2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2012 Chelsio Communications, Inc.
6 * Written by: Navdeep Parhar <np@FreeBSD.org>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
34 #include "opt_inet6.h"
37 #include <sys/param.h>
38 #include <sys/types.h>
39 #include <sys/kernel.h>
41 #include <sys/module.h>
42 #include <sys/protosw.h>
43 #include <sys/refcount.h>
44 #include <sys/domain.h>
45 #include <sys/fnv_hash.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <net/ethernet.h>
51 #include <net/if_types.h>
52 #include <net/if_vlan_var.h>
53 #include <net/route.h>
54 #include <netinet/in.h>
55 #include <netinet/in_fib.h>
56 #include <netinet/in_pcb.h>
57 #include <netinet/ip.h>
58 #include <netinet/ip6.h>
59 #include <netinet6/in6_fib.h>
60 #include <netinet6/scope6_var.h>
61 #include <netinet/tcp_timer.h>
63 #include <netinet/tcp_fsm.h>
64 #include <netinet/tcp_var.h>
65 #include <netinet/toecore.h>
66 #include <netinet/cc/cc.h>
68 #include "common/common.h"
69 #include "common/t4_msg.h"
70 #include "common/t4_regs.h"
72 #include "tom/t4_tom_l2t.h"
73 #include "tom/t4_tom.h"
76 static int alloc_stid(struct adapter *, struct listen_ctx *, int);
77 static struct listen_ctx *lookup_stid(struct adapter *, int);
78 static void free_stid(struct adapter *, struct listen_ctx *);
81 static struct listen_ctx *alloc_lctx(struct adapter *, struct inpcb *,
83 static int free_lctx(struct adapter *, struct listen_ctx *);
84 static void hold_lctx(struct listen_ctx *);
85 static void listen_hash_add(struct adapter *, struct listen_ctx *);
86 static struct listen_ctx *listen_hash_find(struct adapter *, struct inpcb *);
87 static struct listen_ctx *listen_hash_del(struct adapter *, struct inpcb *);
88 static struct inpcb *release_lctx(struct adapter *, struct listen_ctx *);
90 static void send_reset_synqe(struct toedev *, struct synq_entry *);
93 alloc_stid(struct adapter *sc, struct listen_ctx *lctx, int isipv6)
95 struct tid_info *t = &sc->tids;
96 u_int stid, n, f, mask;
97 struct stid_region *sr = &lctx->stid_region;
100 * An IPv6 server needs 2 naturally aligned stids (1 stid = 4 cells) in
101 * the TCAM. The start of the stid region is properly aligned (the chip
102 * requires each region to be 128-cell aligned).
106 KASSERT((t->stid_base & mask) == 0 && (t->nstids & mask) == 0,
107 ("%s: stid region (%u, %u) not properly aligned. n = %u",
108 __func__, t->stid_base, t->nstids, n));
110 mtx_lock(&t->stid_lock);
111 if (n > t->nstids - t->stids_in_use) {
112 mtx_unlock(&t->stid_lock);
116 if (t->nstids_free_head >= n) {
118 * This allocation will definitely succeed because the region
119 * starts at a good alignment and we just checked we have enough
122 f = t->nstids_free_head & mask;
123 t->nstids_free_head -= n + f;
124 stid = t->nstids_free_head;
125 TAILQ_INSERT_HEAD(&t->stids, sr, link);
127 struct stid_region *s;
129 stid = t->nstids_free_head;
130 TAILQ_FOREACH(s, &t->stids, link) {
131 stid += s->used + s->free;
133 if (s->free >= n + f) {
136 TAILQ_INSERT_AFTER(&t->stids, s, sr, link);
141 if (__predict_false(stid != t->nstids)) {
142 panic("%s: stids TAILQ (%p) corrupt."
143 " At %d instead of %d at the end of the queue.",
144 __func__, &t->stids, stid, t->nstids);
147 mtx_unlock(&t->stid_lock);
154 t->stids_in_use += n;
155 t->stid_tab[stid] = lctx;
156 mtx_unlock(&t->stid_lock);
158 KASSERT(((stid + t->stid_base) & mask) == 0,
159 ("%s: EDOOFUS.", __func__));
160 return (stid + t->stid_base);
163 static struct listen_ctx *
164 lookup_stid(struct adapter *sc, int stid)
166 struct tid_info *t = &sc->tids;
168 return (t->stid_tab[stid - t->stid_base]);
172 free_stid(struct adapter *sc, struct listen_ctx *lctx)
174 struct tid_info *t = &sc->tids;
175 struct stid_region *sr = &lctx->stid_region;
176 struct stid_region *s;
178 KASSERT(sr->used > 0, ("%s: nonsense free (%d)", __func__, sr->used));
180 mtx_lock(&t->stid_lock);
181 s = TAILQ_PREV(sr, stid_head, link);
183 s->free += sr->used + sr->free;
185 t->nstids_free_head += sr->used + sr->free;
186 KASSERT(t->stids_in_use >= sr->used,
187 ("%s: stids_in_use (%u) < stids being freed (%u)", __func__,
188 t->stids_in_use, sr->used));
189 t->stids_in_use -= sr->used;
190 TAILQ_REMOVE(&t->stids, sr, link);
191 mtx_unlock(&t->stid_lock);
194 static struct listen_ctx *
195 alloc_lctx(struct adapter *sc, struct inpcb *inp, struct vi_info *vi)
197 struct listen_ctx *lctx;
199 INP_WLOCK_ASSERT(inp);
201 lctx = malloc(sizeof(struct listen_ctx), M_CXGBE, M_NOWAIT | M_ZERO);
205 lctx->stid = alloc_stid(sc, lctx, inp->inp_vflag & INP_IPV6);
206 if (lctx->stid < 0) {
211 if (inp->inp_vflag & INP_IPV6 &&
212 !IN6_ARE_ADDR_EQUAL(&in6addr_any, &inp->in6p_laddr)) {
213 lctx->ce = t4_hold_lip(sc, &inp->in6p_laddr, NULL);
214 if (lctx->ce == NULL) {
220 lctx->ctrlq = &sc->sge.ctrlq[vi->pi->port_id];
221 lctx->ofld_rxq = &sc->sge.ofld_rxq[vi->first_ofld_rxq];
222 refcount_init(&lctx->refcount, 1);
225 lctx->vnet = inp->inp_socket->so_vnet;
231 /* Don't call this directly, use release_lctx instead */
233 free_lctx(struct adapter *sc, struct listen_ctx *lctx)
235 struct inpcb *inp = lctx->inp;
237 INP_WLOCK_ASSERT(inp);
238 KASSERT(lctx->refcount == 0,
239 ("%s: refcount %d", __func__, lctx->refcount));
240 KASSERT(lctx->stid >= 0, ("%s: bad stid %d.", __func__, lctx->stid));
242 CTR4(KTR_CXGBE, "%s: stid %u, lctx %p, inp %p",
243 __func__, lctx->stid, lctx, lctx->inp);
246 t4_release_lip(sc, lctx->ce);
250 return (in_pcbrele_wlocked(inp));
254 hold_lctx(struct listen_ctx *lctx)
257 refcount_acquire(&lctx->refcount);
260 static inline uint32_t
261 listen_hashfn(void *key, u_long mask)
264 return (fnv_32_buf(&key, sizeof(key), FNV1_32_INIT) & mask);
268 * Add a listen_ctx entry to the listen hash table.
