/* This is _not_ the normal way to "unoffload" a socket. */
void
undo_offload_socket(struct socket *so)
{
struct inpcb *inp = sotoinpcb(so);
struct tcpcb *tp = intotcpcb(inp);
struct toepcb *toep = tp->t_toe;
struct tom_data *td = toep->td;
struct sockbuf *sb;
INP_WLOCK_ASSERT(inp);
sb = &so->so_snd;
SOCKBUF_LOCK(sb);
sb->sb_flags &= ~SB_NOCOALESCE;
SOCKBUF_UNLOCK(sb);
sb = &so->so_rcv;
SOCKBUF_LOCK(sb);
sb->sb_flags &= ~SB_NOCOALESCE;
SOCKBUF_UNLOCK(sb);
tp->tod = NULL;
tp->t_toe = NULL;
tp->t_flags &= ~TF_TOE;
toep->inp = NULL;
toep->flags &= ~TPF_ATTACHED;
if (in_pcbrele_wlocked(inp))
panic("%s: inp freed.", __func__);
mtx_lock(&td->toep_list_lock);
TAILQ_REMOVE(&td->toep_list, toep, link);
mtx_unlock(&td->toep_list_lock);
free_toepcb(toep);
}
static void
release_offload_resources(struct toepcb *toep)
{
struct tom_data *td = toep->td;
struct adapter *sc = td_adapter(td);
int tid = toep->tid;
KASSERT(!(toep->flags & TPF_CPL_PENDING),
("%s: %p has CPL pending.", __func__, toep));
KASSERT(!(toep->flags & TPF_ATTACHED),
("%s: %p is still attached.", __func__, toep));
CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
__func__, toep, tid, toep->l2te, toep->ce);
if (toep->ulp_mode == ULP_MODE_TCPDDP)
release_ddp_resources(toep);
if (toep->l2te)
t4_l2t_release(toep->l2te);
if (tid >= 0) {
remove_tid(sc, tid);
release_tid(sc, tid, toep->ctrlq);
}
if (toep->ce)
release_lip(td, toep->ce);
mtx_lock(&td->toep_list_lock);
TAILQ_REMOVE(&td->toep_list, toep, link);
mtx_unlock(&td->toep_list_lock);
free_toepcb(toep);
}
static void
release_offload_resources(struct toepcb *toep)
{
struct tom_data *td = toep->td;
struct adapter *sc = td_adapter(td);
int tid = toep->tid;
KASSERT(toepcb_flag(toep, TPF_CPL_PENDING) == 0,
("%s: %p has CPL pending.", __func__, toep));
KASSERT(toepcb_flag(toep, TPF_ATTACHED) == 0,
("%s: %p is still attached.", __func__, toep));
CTR4(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p)",
__func__, toep, tid, toep->l2te);
if (toep->l2te)
t4_l2t_release(toep->l2te);
if (tid >= 0) {
remove_tid(sc, tid);
release_tid(sc, tid, toep->ctrlq);
}
mtx_lock(&td->toep_list_lock);
TAILQ_REMOVE(&td->toep_list, toep, link);
mtx_unlock(&td->toep_list_lock);
free_toepcb(toep);
}
static void
release_offload_resources(struct toepcb *toep)
{
struct tom_data *td = toep->td;
struct adapter *sc = td_adapter(td);
int tid = toep->tid;
KASSERT(!(toep->flags & TPF_CPL_PENDING),
("%s: %p has CPL pending.", __func__, toep));
KASSERT(!(toep->flags & TPF_ATTACHED),
("%s: %p is still attached.", __func__, toep));
CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
__func__, toep, tid, toep->l2te, toep->ce);
/*
* These queues should have been emptied at approximately the same time
* that a normal connection's socket's so_snd would have been purged or
* drained. Do _not_ clean up here.
*/
MPASS(mbufq_len(&toep->ulp_pduq) == 0);
MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0);
#ifdef INVARIANTS
ddp_assert_empty(toep);
#endif
if (toep->l2te)
t4_l2t_release(toep->l2te);
if (tid >= 0) {
remove_tid(sc, tid, toep->ce ? 2 : 1);
release_tid(sc, tid, toep->ctrlq);
}
if (toep->ce)
release_lip(td, toep->ce);
#ifdef RATELIMIT
if (toep->tc_idx != -1)
t4_release_cl_rl_kbps(sc, toep->vi->pi->port_id, toep->tc_idx);
#endif
mtx_lock(&td->toep_list_lock);
TAILQ_REMOVE(&td->toep_list, toep, link);
mtx_unlock(&td->toep_list_lock);
free_toepcb(toep);
}
/*
* 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);
}