip_hdr, struct tcphdr *tcp_hdr,
 203                         int hash, u_int32_t timestamp, int payload_len)
 204 {
 205         struct lro_flow *flow = NULL;
 206
 207         flow = &lro_flow_list[flow_id];
 208
 209         flow->lr_hash_map = hash;
 210         flow->lr_faddr.s_addr = ip_hdr->ip_src.s_addr;
 211         flow->lr_laddr.s_addr = ip_hdr->ip_dst.s_addr;
 212         flow->lr_fport = tcp_hdr->th_sport;
 213         flow->lr_lport = tcp_hdr->th_dport;
 214         lro_flow_map[hash] = flow_id;
 215         flow->lr_timestamp = timestamp;
 216         flow->lr_seq = ntohl(tcp_hdr->th_seq) + payload_len;
 217         flow->lr_flags = 0;
 218         return;
 219 }
 220
 221 static void
 222 tcp_lro_coalesce(int flow_id, struct mbuf *lro_mb, struct tcphdr *tcphdr,
 223                         int payload_len, int drop_hdrlen, struct tcpopt *topt,
 224                         u_int32_t* tsval, u_int32_t* tsecr, int thflags)
 225 {
 226         struct lro_flow *flow = NULL;
 227         struct mbuf *last;
 228         struct ip *ip = NULL;
 229
 230         flow =  &lro_flow_list[flow_id];
 231         if (flow->lr_mhead) {
 232                 if (lrodebug)
 233                         printf("%s: lr_mhead %x %d \n", __func__, flow->lr_seq,
 234                                 payload_len);
 235                 m_adj(lro_mb, drop_hdrlen);
 236
 237                 last = flow->lr_mtail;
 238                 while (last->m_next != NULL) {
 239                         last = last->m_next;
 240                 }
 241                 last->m_next = lro_mb;
 242
 243                 flow->lr_mtail = lro_mb;
 244
 245                 ip = mtod(flow->lr_mhead, struct ip *);
 246                 ip->ip_len += lro_mb->m_pkthdr.len;
 247                 flow->lr_mhead->m_pkthdr.len += lro_mb->m_pkthdr.len;
 248
 249                 if (flow->lr_len == 0) {
 250                         panic_plain("%s: Inconsistent LRO flow state", __func__);
 251                 }
 252                 flow->lr_len += payload_len;
 253                 flow->lr_seq += payload_len;
 254                 /*
 255                  * This bit is re-OR'd each time a packet is added to the
 256                  * large coalesced packet.
 257                  */
 258                 flow->lr_mhead->m_pkthdr.pkt_flags |= PKTF_SW_LRO_PKT;
 259                 flow->lr_mhead->m_pkthdr.lro_npkts++; /* for tcpstat.tcps_rcvpack */
 260                 if (flow->lr_mhead->m_pkthdr.lro_pktlen <
 261                                 lro_mb->m_pkthdr.lro_pktlen) {
 262                         /*
 263                          * For TCP Inter Arrival Jitter calculation, return max
 264                          * size encountered while coalescing a stream of pkts.
