void
send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
{
struct wrqe *wr;
struct cpl_abort_req *req;
int tid = toep->tid;
struct inpcb *inp = toep->inp;
struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */
INP_WLOCK_ASSERT(inp);
CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
__func__, toep->tid,
inp->inp_flags & INP_DROPPED ? "inp dropped" :
tcpstates[tp->t_state],
toep->flags, inp->inp_flags,
toep->flags & TPF_ABORT_SHUTDOWN ?
" (abort already in progress)" : "");
if (toep->flags & TPF_ABORT_SHUTDOWN)
return; /* abort already in progress */
toep->flags |= TPF_ABORT_SHUTDOWN;
KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
("%s: flowc_wr not sent for tid %d.", __func__, tid));
wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
if (wr == NULL) {
/* XXX */
panic("%s: allocation failure.", __func__);
}
req = wrtod(wr);
INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
if (inp->inp_flags & INP_DROPPED)
req->rsvd0 = htobe32(snd_nxt);
else
req->rsvd0 = htobe32(tp->snd_nxt);
req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
req->cmd = CPL_ABORT_SEND_RST;
/*
* XXX: What's the correct way to tell that the inp hasn't been detached
* from its socket? Should I even be flushing the snd buffer here?
*/
if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
struct socket *so = inp->inp_socket;
if (so != NULL) /* because I'm not sure. See comment above */
sbflush(&so->so_snd);
}
t4_l2t_send(sc, wr, toep->l2te);
}
/*
* Close a connection by sending a CPL_CLOSE_CON_REQ message.
*/
static int
close_conn(struct adapter *sc, struct toepcb *toep)
{
struct wrqe *wr;
struct cpl_close_con_req *req;
unsigned int tid = toep->tid;
CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
if (toep->flags & TPF_FIN_SENT)
return (0);
KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
("%s: flowc_wr not sent for tid %u.", __func__, tid));
wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
if (wr == NULL) {
/* XXX */
panic("%s: allocation failure.", __func__);
}
req = wrtod(wr);
req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
V_FW_WR_FLOWID(tid));
req->wr.wr_lo = cpu_to_be64(0);
OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
req->rsvd = 0;
toep->flags |= TPF_FIN_SENT;
toep->flags &= ~TPF_SEND_FIN;
t4_l2t_send(sc, wr, toep->l2te);
return (0);
}
/*
* active open (soconnect).
*
* State of affairs on entry:
* soisconnecting (so_state |= SS_ISCONNECTING)
* tcbinfo not locked (This has changed - used to be WLOCKed)
* inp WLOCKed
* tp->t_state = TCPS_SYN_SENT
* rtalloc1, RT_UNLOCK on rt.
*/
int
t4_connect(struct toedev *tod, struct socket *so, struct rtentry *rt,
struct sockaddr *nam)
{
struct adapter *sc = tod->tod_softc;