271 listen_hash_add(struct adapter *sc, struct listen_ctx *lctx)
273 struct tom_data *td = sc->tom_softc;
274 int bucket = listen_hashfn(lctx->inp, td->listen_mask);
276 mtx_lock(&td->lctx_hash_lock);
277 LIST_INSERT_HEAD(&td->listen_hash[bucket], lctx, link);
279 mtx_unlock(&td->lctx_hash_lock);
283 * Look for the listening socket's context entry in the hash and return it.
285 static struct listen_ctx *
286 listen_hash_find(struct adapter *sc, struct inpcb *inp)
288 struct tom_data *td = sc->tom_softc;
289 int bucket = listen_hashfn(inp, td->listen_mask);
290 struct listen_ctx *lctx;
292 mtx_lock(&td->lctx_hash_lock);
293 LIST_FOREACH(lctx, &td->listen_hash[bucket], link) {
294 if (lctx->inp == inp)
297 mtx_unlock(&td->lctx_hash_lock);
303 * Removes the listen_ctx structure for inp from the hash and returns it.
305 static struct listen_ctx *
306 listen_hash_del(struct adapter *sc, struct inpcb *inp)
308 struct tom_data *td = sc->tom_softc;
309 int bucket = listen_hashfn(inp, td->listen_mask);
310 struct listen_ctx *lctx, *l;
312 mtx_lock(&td->lctx_hash_lock);
313 LIST_FOREACH_SAFE(lctx, &td->listen_hash[bucket], link, l) {
314 if (lctx->inp == inp) {
315 LIST_REMOVE(lctx, link);
320 mtx_unlock(&td->lctx_hash_lock);
326 * Releases a hold on the lctx. Must be called with the listening socket's inp
327 * locked. The inp may be freed by this function and it returns NULL to
330 static struct inpcb *
331 release_lctx(struct adapter *sc, struct listen_ctx *lctx)
333 struct inpcb *inp = lctx->inp;
336 INP_WLOCK_ASSERT(inp);
337 if (refcount_release(&lctx->refcount))
338 inp_freed = free_lctx(sc, lctx);
340 return (inp_freed ? NULL : inp);
344 send_reset_synqe(struct toedev *tod, struct synq_entry *synqe)
346 struct adapter *sc = tod->tod_softc;
347 struct mbuf *m = synqe->syn;
348 struct ifnet *ifp = m->m_pkthdr.rcvif;
349 struct vi_info *vi = ifp->if_softc;
350 struct port_info *pi = vi->pi;
351 struct l2t_entry *e = &sc->l2t->l2tab[synqe->l2e_idx];
353 struct fw_flowc_wr *flowc;
354 struct cpl_abort_req *req;
356 struct sge_wrq *ofld_txq;
357 struct sge_ofld_rxq *ofld_rxq;
358 const int nparams = 6;
359 const u_int pfvf = sc->pf << S_FW_VIID_PFN;
361 INP_WLOCK_ASSERT(synqe->lctx->inp);
363 CTR5(KTR_CXGBE, "%s: synqe %p (0x%x), tid %d%s",
364 __func__, synqe, synqe->flags, synqe->tid,
365 synqe->flags & TPF_ABORT_SHUTDOWN ?
366 " (abort already in progress)" : "");
367 if (synqe->flags & TPF_ABORT_SHUTDOWN)
368 return; /* abort already in progress */
369 synqe->flags |= TPF_ABORT_SHUTDOWN;
371 ofld_txq = &sc->sge.ofld_txq[synqe->txqid];
372 ofld_rxq = &sc->sge.ofld_rxq[synqe->rxqid];
374 /* The wrqe will have two WRs - a flowc followed by an abort_req */
375 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
377 wr = alloc_wrqe(roundup2(flowclen, EQ_ESIZE) + sizeof(*req), ofld_txq);
380 panic("%s: allocation failure.", __func__);
383 req = (void *)((caddr_t)flowc + roundup2(flowclen, EQ_ESIZE));
385 /* First the flowc ... */
386 memset(flowc, 0, wr->wr_len);
387 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
388 V_FW_FLOWC_WR_NPARAMS(nparams));
389 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
390 V_FW_WR_FLOWID(synqe->tid));
391 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
392 flowc->mnemval[0].val = htobe32(pfvf);
393 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
394 flowc->mnemval[1].val = htobe32(pi->tx_chan);
395 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
396 flowc->mnemval[2].val = htobe32(pi->tx_chan);
397 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
398 flowc->mnemval[3].val = htobe32(ofld_rxq->iq.abs_id);
399 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF;
400 flowc->mnemval[4].val = htobe32(512);
401 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS;
402 flowc->mnemval[5].val = htobe32(512);
403 synqe->flags |= TPF_FLOWC_WR_SENT;
405 /* ... then ABORT request */
406 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, synqe->tid);
407 req->rsvd0 = 0; /* don't have a snd_nxt */
408 req->rsvd1 = 1; /* no data sent yet */
409 req->cmd = CPL_ABORT_SEND_RST;
411 t4_l2t_send(sc, wr, e);
415 create_server(struct adapter *sc, struct listen_ctx *lctx)
418 struct cpl_pass_open_req *req;
419 struct inpcb *inp = lctx->inp;
421 wr = alloc_wrqe(sizeof(*req), lctx->ctrlq);
423 log(LOG_ERR, "%s: allocation failure", __func__);
429 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_PASS_OPEN_REQ, lctx->stid));
430 req->local_port = inp->inp_lport;
432 req->local_ip = inp->inp_laddr.s_addr;
434 req->opt0 = htobe64(V_TX_CHAN(lctx->ctrlq->eq.tx_chan));
435 req->opt1 = htobe64(V_CONN_POLICY(CPL_CONN_POLICY_ASK) |
436 F_SYN_RSS_ENABLE | V_SYN_RSS_QUEUE(lctx->ofld_rxq->iq.abs_id));
443 create_server6(struct adapter *sc, struct listen_ctx *lctx)
446 struct cpl_pass_open_req6 *req;
447 struct inpcb *inp = lctx->inp;
449 wr = alloc_wrqe(sizeof(*req), lctx->ctrlq);
451 log(LOG_ERR, "%s: allocation failure", __func__);
457 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_PASS_OPEN_REQ6, lctx->stid));
458 req->local_port = inp->inp_lport;
460 req->local_ip_hi = *(uint64_t *)&inp->in6p_laddr.s6_addr[0];
461 req->local_ip_lo = *(uint64_t *)&inp->in6p_laddr.s6_addr[8];
464 req->opt0 = htobe64(V_TX_CHAN(lctx->ctrlq->eq.tx_chan));
465 req->opt1 = htobe64(V_CONN_POLICY(CPL_CONN_POLICY_ASK) |
466 F_SYN_RSS_ENABLE | V_SYN_RSS_QUEUE(lctx->ofld_rxq->iq.abs_id));
473 destroy_server(struct adapter *sc, struct listen_ctx *lctx)
476 struct cpl_close_listsvr_req *req;
478 wr = alloc_wrqe(sizeof(*req), lctx->ctrlq);
481 panic("%s: allocation failure.", __func__);
486 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_LISTSRV_REQ,
488 req->reply_ctrl = htobe16(lctx->ofld_rxq->iq.abs_id);
489 req->rsvd = htobe16(0);
496 * Start a listening server by sending a passive open request to HW.