 265                          */
 266                         flow->lr_mhead->m_pkthdr.lro_pktlen =
 267                                                 lro_mb->m_pkthdr.lro_pktlen;
 268                 }
 269                 /* Update the timestamp value */
 270                 if (topt->to_flags & TOF_TS) {
 271                         if ((flow->lr_tsval) &&
 272                                 (TSTMP_GT(topt->to_tsval, ntohl(*(flow->lr_tsval))))) {
 273                                 *(flow->lr_tsval) = htonl(topt->to_tsval);
 274                         }
 275                         if ((flow->lr_tsecr) &&
 276                                 (topt->to_tsecr != 0) &&
 277                                 (TSTMP_GT(topt->to_tsecr, ntohl(*(flow->lr_tsecr))))) {
 278                                 if (lrodebug >= 2) {
 279                                         printf("%s: instantaneous RTT = %d \n", __func__,
 280                                                 topt->to_tsecr - ntohl(*(flow->lr_tsecr)));
 281                                 }
 282                                 *(flow->lr_tsecr) = htonl(topt->to_tsecr);
 283                         }
 284                 }
 285                 /* Coalesce the flags */
 286                 if (thflags) {
 287                         flow->lr_tcphdr->th_flags |= thflags;
 288                 }
 289                 /* Update receive window */
 290                 flow->lr_tcphdr->th_win = tcphdr->th_win;
 291         } else {
 292                 if (lro_mb) {
 293                         flow->lr_mhead = flow->lr_mtail = lro_mb;
 294                         flow->lr_mhead->m_pkthdr.pkt_flags |= PKTF_SW_LRO_PKT;
 295                         flow->lr_tcphdr = tcphdr;
 296                         if ((topt) && (topt->to_flags & TOF_TS)) {
 297                                 ASSERT(tsval != NULL);
 298                                 ASSERT(tsecr != NULL);
 299                                 flow->lr_tsval = tsval;
 300                                 flow->lr_tsecr = tsecr;
 301                         }
 302                         flow->lr_len = payload_len;
 303                         calculate_tcp_clock();
 304                         flow->lr_timestamp = tcp_now;
 305                         tcp_lro_sched_timer(0);
 306                 }
 307                 flow->lr_seq = ntohl(tcphdr->th_seq) + payload_len;
 308         }
 309         if (lro_mb) {
 310                 tcpstat.tcps_coalesced_pack++;
 311         }
 312         return;
 313 }
 314
 315 static struct mbuf *
 316 tcp_lro_eject_flow(int flow_id)
 317 {
 318         struct mbuf *mb = NULL;
 319
 320         mb = lro_flow_list[flow_id].lr_mhead;
 321         ASSERT(lro_flow_map[lro_flow_list[flow_id].lr_hash_map] == flow_id);
 322         lro_flow_map[lro_flow_list[flow_id].lr_hash_map] = TCP_LRO_FLOW_UNINIT;
 323         bzero(&lro_flow_list[flow_id], sizeof(struct lro_flow));
 324
 325         return mb;
 326 }
 327
 328 static struct mbuf*
 329 tcp_lro_eject_coalesced_pkt(int flow_id)
 330 {
 331         struct mbuf *mb = NULL;
 332         mb = lro_flow_list[flow_id].lr_mhead;
 333         lro_flow_list[flow_id].lr_mhead =
 334                 lro_flow_list[flow_id].lr_mtail = NULL;
 335         lro_flow_list[flow_id].lr_tcphdr = NULL;
 336         return mb;
 337 }
 338
 339 static struct mbuf*
 340 tcp_lro_insert_flow(struct mbuf *lro_mb, struct ip *ip_hdr,
 341                         struct tcphdr *tcp_hdr, int payload_len,
 342                         int drop_hdrlen, int hash, struct tcpopt *topt,
 343                         u_int32_t *tsval, u_int32_t *tsecr)
 344 {
 345         int i;
 346         int slot_available = 0;
 347         int candidate_flow = 0;
 348         u_int32_t oldest_timestamp;
 349         struct mbuf *mb = NULL;
 350         int collision = 0;
 351
 352         oldest_timestamp = tcp_now;
 353
 354         /* handle collision */
