sys/dev/cxgbe/tom/t4_connect.c

   1 /*-
   2  * Copyright (c) 2012 Chelsio Communications, Inc.
   3  * All rights reserved.
   4  * Written by: Navdeep Parhar <np@FreeBSD.org>
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  */
  27
  28 #include <sys/cdefs.h>
  29 __FBSDID("$FreeBSD$");
  30
  31 #include "opt_inet.h"
  32 #include "opt_inet6.h"
  33
  34 #ifdef TCP_OFFLOAD
  35 #include <sys/param.h>
  36 #include <sys/types.h>
  37 #include <sys/kernel.h>
  38 #include <sys/ktr.h>
  39 #include <sys/module.h>
  40 #include <sys/protosw.h>
  41 #include <sys/domain.h>
  42 #include <sys/socket.h>
  43 #include <sys/socketvar.h>
  44 #include <net/ethernet.h>
  45 #include <net/if.h>
  46 #include <net/if_types.h>
  47 #include <net/if_vlan_var.h>
  48 #include <net/route.h>
  49 #include <netinet/in.h>
  50 #include <netinet/in_pcb.h>
  51 #include <netinet/ip.h>
  52 #include <netinet/tcp_var.h>
  53 #define TCPSTATES
  54 #include <netinet/tcp_fsm.h>
  55 #include <netinet/toecore.h>
  56
  57 #include "common/common.h"
  58 #include "common/t4_msg.h"
  59 #include "common/t4_regs.h"
  60 #include "tom/t4_tom_l2t.h"
  61 #include "tom/t4_tom.h"
  62
  63 /* atid services */
  64 static int alloc_atid(struct adapter *, void *);
  65 static void *lookup_atid(struct adapter *, int);
  66 static void free_atid(struct adapter *, int);
  67
  68 static int
  69 alloc_atid(struct adapter *sc, void *ctx)
  70 {
  71         struct tid_info *t = &sc->tids;
  72         int atid = -1;
  73
  74         mtx_lock(&t->atid_lock);
  75         if (t->afree) {
  76                 union aopen_entry *p = t->afree;
  77
  78                 atid = p - t->atid_tab;
  79                 t->afree = p->next;
  80                 p->data = ctx;
  81                 t->atids_in_use++;
  82         }
  83         mtx_unlock(&t->atid_lock);
  84         return (atid);
  85 }
  86
  87 static void *
  88 lookup_atid(struct adapter *sc, int atid)
  89 {
  90         struct tid_info *t = &sc->tids;
  91
  92         return (t->atid_tab[atid].data);
  93 }
  94
  95 static void
  96 free_atid(struct adapter *sc, int atid)
  97 {
  98         struct tid_info *t = &sc->tids;
  99         union aopen_entry *p = &t->atid_tab[atid];
 100
 101         mtx_lock(&t->atid_lock);
 102         p->next = t->afree;
 103         t->afree = p;
 104         t->atids_in_use--;
 105         mtx_unlock(&t->atid_lock);
 106 }
 107
 108 /*
 109  * Active open failed.
 110  */
 111 static int
 112 do_act_establish(struct sge_iq *iq, const struct rss_header *rss,
 113     struct mbuf *m)
 114 {
 115         struct adapter *sc = iq->adapter;
 116         const struct cpl_act_establish *cpl = (const void *)(rss + 1);
 117         unsigned int tid = GET_TID(cpl);
 118         unsigned int atid = G_TID_TID(ntohl(cpl->tos_atid));
 119         struct toepcb *toep = lookup_atid(sc, atid);
 120         struct inpcb *inp = toep->inp;
 121
 122         KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
 123         KASSERT(toep->tid == atid, ("%s: toep tid/atid mismatch", __func__));
 124
 125         CTR3(KTR_CXGBE, "%s: atid %u, tid %u", __func__, atid, tid);
 126         free_atid(sc, atid);
 127
 128         INP_WLOCK(inp);
 129         toep->tid = tid;
 130         insert_tid(sc, tid, toep);
 131         if (inp->inp_flags & INP_DROPPED) {
 132
 133                 /* socket closed by the kernel before hw told us it connected */
 134
 135                 send_flowc_wr(toep, NULL);
 136                 send_reset(sc, toep, be32toh(cpl->snd_isn));
 137                 goto done;
 138         }
 139
 140         make_established(toep, cpl->snd_isn, cpl->rcv_isn, cpl->tcp_opt);
 141 done:
 142         INP_WUNLOCK(inp);
 143         return (0);
 144 }
 145
 146 static inline int
 147 act_open_has_tid(unsigned int status)
 148 {
 149
 150         return (status != CPL_ERR_TCAM_FULL &&
 151             status != CPL_ERR_TCAM_PARITY &&
 152             status != CPL_ERR_CONN_EXIST &&
 153             status != CPL_ERR_ARP_MISS);
 154 }
 155
 156 /*
 157  * Convert an ACT_OPEN_RPL status to an errno.