498 * Can't take adapter lock here and access to sc->flags,
499 * sc->offload_map, if_capenable are all race prone.
502 t4_listen_start(struct toedev *tod, struct tcpcb *tp)
504 struct adapter *sc = tod->tod_softc;
506 struct port_info *pi;
507 struct inpcb *inp = tp->t_inpcb;
508 struct listen_ctx *lctx;
510 struct offload_settings settings;
512 INP_WLOCK_ASSERT(inp);
514 rw_rlock(&sc->policy_lock);
515 settings = *lookup_offload_policy(sc, OPEN_TYPE_LISTEN, NULL,
516 EVL_MAKETAG(0xfff, 0, 0), inp);
517 rw_runlock(&sc->policy_lock);
518 if (!settings.offload)
521 /* Don't start a hardware listener for any loopback address. */
522 if (inp->inp_vflag & INP_IPV6 && IN6_IS_ADDR_LOOPBACK(&inp->in6p_laddr))
524 if (!(inp->inp_vflag & INP_IPV6) &&
525 IN_LOOPBACK(ntohl(inp->inp_laddr.s_addr)))
530 log(LOG_ERR, "%s: listen request ignored, %s is busy",
531 __func__, device_get_nameunit(sc->dev));
535 KASSERT(uld_active(sc, ULD_TOM),
536 ("%s: TOM not initialized", __func__));
540 * Find an initialized VI with IFCAP_TOE (4 or 6). We'll use the first
541 * such VI's queues to send the passive open and receive the reply to
544 * XXX: need a way to mark a port in use by offload. if_cxgbe should
545 * then reject any attempt to bring down such a port (and maybe reject
546 * attempts to disable IFCAP_TOE on that port too?).
548 for_each_port(sc, i) {
550 for_each_vi(pi, v, vi) {
551 if (vi->flags & VI_INIT_DONE &&
552 vi->ifp->if_capenable & IFCAP_TOE)
556 goto done; /* no port that's UP with IFCAP_TOE enabled */
559 if (listen_hash_find(sc, inp) != NULL)
560 goto done; /* already setup */
562 lctx = alloc_lctx(sc, inp, vi);
565 "%s: listen request ignored, %s couldn't allocate lctx\n",
566 __func__, device_get_nameunit(sc->dev));
569 listen_hash_add(sc, lctx);
571 CTR6(KTR_CXGBE, "%s: stid %u (%s), lctx %p, inp %p vflag 0x%x",
572 __func__, lctx->stid, tcpstates[tp->t_state], lctx, inp,
575 if (inp->inp_vflag & INP_IPV6)
576 rc = create_server6(sc, lctx);
578 rc = create_server(sc, lctx);
580 log(LOG_ERR, "%s: %s failed to create hw listener: %d.\n",
581 __func__, device_get_nameunit(sc->dev), rc);
582 (void) listen_hash_del(sc, inp);
583 inp = release_lctx(sc, lctx);
584 /* can't be freed, host stack has a reference */
585 KASSERT(inp != NULL, ("%s: inp freed", __func__));
588 lctx->flags |= LCTX_RPL_PENDING;
597 t4_listen_stop(struct toedev *tod, struct tcpcb *tp)
599 struct listen_ctx *lctx;
600 struct adapter *sc = tod->tod_softc;
601 struct inpcb *inp = tp->t_inpcb;
603 INP_WLOCK_ASSERT(inp);
605 lctx = listen_hash_del(sc, inp);
607 return (ENOENT); /* no hardware listener for this inp */
609 CTR4(KTR_CXGBE, "%s: stid %u, lctx %p, flags %x", __func__, lctx->stid,
613 * If the reply to the PASS_OPEN is still pending we'll wait for it to
614 * arrive and clean up when it does.