 355         if (lro_flow_map[hash] != TCP_LRO_FLOW_UNINIT) {
 356                 if (lrodebug) {
 357                         collision = 1;
 358                 }
 359                 candidate_flow = lro_flow_map[hash];
 360                 tcpstat.tcps_flowtbl_collision++;
 361                 goto kick_flow;
 362         }
 363
 364         for (i = 0; i < TCP_LRO_NUM_FLOWS; i++) {
 365                 if (lro_flow_list[i].lr_mhead == NULL) {
 366                         candidate_flow = i;
 367                         slot_available = 1;
 368                         break;
 369                 }
 370                 if (oldest_timestamp >= lro_flow_list[i].lr_timestamp) {
 371                         candidate_flow = i;
 372                         oldest_timestamp = lro_flow_list[i].lr_timestamp;
 373                 }
 374         }
 375
 376         if (!slot_available) {
 377                 tcpstat.tcps_flowtbl_full++;
 378 kick_flow:
 379                 /* kick the oldest flow */
 380                 mb = tcp_lro_eject_flow(candidate_flow);
 381
 382                 if (lrodebug) {
 383                         if (!slot_available) {
 384                                 printf("%s: slot unavailable.\n",__func__);
 385                         }
 386                         if (collision) {
 387                                 printf("%s: collision.\n",__func__);
 388                         }
 389                 }
 390         } else {
 391                 candidate_flow = i; /* this is now the flow to be used */
 392
 393         }
 394
 395         tcp_lro_init_flow(candidate_flow, ip_hdr, tcp_hdr, hash,
 396                                 tcp_now, payload_len);
 397         tcp_lro_coalesce(candidate_flow, lro_mb, tcp_hdr, payload_len,
 398                                 drop_hdrlen, topt, tsval, tsecr, 0);
 399         return mb;
 400 }
 401
 402 struct mbuf*
 403 tcp_lro_process_pkt(struct mbuf *lro_mb, int drop_hdrlen)
 404 {
 405         int flow_id = TCP_LRO_FLOW_UNINIT;
 406         int hash;
 407         unsigned int off = 0;
 408         int eject_flow = 0;
 409         int optlen;
 410         int retval = 0;
 411         struct mbuf *mb = NULL;
 412         int payload_len = 0;
 413         u_char *optp = NULL;
 414         int thflags = 0;
 415         struct tcpopt to;
 416         int ret_response = TCP_LRO_CONSUMED;
 417         int coalesced = 0, tcpflags = 0, unknown_tcpopts = 0;
 418         u_int8_t ecn;
 419         struct ip *ip_hdr;
 420         struct tcphdr *tcp_hdr;
 421
 422         if (lro_mb->m_len < drop_hdrlen) {
 423                 if ((lro_mb = m_pullup(lro_mb, drop_hdrlen)) == NULL) {
 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         ip_hdr = mtod(lro_mb, struct ip*);
 434         tcp_hdr = (struct tcphdr *)((caddr_t)ip_hdr + sizeof(struct ip));
 435
 436         /* Just in case */
 437         lro_mb->m_pkthdr.pkt_flags &= ~PKTF_SW_LRO_DID_CSUM;
 438
 439         if ((lro_mb = lro_tcp_xsum_validate(lro_mb, ip_hdr, tcp_hdr)) == NULL) {
 440                 if (lrodebug) {
 441                         printf("tcp_lro_process_pkt: TCP xsum failed.\n");
 442                 }
 443                 return (NULL);
 444         }
 445
 446         /* Update stats */
 447         lro_pkt_count++;
 448
 449         /* Avoids checksumming in tcp_input */
 450         lro_mb->m_pkthdr.pkt_flags |= PKTF_SW_LRO_DID_CSUM;
 451
 452         off = tcp_hdr->th_off << 2;
 453         optlen = off - sizeof (struct tcphdr);
 454         payload_len = ip_hdr->ip_len - off;
 455         optp = (u_char *)(tcp_hdr + 1);
 456         /*
 457          * Do quick retrieval of timestamp options ("options
 458          * prediction?").  If timestamp is the only option and it's
 459          * formatted as recommended in RFC 1323 appendix A, we
 460          * quickly get the values now and not bother calling
 461          * tcp_dooptions(), etc.