 158  */
 159 static inline int
 160 act_open_rpl_status_to_errno(int status)
 161 {
 162
 163         switch (status) {
 164         case CPL_ERR_CONN_RESET:
 165                 return (ECONNREFUSED);
 166         case CPL_ERR_ARP_MISS:
 167                 return (EHOSTUNREACH);
 168         case CPL_ERR_CONN_TIMEDOUT:
 169                 return (ETIMEDOUT);
 170         case CPL_ERR_TCAM_FULL:
 171                 return (EAGAIN);
 172         case CPL_ERR_CONN_EXIST:
 173                 log(LOG_ERR, "ACTIVE_OPEN_RPL: 4-tuple in use\n");
 174                 return (EAGAIN);
 175         default:
 176                 return (EIO);
 177         }
 178 }
 179
 180 static int
 181 do_act_open_rpl(struct sge_iq *iq, const struct rss_header *rss,
 182     struct mbuf *m)
 183 {
 184         struct adapter *sc = iq->adapter;
 185         const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1);
 186         unsigned int atid = G_TID_TID(G_AOPEN_ATID(be32toh(cpl->atid_status)));
 187         unsigned int status = G_AOPEN_STATUS(be32toh(cpl->atid_status));
 188         struct toepcb *toep = lookup_atid(sc, atid);
 189         struct inpcb *inp = toep->inp;
 190         struct toedev *tod = &toep->td->tod;
 191         int rc;
 192
 193         KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
 194         KASSERT(toep->tid == atid, ("%s: toep tid/atid mismatch", __func__));
 195
 196         CTR3(KTR_CXGBE, "%s: atid %u, status %u ", __func__, atid, status);
 197
 198         /* Ignore negative advice */
 199         if (negative_advice(status))
 200                 return (0);
 201
 202         free_atid(sc, atid);
 203         toep->tid = -1;
 204
 205         if (status && act_open_has_tid(status))
 206                 release_tid(sc, GET_TID(cpl), toep->ctrlq);
 207
 208         rc = act_open_rpl_status_to_errno(status);
 209         if (rc != EAGAIN)
 210                 INP_INFO_WLOCK(&V_tcbinfo);
 211         INP_WLOCK(inp);
 212         toe_connect_failed(tod, inp, rc);
 213         final_cpl_received(toep);       /* unlocks inp */
 214         if (rc != EAGAIN)
 215                 INP_INFO_WUNLOCK(&V_tcbinfo);
 216
 217         return (0);
 218 }
 219
 220 /*
 221  * Options2 for active open.
 222  */
 223 static uint32_t
 224 calc_opt2a(struct socket *so, struct toepcb *toep)
 225 {
 226         struct tcpcb *tp = so_sototcpcb(so);
 227         struct port_info *pi = toep->port;
 228         struct adapter *sc = pi->adapter;
 229         uint32_t opt2 = 0;
 230
 231         if (tp->t_flags & TF_SACK_PERMIT)
 232                 opt2 |= F_SACK_EN;
 233
 234         if (tp->t_flags & TF_REQ_TSTMP)
 235                 opt2 |= F_TSTAMPS_EN;
 236
 237         if (tp->t_flags & TF_REQ_SCALE)
 238                 opt2 |= F_WND_SCALE_EN;
 239
 240         if (V_tcp_do_ecn)
 241                 opt2 |= F_CCTRL_ECN;
 242
 243         opt2 |= V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]);
 244         opt2 |= F_RSS_QUEUE_VALID | V_RSS_QUEUE(toep->ofld_rxq->iq.abs_id);
 245         if (is_t4(sc))
 246                 opt2 |= F_RX_COALESCE_VALID | V_RX_COALESCE(M_RX_COALESCE);
 247         else
 248                 opt2 |= F_T5_OPT_2_VALID | V_RX_COALESCE(M_RX_COALESCE);
 249
 250 #ifdef USE_DDP_RX_FLOW_CONTROL
 251         if (toep->ulp_mode == ULP_MODE_TCPDDP)
 252                 opt2 |= F_RX_FC_VALID | F_RX_FC_DDP;
 253 #endif
 254
 255         return (htobe32(opt2));
 256 }
 257
 258 void
 259 t4_init_connect_cpl_handlers(struct adapter *sc)
 260 {
 261
 262         t4_register_cpl_handler(sc, CPL_ACT_ESTABLISH, do_act_establish);
 263         t4_register_cpl_handler(sc, CPL_ACT_OPEN_RPL, do_act_open_rpl);
 264 }
 265
 266 #define DONT_OFFLOAD_ACTIVE_OPEN(x)     do { \
 267         reason = __LINE__; \
 268         rc = (x); \
 269         goto failed; \
 270 } while (0)
 271
 272 static inline int
 273 act_open_cpl_size(struct adapter *sc, int isipv6)
 274 {
 275         static const int sz_t4[] = {
 276                 sizeof (struct cpl_act_open_req),
 277                 sizeof (struct cpl_act_open_req6)
 278         };
 279         static const int sz_t5[] = {
 280                 sizeof (struct cpl_t5_act_open_req),
 281                 sizeof (struct cpl_t5_act_open_req6)
 282         };
 283
 284         if (is_t4(sc))
 285                 return (sz_t4[!!isipv6]);
 286         else
 287                 return (sz_t5[!!isipv6]);
 288 }
 289
 290 /*
 291  * active open (soconnect).