616 if (lctx->flags & LCTX_RPL_PENDING) {
617 return (EINPROGRESS);
620 destroy_server(sc, lctx);
624 static inline struct synq_entry *
625 alloc_synqe(struct adapter *sc __unused, struct listen_ctx *lctx, int flags)
627 struct synq_entry *synqe;
629 INP_WLOCK_ASSERT(lctx->inp);
630 MPASS(flags == M_WAITOK || flags == M_NOWAIT);
632 synqe = malloc(sizeof(*synqe), M_CXGBE, flags);
633 if (__predict_true(synqe != NULL)) {
634 synqe->flags = TPF_SYNQE;
635 refcount_init(&synqe->refcnt, 1);
637 hold_lctx(lctx); /* Every synqe has a ref on its lctx. */
645 hold_synqe(struct synq_entry *synqe)
648 refcount_acquire(&synqe->refcnt);
651 static inline struct inpcb *
652 release_synqe(struct adapter *sc, struct synq_entry *synqe)
656 MPASS(synqe->flags & TPF_SYNQE);
657 MPASS(synqe->lctx != NULL);
659 inp = synqe->lctx->inp;
661 INP_WLOCK_ASSERT(inp);
663 if (refcount_release(&synqe->refcnt)) {
664 inp = release_lctx(sc, synqe->lctx);
666 free(synqe, M_CXGBE);
673 t4_syncache_added(struct toedev *tod __unused, void *arg)
675 struct synq_entry *synqe = arg;
681 t4_syncache_removed(struct toedev *tod, void *arg)
683 struct adapter *sc = tod->tod_softc;
684 struct synq_entry *synqe = arg;
685 struct inpcb *inp = synqe->lctx->inp;
688 * XXX: this is a LOR but harmless when running from the softclock.
691 inp = release_synqe(sc, synqe);
697 t4_syncache_respond(struct toedev *tod, void *arg, struct mbuf *m)
699 struct synq_entry *synqe = arg;
701 if (atomic_fetchadd_int(&synqe->ok_to_respond, 1) == 0) {
703 struct ip *ip = mtod(m, struct ip *);
706 if (ip->ip_v == IPVERSION)
707 th = (void *)(ip + 1);
709 th = (void *)((struct ip6_hdr *)ip + 1);
710 bzero(&to, sizeof(to));
711 tcp_dooptions(&to, (void *)(th + 1),
712 (th->th_off << 2) - sizeof(*th), TO_SYN);
714 /* save these for later */
715 synqe->iss = be32toh(th->th_seq);
716 synqe->irs = be32toh(th->th_ack) - 1;
717 synqe->ts = to.to_tsval;
720 m_freem(m); /* don't need this any more */
725 do_pass_open_rpl(struct sge_iq *iq, const struct rss_header *rss,
728 struct adapter *sc = iq->adapter;
729 const struct cpl_pass_open_rpl *cpl = (const void *)(rss + 1);
730 int stid = GET_TID(cpl);
731 unsigned int status = cpl->status;
732 struct listen_ctx *lctx = lookup_stid(sc, stid);
733 struct inpcb *inp = lctx->inp;
735 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
738 KASSERT(opcode == CPL_PASS_OPEN_RPL,
739 ("%s: unexpected opcode 0x%x", __func__, opcode));
740 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
741 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__));
745 CTR4(KTR_CXGBE, "%s: stid %d, status %u, flags 0x%x",
746 __func__, stid, status, lctx->flags);
748 lctx->flags &= ~LCTX_RPL_PENDING;
750 if (status != CPL_ERR_NONE)
751 log(LOG_ERR, "listener (stid %u) failed: %d\n", stid, status);
755 * If the inp has been dropped (listening socket closed) then
756 * listen_stop must have run and taken the inp out of the hash.
758 if (inp->inp_flags & INP_DROPPED) {
759 KASSERT(listen_hash_del(sc, inp) == NULL,
760 ("%s: inp %p still in listen hash", __func__, inp));
764 if (inp->inp_flags & INP_DROPPED && status != CPL_ERR_NONE) {
765 if (release_lctx(sc, lctx) != NULL)
771 * Listening socket stopped listening earlier and now the chip tells us
772 * it has started the hardware listener. Stop it; the lctx will be
773 * released in do_close_server_rpl.
775 if (inp->inp_flags & INP_DROPPED) {
776 destroy_server(sc, lctx);
782 * Failed to start hardware listener. Take inp out of the hash and
783 * release our reference on it. An error message has been logged
786 if (status != CPL_ERR_NONE) {
787 listen_hash_del(sc, inp);
788 if (release_lctx(sc, lctx) != NULL)
793 /* hardware listener open for business */
800 do_close_server_rpl(struct sge_iq *iq, const struct rss_header *rss,
803 struct adapter *sc = iq->adapter;
804 const struct cpl_close_listsvr_rpl *cpl = (const void *)(rss + 1);
805 int stid = GET_TID(cpl);
806 unsigned int status = cpl->status;
807 struct listen_ctx *lctx = lookup_stid(sc, stid);
808 struct inpcb *inp = lctx->inp;
810 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
813 KASSERT(opcode == CPL_CLOSE_LISTSRV_RPL,
814 ("%s: unexpected opcode 0x%x", __func__, opcode));
815 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
816 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__));
818 CTR3(KTR_CXGBE, "%s: stid %u, status %u", __func__, stid, status);
820 if (status != CPL_ERR_NONE) {
821 log(LOG_ERR, "%s: failed (%u) to close listener for stid %u\n",
822 __func__, status, stid);
827 inp = release_lctx(sc, lctx);
835 done_with_synqe(struct adapter *sc, struct synq_entry *synqe)
837 struct listen_ctx *lctx = synqe->lctx;
838 struct inpcb *inp = lctx->inp;
839 struct l2t_entry *e = &sc->l2t->l2tab[synqe->l2e_idx];
842 INP_WLOCK_ASSERT(inp);
843 ntids = inp->inp_vflag & INP_IPV6 ? 2 : 1;
845 remove_tid(sc, synqe->tid, ntids);
846 release_tid(sc, synqe->tid, lctx->ctrlq);
848 inp = release_synqe(sc, synqe);
854 synack_failure_cleanup(struct adapter *sc, int tid)
856 struct synq_entry *synqe = lookup_tid(sc, tid);
858 INP_WLOCK(synqe->lctx->inp);
859 done_with_synqe(sc, synqe);
863 do_abort_req_synqe(struct sge_iq *iq, const struct rss_header *rss,
866 struct adapter *sc = iq->adapter;
867 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
868 unsigned int tid = GET_TID(cpl);
869 struct synq_entry *synqe = lookup_tid(sc, tid);
870 struct listen_ctx *lctx = synqe->lctx;