struct tom_data *td = tod_td(tod);
struct toepcb *toep = NULL;
struct wrqe *wr = NULL;
struct ifnet *rt_ifp = rt->rt_ifp;
struct port_info *pi;
int mtu_idx, rscale, qid_atid, rc, isipv6;
struct inpcb *inp = sotoinpcb(so);
struct tcpcb *tp = intotcpcb(inp);
int reason;
INP_WLOCK_ASSERT(inp);
KASSERT(nam->sa_family == AF_INET || nam->sa_family == AF_INET6,
("%s: dest addr %p has family %u", __func__, nam, nam->sa_family));
if (rt_ifp->if_type == IFT_ETHER)
pi = rt_ifp->if_softc;
else if (rt_ifp->if_type == IFT_L2VLAN) {
struct ifnet *ifp = VLAN_COOKIE(rt_ifp);
pi = ifp->if_softc;
} else if (rt_ifp->if_type == IFT_IEEE8023ADLAG)
DONT_OFFLOAD_ACTIVE_OPEN(ENOSYS); /* XXX: implement lagg+TOE */
else
DONT_OFFLOAD_ACTIVE_OPEN(ENOTSUP);
toep = alloc_toepcb(pi, -1, -1, M_NOWAIT);
if (toep == NULL)
DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
toep->tid = alloc_atid(sc, toep);
if (toep->tid < 0)
DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
toep->l2te = t4_l2t_get(pi, rt_ifp,
rt->rt_flags & RTF_GATEWAY ? rt->rt_gateway : nam);
if (toep->l2te == NULL)
DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
isipv6 = nam->sa_family == AF_INET6;
wr = alloc_wrqe(act_open_cpl_size(sc, isipv6), toep->ctrlq);
if (wr == NULL)
DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
if (sc->tt.ddp && (so->so_options & SO_NO_DDP) == 0)
set_tcpddp_ulp_mode(toep);
else
toep->ulp_mode = ULP_MODE_NONE;
SOCKBUF_LOCK(&so->so_rcv);
/* opt0 rcv_bufsiz initially, assumes its normal meaning later */
toep->rx_credits = min(select_rcv_wnd(so) >> 10, M_RCV_BUFSIZ);
SOCKBUF_UNLOCK(&so->so_rcv);
/*
* The kernel sets request_r_scale based on sb_max whereas we need to
* take hardware's MAX_RCV_WND into account too. This is normally a
* no-op as MAX_RCV_WND is much larger than the default sb_max.
*/
if (tp->t_flags & TF_REQ_SCALE)
rscale = tp->request_r_scale = select_rcv_wscale();
else
rscale = 0;
mtu_idx = find_best_mtu_idx(sc, &inp->inp_inc, 0);
qid_atid = (toep->ofld_rxq->iq.abs_id << 14) | toep->tid;
if (isipv6) {
struct cpl_act_open_req6 *cpl = wrtod(wr);
if ((inp->inp_vflag & INP_IPV6) == 0) {
/* XXX think about this a bit more */
log(LOG_ERR,
"%s: time to think about AF_INET6 + vflag 0x%x.\n",
__func__, inp->inp_vflag);
DONT_OFFLOAD_ACTIVE_OPEN(ENOTSUP);
}
toep->ce = hold_lip(td, &inp->in6p_laddr);
if (toep->ce == NULL)
DONT_OFFLOAD_ACTIVE_OPEN(ENOENT);
if (is_t4(sc)) {
INIT_TP_WR(cpl, 0);
cpl->params = select_ntuple(pi, toep->l2te);
} else {
struct cpl_t5_act_open_req6 *c5 = (void *)cpl;
INIT_TP_WR(c5, 0);
c5->iss = htobe32(tp->iss);
c5->params = select_ntuple(pi, toep->l2te);
}
OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ6,
qid_atid));
cpl->local_port = inp->inp_lport;