 292  *
 293  * State of affairs on entry:
 294  * soisconnecting (so_state |= SS_ISCONNECTING)
 295  * tcbinfo not locked (This has changed - used to be WLOCKed)
 296  * inp WLOCKed
 297  * tp->t_state = TCPS_SYN_SENT
 298  * rtalloc1, RT_UNLOCK on rt.
 299  */
 300 int
 301 t4_connect(struct toedev *tod, struct socket *so, struct rtentry *rt,
 302     struct sockaddr *nam)
 303 {
 304         struct adapter *sc = tod->tod_softc;
 305         struct tom_data *td = tod_td(tod);
 306         struct toepcb *toep = NULL;
 307         struct wrqe *wr = NULL;
 308         struct ifnet *rt_ifp = rt->rt_ifp;
 309         struct port_info *pi;
 310         int mtu_idx, rscale, qid_atid, rc, isipv6;
 311         struct inpcb *inp = sotoinpcb(so);
 312         struct tcpcb *tp = intotcpcb(inp);
 313         int reason;
 314
 315         INP_WLOCK_ASSERT(inp);
 316         KASSERT(nam->sa_family == AF_INET || nam->sa_family == AF_INET6,
 317             ("%s: dest addr %p has family %u", __func__, nam, nam->sa_family));
 318
 319         if (rt_ifp->if_type == IFT_ETHER)
 320                 pi = rt_ifp->if_softc;
 321         else if (rt_ifp->if_type == IFT_L2VLAN) {
 322                 struct ifnet *ifp = VLAN_COOKIE(rt_ifp);
 323
 324                 pi = ifp->if_softc;
 325         } else if (rt_ifp->if_type == IFT_IEEE8023ADLAG)
 326                 DONT_OFFLOAD_ACTIVE_OPEN(ENOSYS); /* XXX: implement lagg+TOE */
 327         else
 328                 DONT_OFFLOAD_ACTIVE_OPEN(ENOTSUP);
 329
 330         toep = alloc_toepcb(pi, -1, -1, M_NOWAIT);
 331         if (toep == NULL)
 332                 DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
 333
 334         toep->tid = alloc_atid(sc, toep);
 335         if (toep->tid < 0)
 336                 DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
 337
 338         toep->l2te = t4_l2t_get(pi, rt_ifp,
 339             rt->rt_flags & RTF_GATEWAY ? rt->rt_gateway : nam);
 340         if (toep->l2te == NULL)
 341                 DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
 342
 343         isipv6 = nam->sa_family == AF_INET6;
 344         wr = alloc_wrqe(act_open_cpl_size(sc, isipv6), toep->ctrlq);
 345         if (wr == NULL)
 346                 DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
 347
 348         if (sc->tt.ddp && (so->so_options & SO_NO_DDP) == 0)
 349                 set_tcpddp_ulp_mode(toep);
 350         else
 351                 toep->ulp_mode = ULP_MODE_NONE;
 352         SOCKBUF_LOCK(&so->so_rcv);
 353         /* opt0 rcv_bufsiz initially, assumes its normal meaning later */
 354         toep->rx_credits = min(select_rcv_wnd(so) >> 10, M_RCV_BUFSIZ);
 355         SOCKBUF_UNLOCK(&so->so_rcv);
 356
 357         /*
 358          * The kernel sets request_r_scale based on sb_max whereas we need to
 359          * take hardware's MAX_RCV_WND into account too.  This is normally a
 360          * no-op as MAX_RCV_WND is much larger than the default sb_max.
