void
handle_ddp_close(struct toepcb *toep, struct tcpcb *tp, struct sockbuf *sb,
__be32 rcv_nxt)
{
struct mbuf *m;
int len;
SOCKBUF_LOCK_ASSERT(sb);
INP_WLOCK_ASSERT(toep->inp);
len = be32toh(rcv_nxt) - tp->rcv_nxt;
/* Signal handle_ddp() to break out of its sleep loop. */
toep->ddp_flags &= ~(DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE);
if (len == 0)
return;
tp->rcv_nxt += len;
KASSERT(toep->sb_cc >= sbused(sb),
("%s: sb %p has more data (%d) than last time (%d).",
__func__, sb, sbused(sb), toep->sb_cc));
toep->rx_credits += toep->sb_cc - sbused(sb);
#ifdef USE_DDP_RX_FLOW_CONTROL
toep->rx_credits -= len; /* adjust for F_RX_FC_DDP */
#endif
m = get_ddp_mbuf(len);
sbappendstream_locked(sb, m, 0);
toep->sb_cc = sbused(sb);
}
/* XXX: handle_ddp_data code duplication */
void
insert_ddp_data(struct toepcb *toep, uint32_t n)
{
struct inpcb *inp = toep->inp;
struct tcpcb *tp = intotcpcb(inp);
struct sockbuf *sb = &inp->inp_socket->so_rcv;
struct mbuf *m;
INP_WLOCK_ASSERT(inp);
SOCKBUF_LOCK_ASSERT(sb);
m = get_ddp_mbuf(n);
tp->rcv_nxt += n;
#ifndef USE_DDP_RX_FLOW_CONTROL
KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__));
tp->rcv_wnd -= n;
#endif
KASSERT(toep->sb_cc >= sbused(sb),
("%s: sb %p has more data (%d) than last time (%d).",
__func__, sb, sbused(sb), toep->sb_cc));
toep->rx_credits += toep->sb_cc - sbused(sb);
#ifdef USE_DDP_RX_FLOW_CONTROL
toep->rx_credits -= n; /* adjust for F_RX_FC_DDP */
#endif
sbappendstream_locked(sb, m, 0);
toep->sb_cc = sbused(sb);
}
void
t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp)
{
struct adapter *sc = tod->tod_softc;
struct inpcb *inp = tp->t_inpcb;
struct socket *so = inp->inp_socket;
struct sockbuf *sb = &so->so_rcv;
struct toepcb *toep = tp->t_toe;
int credits;
INP_WLOCK_ASSERT(inp);
SOCKBUF_LOCK_ASSERT(sb);
KASSERT(toep->sb_cc >= sbused(sb),
("%s: sb %p has more data (%d) than last time (%d).",
__func__, sb, sbused(sb), toep->sb_cc));
toep->rx_credits += toep->sb_cc - sbused(sb);
toep->sb_cc = sbused(sb);
if (toep->rx_credits > 0 &&
(tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
(toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
credits = send_rx_credits(sc, toep, toep->rx_credits);
toep->rx_credits -= credits;
tp->rcv_wnd += credits;
tp->rcv_adv += credits;
}
}
static inline int
sdp_post_recvs_needed(struct sdp_sock *ssk)
{
unsigned long bytes_in_process;
unsigned long max_bytes;
int buffer_size;
int posted;
if (!ssk->qp_active || !ssk->socket)
return 0;
posted = rx_ring_posted(ssk);
if (posted >= SDP_RX_SIZE)
return 0;
if (posted < SDP_MIN_TX_CREDITS)
return 1;
buffer_size = ssk->recv_bytes;
max_bytes = max(ssk->socket->so_snd.sb_hiwat,
(1 + SDP_MIN_TX_CREDITS) * buffer_size);
max_bytes *= rcvbuf_scale;
/*
* Compute bytes in the receive queue and socket buffer.