 462          */
 463         bzero(&to, sizeof(to));
 464         if ((optlen == TCPOLEN_TSTAMP_APPA ||
 465                         (optlen > TCPOLEN_TSTAMP_APPA &&
 466                         optp[TCPOLEN_TSTAMP_APPA] == TCPOPT_EOL)) &&
 467                         *(u_int32_t *)optp == htonl(TCPOPT_TSTAMP_HDR) &&
 468                         (tcp_hdr->th_flags & TH_SYN) == 0) {
 469                         to.to_flags |= TOF_TS;
 470                         to.to_tsval = ntohl(*(u_int32_t *)(void *)(optp + 4));
 471                         to.to_tsecr = ntohl(*(u_int32_t *)(void *)(optp + 8));
 472         } else {
 473                 /*
 474                  * If TCP timestamps are not in use, or not the first option,
 475                  * skip LRO path since timestamps are used to avoid LRO
 476                  * from introducing additional latencies for retransmissions
 477                  * and other slow-paced transmissions.
 478                  */
 479                 to.to_flags = to.to_tsecr = 0;
 480                 eject_flow = 1;
 481         }
 482
 483         /* list all the conditions that can trigger a flow ejection here */
 484
 485         thflags = tcp_hdr->th_flags;
 486         if (thflags & (TH_SYN | TH_URG | TH_ECE | TH_CWR | TH_PUSH | TH_RST | TH_FIN)) {
 487                 eject_flow = tcpflags = 1;
 488         }
 489
 490         if (optlen && !((optlen == TCPOLEN_TSTAMP_APPA) &&
 491                         (to.to_flags & TOF_TS))) {
 492                 eject_flow = unknown_tcpopts = 1;
 493         }
 494
 495         if (payload_len <= LRO_MIN_COALESC_SZ) { /* zero payload ACK */
 496                 eject_flow = 1;
 497         }
 498
 499         /* Can't coalesce ECN marked packets. */
 500         ecn = ip_hdr->ip_tos & IPTOS_ECN_MASK;
 501         if (ecn == IPTOS_ECN_CE) {
 502                 /*
 503                  * ECN needs quick notification
 504                  */
 505                 if (lrodebug) {
 506                         printf("%s: ECE bits set.\n", __func__);
 507                 }
 508                 eject_flow = 1;
 509         }
 510
 511         lck_mtx_lock_spin(&tcp_lro_lock);
 512
 513         retval = tcp_lro_matching_tuple(ip_hdr, tcp_hdr, &hash, &flow_id);
 514
 515         switch (retval) {
 516         case TCP_LRO_NAN:
 517                 lck_mtx_unlock(&tcp_lro_lock);
 518                 ret_response = TCP_LRO_FLOW_NOTFOUND;
 519                 break;
 520
 521         case TCP_LRO_COALESCE:
 522                 if ((payload_len != 0) && (unknown_tcpopts == 0) &&
 523                         (tcpflags == 0) && (ecn != IPTOS_ECN_CE) && (to.to_flags & TOF_TS)) {
 524                         tcp_lro_coalesce(flow_id, lro_mb, tcp_hdr, payload_len,
 525                                 drop_hdrlen, &to,
 526                                 (to.to_flags & TOF_TS) ? (u_int32_t *)(void *)(optp + 4) : NULL,
 527                                 (to.to_flags & TOF_TS) ? (u_int32_t *)(void *)(optp + 8) : NULL,
 528                                 thflags);
 529                         if (lrodebug >= 2) {
 530                                 printf("tcp_lro_process_pkt: coalesce len = %d. flow_id = %d payload_len = %d drop_hdrlen = %d optlen = %d lport = %d seqnum = %x.\n",
 531                                         lro_flow_list[flow_id].lr_len, flow_id,
 532                                         payload_len, drop_hdrlen, optlen,
 533                                         ntohs(lro_flow_list[flow_id].lr_lport),
 534                                         ntohl(tcp_hdr->th_seq));
 535                         }
 536                         if (lro_flow_list[flow_id].lr_mhead->m_pkthdr.lro_npkts >= coalesc_sz) {
 537                                 eject_flow = 1;
 538                         }
 539                         coalesced = 1;
 540                 }
 541                 if (eject_flow) {
 542                         mb = tcp_lro_eject_coalesced_pkt(flow_id);
 543                         lro_flow_list[flow_id].lr_seq = ntohl(tcp_hdr->th_seq) +