871 struct inpcb *inp = lctx->inp;
872 struct sge_wrq *ofld_txq;
874 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
877 KASSERT(opcode == CPL_ABORT_REQ_RSS,
878 ("%s: unexpected opcode 0x%x", __func__, opcode));
879 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
880 KASSERT(synqe->tid == tid, ("%s: toep tid mismatch", __func__));
882 CTR6(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x), lctx %p, status %d",
883 __func__, tid, synqe, synqe->flags, synqe->lctx, cpl->status);
885 if (negative_advice(cpl->status))
886 return (0); /* Ignore negative advice */
890 ofld_txq = &sc->sge.ofld_txq[synqe->txqid];
893 * If we'd initiated an abort earlier the reply to it is responsible for
894 * cleaning up resources. Otherwise we tear everything down right here
895 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
897 if (synqe->flags & TPF_ABORT_SHUTDOWN) {
902 done_with_synqe(sc, synqe);
903 /* inp lock released by done_with_synqe */
905 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
910 do_abort_rpl_synqe(struct sge_iq *iq, const struct rss_header *rss,
913 struct adapter *sc = iq->adapter;
914 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
915 unsigned int tid = GET_TID(cpl);
916 struct synq_entry *synqe = lookup_tid(sc, tid);
917 struct listen_ctx *lctx = synqe->lctx;
918 struct inpcb *inp = lctx->inp;
920 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
923 KASSERT(opcode == CPL_ABORT_RPL_RSS,
924 ("%s: unexpected opcode 0x%x", __func__, opcode));
925 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
926 KASSERT(synqe->tid == tid, ("%s: toep tid mismatch", __func__));
928 CTR6(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x), lctx %p, status %d",
929 __func__, tid, synqe, synqe->flags, synqe->lctx, cpl->status);
932 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
933 ("%s: wasn't expecting abort reply for synqe %p (0x%x)",
934 __func__, synqe, synqe->flags));
936 done_with_synqe(sc, synqe);
937 /* inp lock released by done_with_synqe */
943 t4_offload_socket(struct toedev *tod, void *arg, struct socket *so)
945 struct adapter *sc = tod->tod_softc;
946 struct synq_entry *synqe = arg;
948 struct inpcb *inp = sotoinpcb(so);
950 struct toepcb *toep = synqe->toep;
952 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); /* prevents bad race with accept() */
953 INP_WLOCK_ASSERT(inp);
954 KASSERT(synqe->flags & TPF_SYNQE,
955 ("%s: %p not a synq_entry?", __func__, arg));
956 MPASS(toep->tid == synqe->tid);
958 offload_socket(so, toep);
959 make_established(toep, synqe->iss, synqe->irs, synqe->tcp_opt);
960 toep->flags |= TPF_CPL_PENDING;
961 update_tid(sc, synqe->tid, toep);
962 synqe->flags |= TPF_SYNQE_EXPANDED;
966 save_qids_in_synqe(struct synq_entry *synqe, struct vi_info *vi,
967 struct offload_settings *s)
969 uint32_t txqid, rxqid;
971 if (s->txq >= 0 && s->txq < vi->nofldtxq)
974 txqid = arc4random() % vi->nofldtxq;
975 txqid += vi->first_ofld_txq;
977 if (s->rxq >= 0 && s->rxq < vi->nofldrxq)
980 rxqid = arc4random() % vi->nofldrxq;
981 rxqid += vi->first_ofld_rxq;
983 synqe->txqid = txqid;
984 synqe->rxqid = rxqid;
988 t4opt_to_tcpopt(const struct tcp_options *t4opt, struct tcpopt *to)
990 bzero(to, sizeof(*to));
993 to->to_flags |= TOF_MSS;
994 to->to_mss = be16toh(t4opt->mss);
998 to->to_flags |= TOF_SCALE;
999 to->to_wscale = t4opt->wsf;
1003 to->to_flags |= TOF_TS;
1006 to->to_flags |= TOF_SACKPERM;
1010 * Options2 for passive open.
1013 calc_opt2p(struct adapter *sc, struct port_info *pi, int rxqid,
1014 const struct tcp_options *tcpopt, struct tcphdr *th, int ulp_mode,
1015 struct cc_algo *cc, const struct offload_settings *s)
1017 struct sge_ofld_rxq *ofld_rxq = &sc->sge.ofld_rxq[rxqid];
1021 * rx flow control, rx coalesce, congestion control, and tx pace are all
1022 * explicitly set by the driver. On T5+ the ISS is also set by the
1023 * driver to the value picked by the kernel.
1026 opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID;
1027 opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID;
1029 opt2 |= F_T5_OPT_2_VALID; /* all 4 valid */
1030 opt2 |= F_T5_ISS; /* ISS provided in CPL */
1033 if (tcpopt->sack && (s->sack > 0 || (s->sack < 0 && V_tcp_do_rfc1323)))
1036 if (tcpopt->tstamp &&
1037 (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323)))
1038 opt2 |= F_TSTAMPS_EN;
1040 if (tcpopt->wsf < 15 && V_tcp_do_rfc1323)
1041 opt2 |= F_WND_SCALE_EN;
1043 if (th->th_flags & (TH_ECE | TH_CWR) &&
1044 (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn)))
1045 opt2 |= F_CCTRL_ECN;
1047 /* XXX: F_RX_CHANNEL for multiple rx c-chan support goes here. */
1049 opt2 |= V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]);
1051 /* These defaults are subject to ULP specific fixups later. */
1052 opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
1056 if (s->cong_algo >= 0)
1057 opt2 |= V_CONG_CNTRL(s->cong_algo);
1058 else if (sc->tt.cong_algorithm >= 0)
1059 opt2 |= V_CONG_CNTRL(sc->tt.cong_algorithm & M_CONG_CNTRL);
1061 if (strcasecmp(cc->name, "reno") == 0)
1062 opt2 |= V_CONG_CNTRL(CONG_ALG_RENO);
1063 else if (strcasecmp(cc->name, "tahoe") == 0)
1064 opt2 |= V_CONG_CNTRL(CONG_ALG_TAHOE);
1065 if (strcasecmp(cc->name, "newreno") == 0)
1066 opt2 |= V_CONG_CNTRL(CONG_ALG_NEWRENO);
1067 if (strcasecmp(cc->name, "highspeed") == 0)
1068 opt2 |= V_CONG_CNTRL(CONG_ALG_HIGHSPEED);
1071 * Use newreno in case the algorithm selected by the
1072 * host stack is not supported by the hardware.