*/
bytes_in_process = (posted - SDP_MIN_TX_CREDITS) * buffer_size;
bytes_in_process += sbused(&ssk->socket->so_rcv);
return bytes_in_process < max_bytes;
}
static bool_t
svc_vc_ack(SVCXPRT *xprt, uint32_t *ack)
{
*ack = atomic_load_acq_32(&xprt->xp_snt_cnt);
*ack -= sbused(&xprt->xp_socket->so_snd);
return (TRUE);
}
static int
do_rx_iscsi_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
{
struct adapter *sc = iq->adapter;
struct cxgbei_data *ci = sc->iscsi_ulp_softc;
const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1);
u_int tid = GET_TID(cpl);
struct toepcb *toep = lookup_tid(sc, tid);
struct inpcb *inp = toep->inp;
struct socket *so;
struct sockbuf *sb;
struct tcpcb *tp;
struct icl_cxgbei_conn *icc;
struct icl_conn *ic;
struct icl_cxgbei_pdu *icp = toep->ulpcb2;
struct icl_pdu *ip;
u_int pdu_len, val;
MPASS(m == NULL);
/* Must already be assembling a PDU. */
MPASS(icp != NULL);
MPASS(icp->icp_flags & ICPF_RX_HDR); /* Data is optional. */
MPASS((icp->icp_flags & ICPF_RX_STATUS) == 0);
pdu_len = be16toh(cpl->len); /* includes everything. */
val = be32toh(cpl->ddpvld);
#if 0
CTR5(KTR_CXGBE,
"%s: tid %u, cpl->len %u, ddpvld 0x%08x, icp_flags 0x%08x",
__func__, tid, pdu_len, val, icp->icp_flags);
#endif
icp->icp_flags |= ICPF_RX_STATUS;
ip = &icp->ip;
if (val & F_DDP_PADDING_ERR)
icp->icp_flags |= ICPF_PAD_ERR;
if (val & F_DDP_HDRCRC_ERR)
icp->icp_flags |= ICPF_HCRC_ERR;
if (val & F_DDP_DATACRC_ERR)
icp->icp_flags |= ICPF_DCRC_ERR;
if (val & F_DDP_PDU && ip->ip_data_mbuf == NULL) {
MPASS((icp->icp_flags & ICPF_RX_FLBUF) == 0);
MPASS(ip->ip_data_len > 0);
icp->icp_flags |= ICPF_RX_DDP;
counter_u64_add(ci->ddp_pdus, 1);
counter_u64_add(ci->ddp_bytes, ip->ip_data_len);
}
INP_WLOCK(inp);
if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
__func__, tid, pdu_len, inp->inp_flags);
INP_WUNLOCK(inp);
icl_cxgbei_conn_pdu_free(NULL, ip);
#ifdef INVARIANTS
toep->ulpcb2 = NULL;
#endif
return (0);
}
tp = intotcpcb(inp);
MPASS(icp->icp_seq == tp->rcv_nxt);
MPASS(tp->rcv_wnd >= pdu_len);
tp->rcv_nxt += pdu_len;
tp->rcv_wnd -= pdu_len;
tp->t_rcvtime = ticks;
/* update rx credits */
toep->rx_credits += pdu_len;
t4_rcvd(&toep->td->tod, tp); /* XXX: sc->tom_softc.tod */
so = inp->inp_socket;
sb = &so->so_rcv;
SOCKBUF_LOCK(sb);
icc = toep->ulpcb;
if (__predict_false(icc == NULL || sb->sb_state & SBS_CANTRCVMORE)) {
CTR5(KTR_CXGBE,
"%s: tid %u, excess rx (%d bytes), icc %p, sb_state 0x%x",
__func__, tid, pdu_len, icc, sb->sb_state);
SOCKBUF_UNLOCK(sb);
INP_WUNLOCK(inp);
INP_INFO_RLOCK(&V_tcbinfo);
INP_WLOCK(inp);
tp = tcp_drop(tp, ECONNRESET);
if (tp)
INP_WUNLOCK(inp);
INP_INFO_RUNLOCK(&V_tcbinfo);
icl_cxgbei_conn_pdu_free(NULL, ip);
#ifdef INVARIANTS
toep->ulpcb2 = NULL;
#endif
return (0);
}
MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
ic = &icc->ic;
icl_cxgbei_new_pdu_set_conn(ip, ic);
MPASS(m == NULL); /* was unused, we'll use it now. */
m = sbcut_locked(sb, sbused(sb)); /* XXXNP: toep->sb_cc accounting? */
if (__predict_false(m != NULL)) {
int len = m_length(m, NULL);
/*
* PDUs were received before the tid transitioned to ULP mode.
* Convert them to icl_cxgbei_pdus and send them to ICL before
* the PDU in icp/ip.