 544                                                                 payload_len;
 545                         calculate_tcp_clock();
 546                         u_int8_t timestamp = tcp_now - lro_flow_list[flow_id].lr_timestamp;
 547                         lck_mtx_unlock(&tcp_lro_lock);
 548                         if (mb) {
 549                                 mb->m_pkthdr.lro_elapsed = timestamp;
 550                                 lro_proto_input(mb);
 551                         }
 552                         if (!coalesced) {
 553                                 if (lrodebug >= 2) {
 554                                         printf("%s: pkt payload_len = %d \n", __func__, payload_len);
 555                                 }
 556                                 lro_proto_input(lro_mb);
 557                         }
 558                 } else {
 559                         lck_mtx_unlock(&tcp_lro_lock);
 560                 }
 561                 break;
 562
 563         case TCP_LRO_EJECT_FLOW:
 564                 mb = tcp_lro_eject_coalesced_pkt(flow_id);
 565                 calculate_tcp_clock();
 566                 u_int8_t timestamp = tcp_now - lro_flow_list[flow_id].lr_timestamp;
 567                 lck_mtx_unlock(&tcp_lro_lock);
 568                 if (mb) {
 569                         if (lrodebug)
 570                                 printf("tcp_lro_process_pkt eject_flow, len = %d\n", mb->m_pkthdr.len);
 571                         mb->m_pkthdr.lro_elapsed = timestamp;
 572                         lro_proto_input(mb);
 573                 }
 574
 575                 lro_proto_input(lro_mb);
 576                 break;
 577
 578         case TCP_LRO_COLLISION:
 579                 lck_mtx_unlock(&tcp_lro_lock);
 580                 ret_response = TCP_LRO_FLOW_NOTFOUND;
 581                 break;
 582
 583         default:
 584                 lck_mtx_unlock(&tcp_lro_lock);
 585                 panic_plain("%s: unrecognized type %d", __func__, retval);
 586                 break;
 587         }
 588
 589         if (ret_response == TCP_LRO_FLOW_NOTFOUND) {
 590                 lro_proto_input(lro_mb);
 591         }
 592         return (NULL);
 593 }
 594
 595 static void
 596 tcp_lro_timer_proc(void *arg1, void *arg2)
 597 {
 598 #pragma unused(arg1, arg2)
 599
 600         lck_mtx_lock_spin(&tcp_lro_lock);
 601         lro_timer_set = 0;
 602         lck_mtx_unlock(&tcp_lro_lock);
 603         tcp_lro_flush_flows();
 604 }
 605
 606 static void
 607 tcp_lro_flush_flows(void)
 608 {
 609         int i = 0;
 610         struct mbuf *mb;
 611         struct lro_flow *flow;
 612         int tcpclock_updated = 0;
 613
 614         lck_mtx_lock(&tcp_lro_lock);
 615
 616         while (i < TCP_LRO_NUM_FLOWS) {
 617                 flow = &lro_flow_list[i];
 618                 if (flow->lr_mhead != NULL) {
 619
 620                         if (!tcpclock_updated) {
 621                                 calculate_tcp_clock();
 622                                 tcpclock_updated = 1;
 623                         }
 624
 625                         if (lrodebug >= 2)
 626                                 printf("tcp_lro_flush_flows: len =%d n_pkts = %d %d %d \n",
 627                                         flow->lr_len,
 628                                         flow->lr_mhead->m_pkthdr.lro_npkts,
 629                                         flow->lr_timestamp, tcp_now);
 630
 631                         u_int8_t timestamp = tcp_now - flow->lr_timestamp;
 632
 633                         mb = tcp_lro_eject_flow(i);
 634
 635                         if (mb) {
 636                                 mb->m_pkthdr.lro_elapsed = timestamp;
 637                                 lck_mtx_unlock(&tcp_lro_lock);
 638                                 lro_update_flush_stats(mb);
 639                                 lro_proto_input(mb);
 640                                 lck_mtx_lock(&tcp_lro_lock);
 641                         }
 642                 }
 643                 i++;
 644         }
 645         lck_mtx_unlock(&tcp_lro_lock);
 646 }
 647
 648 /*
 649  * Must be called with tcp_lro_lock held.