cpl->local_ip_hi = *(uint64_t *)&inp->in6p_laddr.s6_addr[0];
cpl->local_ip_lo = *(uint64_t *)&inp->in6p_laddr.s6_addr[8];
cpl->peer_port = inp->inp_fport;
cpl->peer_ip_hi = *(uint64_t *)&inp->in6p_faddr.s6_addr[0];
cpl->peer_ip_lo = *(uint64_t *)&inp->in6p_faddr.s6_addr[8];
cpl->opt0 = calc_opt0(so, pi, toep->l2te, mtu_idx, rscale,
toep->rx_credits, toep->ulp_mode);
cpl->opt2 = calc_opt2a(so, toep);
} else {
struct cpl_act_open_req *cpl = wrtod(wr);
if (is_t4(sc)) {
INIT_TP_WR(cpl, 0);
cpl->params = select_ntuple(pi, toep->l2te);
} else {
struct cpl_t5_act_open_req *c5 = (void *)cpl;
INIT_TP_WR(c5, 0);
c5->iss = htobe32(tp->iss);
c5->params = select_ntuple(pi, toep->l2te);
}
OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ,
qid_atid));
inp_4tuple_get(inp, &cpl->local_ip, &cpl->local_port,
&cpl->peer_ip, &cpl->peer_port);
cpl->opt0 = calc_opt0(so, pi, toep->l2te, mtu_idx, rscale,
toep->rx_credits, toep->ulp_mode);
cpl->opt2 = calc_opt2a(so, toep);
}
CTR5(KTR_CXGBE, "%s: atid %u (%s), toep %p, inp %p", __func__,
toep->tid, tcpstates[tp->t_state], toep, inp);
offload_socket(so, toep);
rc = t4_l2t_send(sc, wr, toep->l2te);
if (rc == 0) {
toep->flags |= TPF_CPL_PENDING;
return (0);
}
undo_offload_socket(so);
reason = __LINE__;
failed:
CTR3(KTR_CXGBE, "%s: not offloading (%d), rc %d", __func__, reason, rc);
if (wr)
free_wrqe(wr);
if (toep) {
if (toep->tid >= 0)
free_atid(sc, toep->tid);
if (toep->l2te)
t4_l2t_release(toep->l2te);
if (toep->ce)
release_lip(td, toep->ce);
free_toepcb(toep);
}
return (rc);
}
/*
* active open (soconnect).
*
* State of affairs on entry:
* soisconnecting (so_state |= SS_ISCONNECTING)
* tcbinfo not locked (This has changed - used to be WLOCKed)
* inp WLOCKed
* tp->t_state = TCPS_SYN_SENT
* rtalloc1, RT_UNLOCK on rt.
*/
int
t4_connect(struct toedev *tod, struct socket *so, struct rtentry *rt,
struct sockaddr *nam)
{
struct adapter *sc = tod->tod_softc;
struct toepcb *toep = NULL;
struct wrqe *wr = NULL;
struct ifnet *rt_ifp = rt->rt_ifp;
struct vi_info *vi;
int mtu_idx, rscale, qid_atid, rc, isipv6, txqid, rxqid;
struct inpcb *inp = sotoinpcb(so);
struct tcpcb *tp = intotcpcb(inp);
int reason;
struct offload_settings settings;
uint16_t vid = 0xfff, pcp = 0;
INP_WLOCK_ASSERT(inp);
KASSERT(nam->sa_family == AF_INET || nam->sa_family == AF_INET6,
("%s: dest addr %p has family %u", __func__, nam, nam->sa_family));
if (rt_ifp->if_type == IFT_ETHER)
vi = rt_ifp->if_softc;
else if (rt_ifp->if_type == IFT_L2VLAN) {
struct ifnet *ifp = VLAN_TRUNKDEV(rt_ifp);
vi = ifp->if_softc;
VLAN_TAG(rt_ifp, &vid);
VLAN_PCP(rt_ifp, &pcp);
} else if (rt_ifp->if_type == IFT_IEEE8023ADLAG)
DONT_OFFLOAD_ACTIVE_OPEN(ENOSYS); /* XXX: implement lagg+TOE */