1074 opt2 |= V_CONG_CNTRL(CONG_ALG_NEWRENO);
1078 if (s->rx_coalesce > 0 || (s->rx_coalesce < 0 && sc->tt.rx_coalesce))
1079 opt2 |= V_RX_COALESCE(M_RX_COALESCE);
1081 /* Note that ofld_rxq is already set according to s->rxq. */
1082 opt2 |= F_RSS_QUEUE_VALID;
1083 opt2 |= V_RSS_QUEUE(ofld_rxq->iq.abs_id);
1085 #ifdef USE_DDP_RX_FLOW_CONTROL
1086 if (ulp_mode == ULP_MODE_TCPDDP)
1087 opt2 |= F_RX_FC_DDP;
1090 if (ulp_mode == ULP_MODE_TLS) {
1091 opt2 &= ~V_RX_COALESCE(M_RX_COALESCE);
1092 opt2 |= F_RX_FC_DISABLE;
1095 return (htobe32(opt2));
1099 pass_accept_req_to_protohdrs(struct adapter *sc, const struct mbuf *m,
1100 struct in_conninfo *inc, struct tcphdr *th)
1102 const struct cpl_pass_accept_req *cpl = mtod(m, const void *);
1103 const struct ether_header *eh;
1104 unsigned int hlen = be32toh(cpl->hdr_len);
1106 const struct tcphdr *tcp;
1108 eh = (const void *)(cpl + 1);
1109 if (chip_id(sc) >= CHELSIO_T6) {
1110 l3hdr = ((uintptr_t)eh + G_T6_ETH_HDR_LEN(hlen));
1111 tcp = (const void *)(l3hdr + G_T6_IP_HDR_LEN(hlen));
1113 l3hdr = ((uintptr_t)eh + G_ETH_HDR_LEN(hlen));
1114 tcp = (const void *)(l3hdr + G_IP_HDR_LEN(hlen));
1118 bzero(inc, sizeof(*inc));
1119 inc->inc_fport = tcp->th_sport;
1120 inc->inc_lport = tcp->th_dport;
1121 if (((struct ip *)l3hdr)->ip_v == IPVERSION) {
1122 const struct ip *ip = (const void *)l3hdr;
1124 inc->inc_faddr = ip->ip_src;
1125 inc->inc_laddr = ip->ip_dst;
1127 const struct ip6_hdr *ip6 = (const void *)l3hdr;
1129 inc->inc_flags |= INC_ISIPV6;
1130 inc->inc6_faddr = ip6->ip6_src;
1131 inc->inc6_laddr = ip6->ip6_dst;
1136 bcopy(tcp, th, sizeof(*th));
1137 tcp_fields_to_host(th); /* just like tcp_input */
1141 static struct l2t_entry *
1142 get_l2te_for_nexthop(struct port_info *pi, struct ifnet *ifp,
1143 struct in_conninfo *inc)
1145 struct l2t_entry *e;
1146 struct sockaddr_in6 sin6;
1147 struct sockaddr *dst = (void *)&sin6;
1149 if (inc->inc_flags & INC_ISIPV6) {
1150 struct nhop6_basic nh6;
1152 bzero(dst, sizeof(struct sockaddr_in6));
1153 dst->sa_len = sizeof(struct sockaddr_in6);
1154 dst->sa_family = AF_INET6;
1156 if (IN6_IS_ADDR_LINKLOCAL(&inc->inc6_laddr)) {
1157 /* no need for route lookup */
1158 e = t4_l2t_get(pi, ifp, dst);
1162 if (fib6_lookup_nh_basic(RT_DEFAULT_FIB, &inc->inc6_faddr,
1163 0, 0, 0, &nh6) != 0)
1165 if (nh6.nh_ifp != ifp)
1167 ((struct sockaddr_in6 *)dst)->sin6_addr = nh6.nh_addr;
1169 struct nhop4_basic nh4;
1171 dst->sa_len = sizeof(struct sockaddr_in);
1172 dst->sa_family = AF_INET;
1174 if (fib4_lookup_nh_basic(RT_DEFAULT_FIB, inc->inc_faddr, 0, 0,
1177 if (nh4.nh_ifp != ifp)
1179 ((struct sockaddr_in *)dst)->sin_addr = nh4.nh_addr;
1182 e = t4_l2t_get(pi, ifp, dst);
1187 send_synack(struct adapter *sc, struct synq_entry *synqe, uint64_t opt0,
1188 uint32_t opt2, int tid)
1191 struct cpl_pass_accept_rpl *rpl;
1192 struct l2t_entry *e = &sc->l2t->l2tab[synqe->l2e_idx];
1194 wr = alloc_wrqe(is_t4(sc) ? sizeof(struct cpl_pass_accept_rpl) :
1195 sizeof(struct cpl_t5_pass_accept_rpl), &sc->sge.ctrlq[0]);
1201 INIT_TP_WR_MIT_CPL(rpl, CPL_PASS_ACCEPT_RPL, tid);
1203 struct cpl_t5_pass_accept_rpl *rpl5 = (void *)rpl;
1205 INIT_TP_WR_MIT_CPL(rpl5, CPL_PASS_ACCEPT_RPL, tid);
1206 rpl5->iss = htobe32(synqe->iss);
1211 return (t4_l2t_send(sc, wr, e));
1214 #define REJECT_PASS_ACCEPT_REQ(tunnel) do { \
1219 reject_reason = __LINE__; \
1224 * The context associated with a tid entry via insert_tid could be a synq_entry
1225 * or a toepcb. The only way CPL handlers can tell is via a bit in these flags.
1227 CTASSERT(offsetof(struct toepcb, flags) == offsetof(struct synq_entry, flags));
1230 * Incoming SYN on a listening socket.
1232 * XXX: Every use of ifp in this routine has a bad race with up/down, toe/-toe,
1236 do_pass_accept_req(struct sge_iq *iq, const struct rss_header *rss,
1239 struct adapter *sc = iq->adapter;
1241 const struct cpl_pass_accept_req *cpl = mtod(m, const void *);
1242 unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid));
1243 unsigned int tid = GET_TID(cpl);
1244 struct listen_ctx *lctx = lookup_stid(sc, stid);
1247 struct in_conninfo inc;
1250 struct port_info *pi;
1252 struct ifnet *hw_ifp, *ifp;
1253 struct l2t_entry *e = NULL;
1254 struct synq_entry *synqe = NULL;
1255 int reject_reason, v, ntids;
1256 uint16_t vid, l2info;
1257 struct epoch_tracker et;
1259 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1261 struct offload_settings settings;
1263 KASSERT(opcode == CPL_PASS_ACCEPT_REQ,
1264 ("%s: unexpected opcode 0x%x", __func__, opcode));
1265 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__));
1267 CTR4(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p", __func__, stid, tid,
1270 CURVNET_SET(lctx->vnet); /* before any potential REJECT */
1273 * Use the MAC index to lookup the associated VI. If this SYN didn't
1274 * match a perfect MAC filter, punt.