 650  * The hint is non-zero for longer waits. The wait time dictated by coalesc_time
 651  * takes precedence, so lro_timer_set is not set for the hint case
 652  */
 653 static void
 654 tcp_lro_sched_timer(uint64_t hint)
 655 {
 656         if (lro_timer_set) {
 657                 return;
 658         }
 659
 660         lro_timer_set = 1;
 661         if (!hint) {
 662                 /* the intent is to wake up every coalesc_time msecs */
 663                 clock_interval_to_deadline(coalesc_time,
 664                         (NSEC_PER_SEC / TCP_RETRANSHZ), &lro_deadline);
 665         } else {
 666                 clock_interval_to_deadline(hint, NSEC_PER_SEC / TCP_RETRANSHZ,
 667                         &lro_deadline);
 668         }
 669         thread_call_enter_delayed(tcp_lro_timer, lro_deadline);
 670 }
 671
 672 struct mbuf*
 673 tcp_lro(struct mbuf *m, unsigned int hlen)
 674 {
 675         struct ip *ip_hdr;
 676         unsigned int tlen;
 677         struct tcphdr * tcp_hdr = NULL;
 678         unsigned int off = 0;
 679
 680         if (kipf_count != 0)
 681                 return (m);
 682
 683         /*
 684          * Experiments on cellular show that the RTT is much higher
 685          * than the coalescing time of 5 msecs, causing lro to flush
 686          * 80% of the time on a single packet. Increasing
 687          * coalescing time for cellular does not show marked
 688          * improvement to throughput either. Loopback perf is hurt
 689          * by the 5 msec latency and it already sends large packets.
 690          */
 691         if (IFNET_IS_CELLULAR(m->m_pkthdr.rcvif) ||
 692                 (m->m_pkthdr.rcvif->if_type == IFT_LOOP)) {
 693                 return (m);
 694         }
 695
 696         ip_hdr = mtod(m, struct ip*);
 697
 698         /* don't deal with IP options */
 699         if (hlen != sizeof (struct ip))
 700                 return (m);
 701
 702         /* only TCP is coalesced */
 703         if (ip_hdr->ip_p != IPPROTO_TCP) {
 704                 return (m);
 705         }
 706
 707         if (m->m_len < (int32_t) sizeof (struct tcpiphdr)) {
 708                 if (lrodebug) printf("tcp_lro m_pullup \n");
 709                 if ((m = m_pullup(m, sizeof (struct tcpiphdr))) == NULL) {
 710                         tcpstat.tcps_rcvshort++;
 711                         if (lrodebug) {
 712                                 printf("ip_lro: rcvshort.\n");
 713                         }
 714                         return (NULL);
 715                 }
 716                 ip_hdr = mtod(m, struct ip*);
 717         }
 718
 719         tcp_hdr = (struct tcphdr *)((caddr_t)ip_hdr + hlen);
 720         tlen = ip_hdr->ip_len ; //ignore IP header bytes len
 721         m->m_pkthdr.lro_pktlen = tlen; /* Used to return max pkt encountered to tcp */
 722         m->m_pkthdr.lro_npkts = 1; /* Initialize a counter to hold num pkts coalesced */
 723         m->m_pkthdr.lro_elapsed = 0; /* Initialize the field to carry elapsed time */
 724         off = tcp_hdr->th_off << 2;
 725         if (off < sizeof (struct tcphdr) || off > tlen) {
 726                 tcpstat.tcps_rcvbadoff++;
 727                 if (lrodebug) {
 728                         printf("ip_lro: TCP off greater than TCP header.\n");
 729                 }
 730                 return (m);
 731         }
 732
 733         return (tcp_lro_process_pkt(m, hlen + off));