else
DONT_OFFLOAD_ACTIVE_OPEN(ENOTSUP);
rw_rlock(&sc->policy_lock);
settings = *lookup_offload_policy(sc, OPEN_TYPE_ACTIVE, NULL,
EVL_MAKETAG(vid, pcp, 0), inp);
rw_runlock(&sc->policy_lock);
if (!settings.offload)
DONT_OFFLOAD_ACTIVE_OPEN(EPERM);
if (settings.txq >= 0 && settings.txq < vi->nofldtxq)
txqid = settings.txq;
else
txqid = arc4random() % vi->nofldtxq;
txqid += vi->first_ofld_txq;
if (settings.rxq >= 0 && settings.rxq < vi->nofldrxq)
rxqid = settings.rxq;
else
rxqid = arc4random() % vi->nofldrxq;
rxqid += vi->first_ofld_rxq;
toep = alloc_toepcb(vi, txqid, rxqid, M_NOWAIT | M_ZERO);
if (toep == NULL)
DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
toep->tid = alloc_atid(sc, toep);
if (toep->tid < 0)
DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
toep->l2te = t4_l2t_get(vi->pi, rt_ifp,
rt->rt_flags & RTF_GATEWAY ? rt->rt_gateway : nam);
if (toep->l2te == NULL)
DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
isipv6 = nam->sa_family == AF_INET6;
wr = alloc_wrqe(act_open_cpl_size(sc, isipv6), toep->ctrlq);
if (wr == NULL)
DONT_OFFLOAD_ACTIVE_OPEN(ENOMEM);
toep->vnet = so->so_vnet;
set_ulp_mode(toep, select_ulp_mode(so, sc, &settings));
SOCKBUF_LOCK(&so->so_rcv);
/* opt0 rcv_bufsiz initially, assumes its normal meaning later */
toep->rx_credits = min(select_rcv_wnd(so) >> 10, M_RCV_BUFSIZ);
SOCKBUF_UNLOCK(&so->so_rcv);
/*
* The kernel sets request_r_scale based on sb_max whereas we need to
* take hardware's MAX_RCV_WND into account too. This is normally a
* no-op as MAX_RCV_WND is much larger than the default sb_max.
*/
if (tp->t_flags & TF_REQ_SCALE)
rscale = tp->request_r_scale = select_rcv_wscale();
else
rscale = 0;
mtu_idx = find_best_mtu_idx(sc, &inp->inp_inc, &settings);
qid_atid = V_TID_QID(toep->ofld_rxq->iq.abs_id) | V_TID_TID(toep->tid) |
V_TID_COOKIE(CPL_COOKIE_TOM);
if (isipv6) {
struct cpl_act_open_req6 *cpl = wrtod(wr);
struct cpl_t5_act_open_req6 *cpl5 = (void *)cpl;
struct cpl_t6_act_open_req6 *cpl6 = (void *)cpl;
if ((inp->inp_vflag & INP_IPV6) == 0)
DONT_OFFLOAD_ACTIVE_OPEN(ENOTSUP);
toep->ce = t4_hold_lip(sc, &inp->in6p_laddr, NULL);
if (toep->ce == NULL)
DONT_OFFLOAD_ACTIVE_OPEN(ENOENT);
switch (chip_id(sc)) {
case CHELSIO_T4:
INIT_TP_WR(cpl, 0);
cpl->params = select_ntuple(vi, toep->l2te);
break;
case CHELSIO_T5:
INIT_TP_WR(cpl5, 0);
cpl5->iss = htobe32(tp->iss);
cpl5->params = select_ntuple(vi, toep->l2te);
break;
case CHELSIO_T6:
default:
INIT_TP_WR(cpl6, 0);
cpl6->iss = htobe32(tp->iss);
cpl6->params = select_ntuple(vi, toep->l2te);
break;
}
OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ6,
qid_atid));
cpl->local_port = inp->inp_lport;
cpl->local_ip_hi = *(uint64_t *)&inp->in6p_laddr.s6_addr[0];