 361          */
 362         if (tp->t_flags & TF_REQ_SCALE)
 363                 rscale = tp->request_r_scale = select_rcv_wscale();
 364         else
 365                 rscale = 0;
 366         mtu_idx = find_best_mtu_idx(sc, &inp->inp_inc, 0);
 367         qid_atid = (toep->ofld_rxq->iq.abs_id << 14) | toep->tid;
 368
 369         if (isipv6) {
 370                 struct cpl_act_open_req6 *cpl = wrtod(wr);
 371
 372                 if ((inp->inp_vflag & INP_IPV6) == 0) {
 373                         /* XXX think about this a bit more */
 374                         log(LOG_ERR,
 375                             "%s: time to think about AF_INET6 + vflag 0x%x.\n",
 376                             __func__, inp->inp_vflag);
 377                         DONT_OFFLOAD_ACTIVE_OPEN(ENOTSUP);
 378                 }
 379
 380                 toep->ce = hold_lip(td, &inp->in6p_laddr);
 381                 if (toep->ce == NULL)
 382                         DONT_OFFLOAD_ACTIVE_OPEN(ENOENT);
 383
 384                 INIT_TP_WR(cpl, 0);
 385                 OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ6,
 386                     qid_atid));
 387
 388                 cpl->local_port = inp->inp_lport;
 389                 cpl->local_ip_hi = *(uint64_t *)&inp->in6p_laddr.s6_addr[0];
 390                 cpl->local_ip_lo = *(uint64_t *)&inp->in6p_laddr.s6_addr[8];
 391                 cpl->peer_port = inp->inp_fport;
 392                 cpl->peer_ip_hi = *(uint64_t *)&inp->in6p_faddr.s6_addr[0];
 393                 cpl->peer_ip_lo = *(uint64_t *)&inp->in6p_faddr.s6_addr[8];
 394                 cpl->opt0 = calc_opt0(so, pi, toep->l2te, mtu_idx, rscale,
 395                     toep->rx_credits, toep->ulp_mode);
 396                 cpl->opt2 = calc_opt2a(so, toep);
 397                 if (is_t4(sc)) {
 398                         cpl->params = select_ntuple(pi, toep->l2te,
 399                             sc->filter_mode);
 400                 } else {
 401                         struct cpl_t5_act_open_req6 *c5 = (void *)cpl;
 402
 403                         c5->rsvd = 0;
 404                         c5->params = select_ntuple(pi, toep->l2te,
 405                             sc->filter_mode);
 406                 }
 407         } else {
 408                 struct cpl_act_open_req *cpl = wrtod(wr);
 409
 410                 INIT_TP_WR(cpl, 0);
 411                 OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ,
 412                     qid_atid));
 413                 inp_4tuple_get(inp, &cpl->local_ip, &cpl->local_port,
 414                     &cpl->peer_ip, &cpl->peer_port);
 415                 cpl->opt0 = calc_opt0(so, pi, toep->l2te, mtu_idx, rscale,
 416                     toep->rx_credits, toep->ulp_mode);
 417                 cpl->opt2 = calc_opt2a(so, toep);
 418                 if (is_t4(sc)) {
 419                         cpl->params = select_ntuple(pi, toep->l2te,
 420                             sc->filter_mode);
 421                 } else {
 422                         struct cpl_t5_act_open_req6 *c5 = (void *)cpl;
 423
 424                         c5->rsvd = 0;
 425                         c5->params = select_ntuple(pi, toep->l2te,
 426                             sc->filter_mode);
 427                 }
 428         }
 429
 430         CTR5(KTR_CXGBE, "%s: atid %u (%s), toep %p, inp %p", __func__,
 431             toep->tid, tcpstates[tp->t_state], toep, inp);
 432
 433         offload_socket(so, toep);
 434         rc = t4_l2t_send(sc, wr, toep->l2te);
 435         if (rc == 0) {
 436                 toep->flags |= TPF_CPL_PENDING;
 437                 return (0);
 438         }
 439
 440         undo_offload_socket(so);
 441         reason = __LINE__;
 442 failed:
 443         CTR3(KTR_CXGBE, "%s: not offloading (%d), rc %d", __func__, reason, rc);
 444
 445         if (wr)
 446                 free_wrqe(wr);
 447
 448         if (toep) {
 449                 if (toep->tid >= 0)
 450                         free_atid(sc, toep->tid);
 451                 if (toep->l2te)
 452                         t4_l2t_release(toep->l2te);
 453                 if (toep->ce)
 454                         release_lip(td, toep->ce);
 455                 free_toepcb(toep);
 456         }
 457
 458         return (rc);
 459 }
 460 #endif