1276 l2info = be16toh(cpl->l2info);
1277 pi = sc->port[G_SYN_INTF(l2info)];
1278 if (!(l2info & F_SYN_XACT_MATCH)) {
1279 REJECT_PASS_ACCEPT_REQ(false);
1281 for_each_vi(pi, v, vi) {
1282 if (vi->xact_addr_filt == G_SYN_MAC_IDX(l2info))
1285 REJECT_PASS_ACCEPT_REQ(false);
1287 hw_ifp = vi->ifp; /* the cxgbe ifnet */
1288 m->m_pkthdr.rcvif = hw_ifp;
1289 tod = TOEDEV(hw_ifp);
1292 * Don't offload if the peer requested a TCP option that's not known to
1293 * the silicon. Send the SYN to the kernel instead.
1295 if (__predict_false(cpl->tcpopt.unknown))
1296 REJECT_PASS_ACCEPT_REQ(true);
1299 * Figure out if there is a pseudo interface (vlan, lagg, etc.)
1300 * involved. Don't offload if the SYN had a VLAN tag and the vid
1301 * doesn't match anything on this interface.
1303 * XXX: lagg support, lagg + vlan support.
1305 vid = EVL_VLANOFTAG(be16toh(cpl->vlan));
1306 if (vid != 0xfff && vid != 0) {
1307 ifp = VLAN_DEVAT(hw_ifp, vid);
1309 REJECT_PASS_ACCEPT_REQ(true);
1314 * Don't offload if the ifnet that the SYN came in on is not in the same
1315 * vnet as the listening socket.
1317 if (lctx->vnet != ifp->if_vnet)
1318 REJECT_PASS_ACCEPT_REQ(true);
1320 pass_accept_req_to_protohdrs(sc, m, &inc, &th);
1321 if (inc.inc_flags & INC_ISIPV6) {
1323 /* Don't offload if the ifcap isn't enabled */
1324 if ((ifp->if_capenable & IFCAP_TOE6) == 0)
1325 REJECT_PASS_ACCEPT_REQ(true);
1328 * SYN must be directed to an IP6 address on this ifnet. This
1329 * is more restrictive than in6_localip.
1331 if (!in6_ifhasaddr(ifp, &inc.inc6_laddr))
1332 REJECT_PASS_ACCEPT_REQ(true);
1337 /* Don't offload if the ifcap isn't enabled */
1338 if ((ifp->if_capenable & IFCAP_TOE4) == 0)
1339 REJECT_PASS_ACCEPT_REQ(true);
1342 * SYN must be directed to an IP address on this ifnet. This
1343 * is more restrictive than in_localip.
1345 if (!in_ifhasaddr(ifp, inc.inc_laddr))
1346 REJECT_PASS_ACCEPT_REQ(true);
1351 e = get_l2te_for_nexthop(pi, ifp, &inc);
1353 REJECT_PASS_ACCEPT_REQ(true);
1355 /* Don't offload if the 4-tuple is already in use */
1356 INP_INFO_RLOCK_ET(&V_tcbinfo, et); /* for 4-tuple check */
1357 if (toe_4tuple_check(&inc, &th, ifp) != 0) {
1358 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1359 REJECT_PASS_ACCEPT_REQ(false);
1361 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1363 inp = lctx->inp; /* listening socket, not owned by TOE */
1366 /* Don't offload if the listening socket has closed */
1367 if (__predict_false(inp->inp_flags & INP_DROPPED)) {
1369 REJECT_PASS_ACCEPT_REQ(false);
1371 so = inp->inp_socket;
1372 rw_rlock(&sc->policy_lock);
1373 settings = *lookup_offload_policy(sc, OPEN_TYPE_PASSIVE, m,
1374 EVL_MAKETAG(0xfff, 0, 0), inp);
1375 rw_runlock(&sc->policy_lock);
1376 if (!settings.offload) {
1378 REJECT_PASS_ACCEPT_REQ(true); /* Rejected by COP. */
1381 synqe = alloc_synqe(sc, lctx, M_NOWAIT);
1382 if (synqe == NULL) {
1384 REJECT_PASS_ACCEPT_REQ(true);
1386 atomic_store_int(&synqe->ok_to_respond, 0);
1389 * If all goes well t4_syncache_respond will get called during
1390 * syncache_add. Note that syncache_add releases the pcb lock.
1392 t4opt_to_tcpopt(&cpl->tcpopt, &to);
1393 toe_syncache_add(&inc, &to, &th, inp, tod, synqe);
1395 if (atomic_load_int(&synqe->ok_to_respond) > 0) {
1399 int rscale, mtu_idx, rx_credits;
1401 mtu_idx = find_best_mtu_idx(sc, &inc, &settings);
1402 rscale = cpl->tcpopt.wsf && V_tcp_do_rfc1323 ? select_rcv_wscale() : 0;
1403 wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND);
1404 wnd = min(wnd, MAX_RCV_WND);
1405 rx_credits = min(wnd >> 10, M_RCV_BUFSIZ);
1407 save_qids_in_synqe(synqe, vi, &settings);
1408 synqe->ulp_mode = select_ulp_mode(so, sc, &settings);
1410 opt0 = calc_opt0(so, vi, e, mtu_idx, rscale, rx_credits,
1411 synqe->ulp_mode, &settings);
1412 opt2 = calc_opt2p(sc, pi, synqe->rxqid, &cpl->tcpopt, &th,
1413 synqe->ulp_mode, CC_ALGO(intotcpcb(inp)), &settings);
1415 insert_tid(sc, tid, synqe, ntids);
1417 synqe->l2e_idx = e->idx;
1418 synqe->rcv_bufsize = rx_credits;
1422 if (send_synack(sc, synqe, opt0, opt2, tid) != 0) {
1423 remove_tid(sc, tid, ntids);
1426 REJECT_PASS_ACCEPT_REQ(true);
1430 "%s: stid %u, tid %u, lctx %p, synqe %p, mode %d, SYNACK",
1431 __func__, stid, tid, lctx, synqe, synqe->ulp_mode);
1433 REJECT_PASS_ACCEPT_REQ(false);
1439 CTR4(KTR_CXGBE, "%s: stid %u, tid %u, REJECT (%d)", __func__, stid, tid,
1444 release_tid(sc, tid, lctx->ctrlq);
1446 inp = synqe->lctx->inp;
1448 inp = release_synqe(sc, synqe);
1455 * The connection request hit a TOE listener but is being passed
1456 * on to the kernel sw stack instead of getting offloaded.