cpl->local_ip_lo = *(uint64_t *)&inp->in6p_laddr.s6_addr[8];
cpl->peer_port = inp->inp_fport;
cpl->peer_ip_hi = *(uint64_t *)&inp->in6p_faddr.s6_addr[0];
cpl->peer_ip_lo = *(uint64_t *)&inp->in6p_faddr.s6_addr[8];
cpl->opt0 = calc_opt0(so, vi, toep->l2te, mtu_idx, rscale,
toep->rx_credits, toep->ulp_mode, &settings);
cpl->opt2 = calc_opt2a(so, toep, &settings);
} else {
struct cpl_act_open_req *cpl = wrtod(wr);
struct cpl_t5_act_open_req *cpl5 = (void *)cpl;
struct cpl_t6_act_open_req *cpl6 = (void *)cpl;
switch (chip_id(sc)) {
case CHELSIO_T4:
INIT_TP_WR(cpl, 0);
cpl->params = select_ntuple(vi, toep->l2te);
break;
case CHELSIO_T5:
INIT_TP_WR(cpl5, 0);
cpl5->iss = htobe32(tp->iss);
cpl5->params = select_ntuple(vi, toep->l2te);
break;
case CHELSIO_T6:
default:
INIT_TP_WR(cpl6, 0);
cpl6->iss = htobe32(tp->iss);
cpl6->params = select_ntuple(vi, toep->l2te);
break;
}
OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ,
qid_atid));
inp_4tuple_get(inp, &cpl->local_ip, &cpl->local_port,
&cpl->peer_ip, &cpl->peer_port);
cpl->opt0 = calc_opt0(so, vi, toep->l2te, mtu_idx, rscale,
toep->rx_credits, toep->ulp_mode, &settings);
cpl->opt2 = calc_opt2a(so, toep, &settings);
}
CTR5(KTR_CXGBE, "%s: atid %u (%s), toep %p, inp %p", __func__,
toep->tid, tcpstates[tp->t_state], toep, inp);
offload_socket(so, toep);
rc = t4_l2t_send(sc, wr, toep->l2te);
if (rc == 0) {
toep->flags |= TPF_CPL_PENDING;
return (0);
}
undo_offload_socket(so);
reason = __LINE__;
failed:
CTR3(KTR_CXGBE, "%s: not offloading (%d), rc %d", __func__, reason, rc);
if (wr)
free_wrqe(wr);
if (toep) {
if (toep->tid >= 0)
free_atid(sc, toep->tid);
if (toep->l2te)
t4_l2t_release(toep->l2te);
if (toep->ce)
t4_release_lip(sc, toep->ce);
free_toepcb(toep);
}
return (rc);
}
/*
* Send data and/or a FIN to the peer.
*
* The socket's so_snd buffer consists of a stream of data starting with sb_mb
* and linked together with m_next. sb_sndptr, if set, is the last mbuf that
* was transmitted.
*/
static void
t4_push_frames(struct adapter *sc, struct toepcb *toep)
{
struct mbuf *sndptr, *m, *sb_sndptr;
struct fw_ofld_tx_data_wr *txwr;
struct wrqe *wr;
unsigned int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
struct inpcb *inp = toep->inp;
struct tcpcb *tp = intotcpcb(inp);
struct socket *so = inp->inp_socket;
struct sockbuf *sb = &so->so_snd;
int tx_credits;
struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
INP_WLOCK_ASSERT(inp);
KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
if (__predict_false(toep->ulp_mode != ULP_MODE_NONE &&
toep->ulp_mode != ULP_MODE_TCPDDP))
CXGBE_UNIMPLEMENTED("ulp_mode");
/*
* This function doesn't resume by itself. Someone else must clear the
* flag and call this function.