 734 }
 735
 736 static void
 737 lro_proto_input(struct mbuf *m)
 738 {
 739         struct ip* ip_hdr = mtod(m, struct ip*);
 740
 741         if (lrodebug >= 3) {
 742                 printf("lro_proto_input: ip_len = %d \n",
 743                         ip_hdr->ip_len);
 744         }
 745         lro_update_stats(m);
 746         ip_proto_dispatch_in_wrapper(m, ip_hdr->ip_hl << 2, ip_hdr->ip_p);
 747 }
 748
 749 static struct mbuf *
 750 lro_tcp_xsum_validate(struct mbuf *m, struct ip *ip, struct tcphdr * th)
 751 {
 752         /* Expect 32-bit aligned data pointer on strict-align platforms */
 753         MBUF_STRICT_DATA_ALIGNMENT_CHECK_32(m);
 754
 755         /* we shouldn't get here for IP with options; hence sizeof (ip) */
 756         if (tcp_input_checksum(AF_INET, m, th, sizeof (*ip), ip->ip_len)) {
 757                 if (lrodebug)
 758                         printf("%s: bad xsum and drop m = 0x%llx.\n", __func__,
 759                         (uint64_t)VM_KERNEL_ADDRPERM(m));
 760                 m_freem(m);
 761                 return (NULL);
 762         }
 763
 764         return (m);
 765 }
 766
 767 /*
 768  * When TCP detects a stable, steady flow without out of ordering,
 769  * with a sufficiently high cwnd, it invokes LRO.
 770  */
 771 int
 772 tcp_start_coalescing(struct ip *ip_hdr, struct tcphdr *tcp_hdr, int tlen)
 773 {
 774         int hash;
 775         int flow_id;
 776         struct mbuf *eject_mb;
 777         struct lro_flow *lf;
 778
 779         hash = LRO_HASH(ip_hdr->ip_src.s_addr, ip_hdr->ip_dst.s_addr,
 780                 tcp_hdr->th_sport, tcp_hdr->th_dport,
 781                 (TCP_LRO_FLOW_MAP - 1));
 782
 783
 784         lck_mtx_lock_spin(&tcp_lro_lock);
 785         flow_id = lro_flow_map[hash];
 786         if (flow_id != TCP_LRO_FLOW_NOTFOUND) {
 787                 lf = &lro_flow_list[flow_id];
 788                 if ((lf->lr_faddr.s_addr == ip_hdr->ip_src.s_addr) &&
 789                     (lf->lr_laddr.s_addr == ip_hdr->ip_dst.s_addr) &&
 790                     (lf->lr_fport == tcp_hdr->th_sport) &&
 791                     (lf->lr_lport == tcp_hdr->th_dport)) {
 792                         if ((lf->lr_tcphdr == NULL) &&
 793                                 (lf->lr_seq != (tcp_hdr->th_seq + tlen))) {
 794                                 lf->lr_seq = tcp_hdr->th_seq + tlen;
 795                         }
 796                         lf->lr_flags &= ~LRO_EJECT_REQ;
 797                 }
 798                 lck_mtx_unlock(&tcp_lro_lock);
 799                 return 0;
 800         }
 801
 802         HTONL(tcp_hdr->th_seq);
 803         HTONL(tcp_hdr->th_ack);
 804         eject_mb =
 805                 tcp_lro_insert_flow(NULL, ip_hdr, tcp_hdr, tlen, 0, hash,
 806                 NULL, NULL, NULL);
 807
 808         lck_mtx_unlock(&tcp_lro_lock);
 809
 810         NTOHL(tcp_hdr->th_seq);
 811         NTOHL(tcp_hdr->th_ack);
 812         if (lrodebug >= 3) {
 813                 printf("%s: src = %x dst = %x sport = %d dport = %d seq %x \n",
 814                         __func__, ip_hdr->ip_src.s_addr, ip_hdr->ip_dst.s_addr,
 815                         tcp_hdr->th_sport, tcp_hdr->th_dport, tcp_hdr->th_seq);
 816         }
 817         ASSERT(eject_mb == NULL);
 818         return 0;
 819 }
 820
 821 /*
 822  * When TCP detects loss or idle condition, it stops offloading
 823  * to LRO.