1458 m_adj(m, sizeof(*cpl));
1459 m->m_pkthdr.csum_flags |= (CSUM_IP_CHECKED | CSUM_IP_VALID |
1460 CSUM_DATA_VALID | CSUM_PSEUDO_HDR);
1461 m->m_pkthdr.csum_data = 0xffff;
1462 hw_ifp->if_input(hw_ifp, m);
1465 return (reject_reason);
1469 synqe_to_protohdrs(struct adapter *sc, struct synq_entry *synqe,
1470 const struct cpl_pass_establish *cpl, struct in_conninfo *inc,
1471 struct tcphdr *th, struct tcpopt *to)
1473 uint16_t tcp_opt = be16toh(cpl->tcp_opt);
1475 /* start off with the original SYN */
1476 pass_accept_req_to_protohdrs(sc, synqe->syn, inc, th);
1478 /* modify parts to make it look like the ACK to our SYN|ACK */
1479 th->th_flags = TH_ACK;
1480 th->th_ack = synqe->iss + 1;
1481 th->th_seq = be32toh(cpl->rcv_isn);
1482 bzero(to, sizeof(*to));
1483 if (G_TCPOPT_TSTAMP(tcp_opt)) {
1484 to->to_flags |= TOF_TS;
1485 to->to_tsecr = synqe->ts;
1490 do_pass_establish(struct sge_iq *iq, const struct rss_header *rss,
1493 struct adapter *sc = iq->adapter;
1496 const struct cpl_pass_establish *cpl = (const void *)(rss + 1);
1497 #if defined(KTR) || defined(INVARIANTS)
1498 unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid));
1500 unsigned int tid = GET_TID(cpl);
1501 struct synq_entry *synqe = lookup_tid(sc, tid);
1502 struct listen_ctx *lctx = synqe->lctx;
1503 struct inpcb *inp = lctx->inp, *new_inp;
1507 struct in_conninfo inc;
1508 struct toepcb *toep;
1509 struct epoch_tracker et;
1511 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1514 KASSERT(opcode == CPL_PASS_ESTABLISH,
1515 ("%s: unexpected opcode 0x%x", __func__, opcode));
1516 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1517 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__));
1518 KASSERT(synqe->flags & TPF_SYNQE,
1519 ("%s: tid %u (ctx %p) not a synqe", __func__, tid, synqe));
1521 CURVNET_SET(lctx->vnet);
1522 INP_INFO_RLOCK_ET(&V_tcbinfo, et); /* for syncache_expand */
1526 "%s: stid %u, tid %u, synqe %p (0x%x), inp_flags 0x%x",
1527 __func__, stid, tid, synqe, synqe->flags, inp->inp_flags);
1529 ifp = synqe->syn->m_pkthdr.rcvif;
1531 KASSERT(vi->pi->adapter == sc,
1532 ("%s: vi %p, sc %p mismatch", __func__, vi, sc));
1534 if (__predict_false(inp->inp_flags & INP_DROPPED)) {
1536 send_reset_synqe(TOEDEV(ifp), synqe);
1538 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1543 KASSERT(synqe->rxqid == iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0],
1544 ("%s: CPL arrived on unexpected rxq. %d %d", __func__,
1545 synqe->rxqid, (int)(iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0])));
1547 toep = alloc_toepcb(vi, synqe->txqid, synqe->rxqid, M_NOWAIT);
1551 toep->l2te = &sc->l2t->l2tab[synqe->l2e_idx];
1552 toep->vnet = lctx->vnet;
1553 set_ulp_mode(toep, synqe->ulp_mode);
1554 toep->opt0_rcv_bufsize = synqe->rcv_bufsize;
1556 MPASS(be32toh(cpl->snd_isn) - 1 == synqe->iss);
1557 MPASS(be32toh(cpl->rcv_isn) - 1 == synqe->irs);
1558 synqe->tcp_opt = cpl->tcp_opt;
1561 /* Come up with something that syncache_expand should be ok with. */
1562 synqe_to_protohdrs(sc, synqe, cpl, &inc, &th, &to);
1563 if (inc.inc_flags & INC_ISIPV6)
1564 toep->ce = t4_hold_lip(sc, &inc.inc6_laddr, lctx->ce);
1565 so = inp->inp_socket;
1566 KASSERT(so != NULL, ("%s: socket is NULL", __func__));
1568 if (!toe_syncache_expand(&inc, &to, &th, &so) || so == NULL) {
1573 /* New connection inpcb is already locked by syncache_expand(). */
1574 new_inp = sotoinpcb(so);
1575 INP_WLOCK_ASSERT(new_inp);
1576 MPASS(so->so_vnet == lctx->vnet);
1579 * This is for expansion from syncookies.
1581 * XXX: we've held the tcbinfo lock throughout so there's no risk of
1582 * anyone accept'ing a connection before we've installed our hooks, but
1583 * this somewhat defeats the purpose of having a tod_offload_socket :-(
1585 if (__predict_false(!(synqe->flags & TPF_SYNQE_EXPANDED))) {
1586 tcp_timer_activate(intotcpcb(new_inp), TT_KEEP, 0);
1587 t4_offload_socket(TOEDEV(ifp), synqe, so);
1590 INP_WUNLOCK(new_inp);
1592 /* Done with the synqe */
1593 inp = release_synqe(sc, synqe);
1596 INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1603 t4_init_listen_cpl_handlers(void)
1606 t4_register_cpl_handler(CPL_PASS_OPEN_RPL, do_pass_open_rpl);
1607 t4_register_cpl_handler(CPL_CLOSE_LISTSRV_RPL, do_close_server_rpl);
1608 t4_register_cpl_handler(CPL_PASS_ACCEPT_REQ, do_pass_accept_req);
1609 t4_register_cpl_handler(CPL_PASS_ESTABLISH, do_pass_establish);
1613 t4_uninit_listen_cpl_handlers(void)
1616 t4_register_cpl_handler(CPL_PASS_OPEN_RPL, NULL);
1617 t4_register_cpl_handler(CPL_CLOSE_LISTSRV_RPL, NULL);
1618 t4_register_cpl_handler(CPL_PASS_ACCEPT_REQ, NULL);
1619 t4_register_cpl_handler(CPL_PASS_ESTABLISH, NULL);