*/
if (__predict_false(toep->flags & TPF_TX_SUSPENDED))
return;
do {
tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
max_imm = max_imm_payload(tx_credits);
max_nsegs = max_dsgl_nsegs(tx_credits);
SOCKBUF_LOCK(sb);
sb_sndptr = sb->sb_sndptr;
sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
plen = 0;
nsegs = 0;
max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
for (m = sndptr; m != NULL; m = m->m_next) {
int n = sglist_count(mtod(m, void *), m->m_len);
nsegs += n;
plen += m->m_len;
/* This mbuf sent us _over_ the nsegs limit, back out */
if (plen > max_imm && nsegs > max_nsegs) {
nsegs -= n;
plen -= m->m_len;
if (plen == 0) {
/* Too few credits */
toep->flags |= TPF_TX_SUSPENDED;
SOCKBUF_UNLOCK(sb);
return;
}
break;
}
if (max_nsegs_1mbuf < n)
max_nsegs_1mbuf = n;
sb_sndptr = m; /* new sb->sb_sndptr if all goes well */
/* This mbuf put us right at the max_nsegs limit */
if (plen > max_imm && nsegs == max_nsegs) {
m = m->m_next;
break;
}
}
if (sb->sb_flags & SB_AUTOSIZE &&
V_tcp_do_autosndbuf &&
sb->sb_hiwat < V_tcp_autosndbuf_max &&
sbspace(sb) < sb->sb_hiwat / 8 * 7) {
int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
V_tcp_autosndbuf_max);
if (!sbreserve_locked(sb, newsize, so, NULL))
sb->sb_flags &= ~SB_AUTOSIZE;
else {
sowwakeup_locked(so); /* room available */
SOCKBUF_UNLOCK_ASSERT(sb);
goto unlocked;
}
}
SOCKBUF_UNLOCK(sb);
unlocked:
/* nothing to send */
if (plen == 0) {
KASSERT(m == NULL,
("%s: nothing to send, but m != NULL", __func__));
break;
}
if (__predict_false(toep->flags & TPF_FIN_SENT))
panic("%s: excess tx.", __func__);
if (plen <= max_imm) {
/* Immediate data tx */
wr = alloc_wrqe(roundup(sizeof(*txwr) + plen, 16),
toep->ofld_txq);
if (wr == NULL) {
/* XXX: how will we recover from this? */
toep->flags |= TPF_TX_SUSPENDED;
return;
}
txwr = wrtod(wr);
credits = howmany(wr->wr_len, 16);
write_tx_wr(txwr, toep, plen, plen, credits,
tp->t_flags & TF_MORETOCOME);
m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
} else {
int wr_len;
/* DSGL tx */
wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
wr = alloc_wrqe(roundup(wr_len, 16), toep->ofld_txq);
if (wr == NULL) {
/* XXX: how will we recover from this? */
toep->flags |= TPF_TX_SUSPENDED;
return;
}
txwr = wrtod(wr);
credits = howmany(wr_len, 16);
write_tx_wr(txwr, toep, 0, plen, credits,
tp->t_flags & TF_MORETOCOME);
write_tx_sgl(txwr + 1, sndptr, m, nsegs,
max_nsegs_1mbuf);
if (wr_len & 0xf) {
uint64_t *pad = (uint64_t *)
((uintptr_t)txwr + wr_len);
*pad = 0;
}
}
KASSERT(toep->tx_credits >= credits,
("%s: not enough credits", __func__));
toep->tx_credits -= credits;
tp->snd_nxt += plen;
tp->snd_max += plen;
SOCKBUF_LOCK(sb);
KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
sb->sb_sndptr = sb_sndptr;
SOCKBUF_UNLOCK(sb);
toep->flags |= TPF_TX_DATA_SENT;
KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
txsd->plen = plen;
txsd->tx_credits = credits;
txsd++;
if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
toep->txsd_pidx = 0;
txsd = &toep->txsd[0];
}
toep->txsd_avail--;
t4_l2t_send(sc, wr, toep->l2te);
} while (m != NULL);
/* Send a FIN if requested, but only if there's no more data to send */
if (m == NULL && toep->flags & TPF_SEND_FIN)
close_conn(sc, toep);
}
void
t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
{
struct mbuf *sndptr, *m;
struct fw_ofld_tx_data_wr *txwr;
struct wrqe *wr;
u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
u_int adjusted_plen, ulp_submode;
struct inpcb *inp = toep->inp;
struct tcpcb *tp = intotcpcb(inp);
int tx_credits, shove;
struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
struct mbufq *pduq = &toep->ulp_pduq;
static const u_int ulp_extra_len[] = {0, 4, 4, 8};
INP_WLOCK_ASSERT(inp);
KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
KASSERT(toep->ulp_mode == ULP_MODE_ISCSI,
("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
return;
/*
* This function doesn't resume by itself. Someone else must clear the
* flag and call this function.