 824  */
 825 int
 826 tcp_lro_remove_state(struct in_addr saddr, struct in_addr daddr,
 827                 unsigned short sport, unsigned short dport)
 828 {
 829         int hash, flow_id;
 830         struct lro_flow *lf;
 831
 832         hash = LRO_HASH(daddr.s_addr, saddr.s_addr, dport, sport,
 833                 (TCP_LRO_FLOW_MAP - 1));
 834         lck_mtx_lock_spin(&tcp_lro_lock);
 835         flow_id = lro_flow_map[hash];
 836         if (flow_id == TCP_LRO_FLOW_UNINIT) {
 837                 lck_mtx_unlock(&tcp_lro_lock);
 838                 return 0;
 839         }
 840         lf = &lro_flow_list[flow_id];
 841         if ((lf->lr_faddr.s_addr == daddr.s_addr) &&
 842             (lf->lr_laddr.s_addr == saddr.s_addr) &&
 843             (lf->lr_fport == dport) &&
 844             (lf->lr_lport == sport)) {
 845                 if (lrodebug) {
 846                         printf("%s: %x %x\n", __func__,
 847                                 lf->lr_flags, lf->lr_seq);
 848                 }
 849                 lf->lr_flags |= LRO_EJECT_REQ;
 850         }
 851         lck_mtx_unlock(&tcp_lro_lock);
 852         return 0;
 853 }
 854
 855 void
 856 tcp_update_lro_seq(__uint32_t rcv_nxt, struct in_addr saddr, struct in_addr daddr,
 857                 unsigned short sport, unsigned short dport)
 858 {
 859         int hash, flow_id;
 860         struct lro_flow *lf;
 861
 862         hash = LRO_HASH(daddr.s_addr, saddr.s_addr, dport, sport,
 863                 (TCP_LRO_FLOW_MAP - 1));
 864         lck_mtx_lock_spin(&tcp_lro_lock);
 865         flow_id = lro_flow_map[hash];
 866         if (flow_id == TCP_LRO_FLOW_UNINIT) {
 867                 lck_mtx_unlock(&tcp_lro_lock);
 868                 return;
 869         }
 870         lf = &lro_flow_list[flow_id];
 871         if ((lf->lr_faddr.s_addr == daddr.s_addr) &&
 872             (lf->lr_laddr.s_addr == saddr.s_addr) &&
 873             (lf->lr_fport == dport) &&
 874             (lf->lr_lport == sport) &&
 875             (lf->lr_tcphdr == NULL)) {
 876                 lf->lr_seq = (tcp_seq)rcv_nxt;
 877         }
 878         lck_mtx_unlock(&tcp_lro_lock);
 879         return;
 880 }
 881
 882 static void
 883 lro_update_stats(struct mbuf *m)
 884 {
 885         switch(m->m_pkthdr.lro_npkts) {
 886         case 0: /* fall through */
 887         case 1:
 888                 break;
 889
 890         case 2:
 891                 tcpstat.tcps_lro_twopack++;
 892                 break;
 893
 894         case 3: /* fall through */
 895         case 4:
 896                 tcpstat.tcps_lro_multpack++;
 897                 break;
 898
 899         default:
 900                 tcpstat.tcps_lro_largepack++;
 901                 break;
 902         }
 903         return;
 904 }
 905
 906 static void
 907 lro_update_flush_stats(struct mbuf *m)
 908 {
 909         lro_flushes++;
 910         switch(m->m_pkthdr.lro_npkts) {
 911         case 0: ASSERT(0);
 912         case 1: lro_single_flushes++;
 913                 break;
 914         case 2: lro_double_flushes++;
 915                 break;
 916         default: lro_good_flushes++;
 917                 break;
 918         }
 919         return;
 920 }