*/
if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
KASSERT(drop == 0,
("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
return;
}
if (drop)
rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
while ((sndptr = mbufq_first(pduq)) != NULL) {
M_ASSERTPKTHDR(sndptr);
tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
max_imm = max_imm_payload(tx_credits);
max_nsegs = max_dsgl_nsegs(tx_credits);
plen = 0;
nsegs = 0;
max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
for (m = sndptr; m != NULL; m = m->m_next) {
int n = sglist_count(mtod(m, void *), m->m_len);
nsegs += n;
plen += m->m_len;
/*
* This mbuf would send us _over_ the nsegs limit.
* Suspend tx because the PDU can't be sent out.
*/
if (plen > max_imm && nsegs > max_nsegs) {
toep->flags |= TPF_TX_SUSPENDED;
return;
}
if (max_nsegs_1mbuf < n)
max_nsegs_1mbuf = n;
}
if (__predict_false(toep->flags & TPF_FIN_SENT))
panic("%s: excess tx.", __func__);
/*
* We have a PDU to send. All of it goes out in one WR so 'm'
* is NULL. A PDU's length is always a multiple of 4.
*/
MPASS(m == NULL);
MPASS((plen & 3) == 0);
MPASS(sndptr->m_pkthdr.len == plen);
shove = !(tp->t_flags & TF_MORETOCOME);
ulp_submode = mbuf_ulp_submode(sndptr);
MPASS(ulp_submode < nitems(ulp_extra_len));
/*
* plen doesn't include header and data digests, which are
* generated and inserted in the right places by the TOE, but
* they do occupy TCP sequence space and need to be accounted
* for.
*/
adjusted_plen = plen + ulp_extra_len[ulp_submode];
if (plen <= max_imm) {
/* Immediate data tx */
wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
toep->ofld_txq);
if (wr == NULL) {
/* XXX: how will we recover from this? */
toep->flags |= TPF_TX_SUSPENDED;
return;
}
txwr = wrtod(wr);
credits = howmany(wr->wr_len, 16);
write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
shove, ulp_submode, sc->tt.tx_align);
m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
nsegs = 0;
} else {
int wr_len;
/* DSGL tx */
wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
if (wr == NULL) {
/* XXX: how will we recover from this? */
toep->flags |= TPF_TX_SUSPENDED;
return;
}
txwr = wrtod(wr);
credits = howmany(wr_len, 16);
write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
shove, ulp_submode, sc->tt.tx_align);
write_tx_sgl(txwr + 1, sndptr, m, nsegs,
max_nsegs_1mbuf);
if (wr_len & 0xf) {
uint64_t *pad = (uint64_t *)
((uintptr_t)txwr + wr_len);
*pad = 0;
}
}
KASSERT(toep->tx_credits >= credits,
("%s: not enough credits", __func__));
m = mbufq_dequeue(pduq);
MPASS(m == sndptr);
mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
toep->tx_credits -= credits;
toep->tx_nocompl += credits;
toep->plen_nocompl += plen;
if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
toep->tx_nocompl >= toep->tx_total / 4) {
txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
toep->tx_nocompl = 0;
toep->plen_nocompl = 0;
}
tp->snd_nxt += adjusted_plen;
tp->snd_max += adjusted_plen;
toep->flags |= TPF_TX_DATA_SENT;
if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
toep->flags |= TPF_TX_SUSPENDED;
KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
txsd->plen = plen;
txsd->tx_credits = credits;
txsd++;
if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
toep->txsd_pidx = 0;
txsd = &toep->txsd[0];
}
toep->txsd_avail--;
t4_l2t_send(sc, wr, toep->l2te);
}
/* Send a FIN if requested, but only if there are no more PDUs to send */
if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
close_conn(sc, toep);
}