*/
CTR3(KTR_CXGBE, "%s: tid %u, %u bytes in so_rcv", __func__, tid,
len);
/* XXXNP: needs to be rewritten. */
if (len == sizeof(struct iscsi_bhs) || len == 4 + sizeof(struct
iscsi_bhs)) {
struct icl_cxgbei_pdu *icp0;
struct icl_pdu *ip0;
ip0 = icl_cxgbei_new_pdu(M_NOWAIT);
icl_cxgbei_new_pdu_set_conn(ip0, ic);
if (ip0 == NULL)
CXGBE_UNIMPLEMENTED("PDU allocation failure");
icp0 = ip_to_icp(ip0);
icp0->icp_seq = 0; /* XXX */
icp0->icp_flags = ICPF_RX_HDR | ICPF_RX_STATUS;
m_copydata(m, 0, sizeof(struct iscsi_bhs), (void *)ip0->ip_bhs);
STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip0, ip_next);
}
m_freem(m);
}
STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip, ip_next);
if ((icc->rx_flags & RXF_ACTIVE) == 0) {
struct cxgbei_worker_thread_softc *cwt = &cwt_softc[icc->cwt];
mtx_lock(&cwt->cwt_lock);
icc->rx_flags |= RXF_ACTIVE;
TAILQ_INSERT_TAIL(&cwt->rx_head, icc, rx_link);
if (cwt->cwt_state == CWT_SLEEPING) {
cwt->cwt_state = CWT_RUNNING;
cv_signal(&cwt->cwt_cv);
}
mtx_unlock(&cwt->cwt_lock);
}
SOCKBUF_UNLOCK(sb);
INP_WUNLOCK(inp);
#ifdef INVARIANTS
toep->ulpcb2 = NULL;
#endif
return (0);
}
static int
soo_ioctl(struct file *fp, u_long cmd, void *data, struct ucred *active_cred,
struct thread *td)
{
struct socket *so = fp->f_data;
int error = 0;
switch (cmd) {
case FIONBIO:
SOCK_LOCK(so);
if (*(int *)data)
so->so_state |= SS_NBIO;
else
so->so_state &= ~SS_NBIO;
SOCK_UNLOCK(so);
break;
case FIOASYNC:
/*
* XXXRW: This code separately acquires SOCK_LOCK(so) and
* SOCKBUF_LOCK(&so->so_rcv) even though they are the same
* mutex to avoid introducing the assumption that they are
* the same.
*/
if (*(int *)data) {
SOCK_LOCK(so);
so->so_state |= SS_ASYNC;
SOCK_UNLOCK(so);
SOCKBUF_LOCK(&so->so_rcv);
so->so_rcv.sb_flags |= SB_ASYNC;
SOCKBUF_UNLOCK(&so->so_rcv);
SOCKBUF_LOCK(&so->so_snd);
so->so_snd.sb_flags |= SB_ASYNC;
SOCKBUF_UNLOCK(&so->so_snd);
} else {
SOCK_LOCK(so);
so->so_state &= ~SS_ASYNC;
SOCK_UNLOCK(so);
SOCKBUF_LOCK(&so->so_rcv);
so->so_rcv.sb_flags &= ~SB_ASYNC;
SOCKBUF_UNLOCK(&so->so_rcv);
SOCKBUF_LOCK(&so->so_snd);
so->so_snd.sb_flags &= ~SB_ASYNC;
SOCKBUF_UNLOCK(&so->so_snd);
}
break;
case FIONREAD:
/* Unlocked read. */
*(int *)data = sbavail(&so->so_rcv);
break;
case FIONWRITE:
/* Unlocked read. */
*(int *)data = sbavail(&so->so_snd);
break;
case FIONSPACE:
/* Unlocked read. */
if ((so->so_snd.sb_hiwat < sbused(&so->so_snd)) ||
(so->so_snd.sb_mbmax < so->so_snd.sb_mbcnt))
*(int *)data = 0;
else
*(int *)data = sbspace(&so->so_snd);
break;
case FIOSETOWN:
error = fsetown(*(int *)data, &so->so_sigio);
break;
case FIOGETOWN:
*(int *)data = fgetown(&so->so_sigio);
break;
case SIOCSPGRP:
error = fsetown(-(*(int *)data), &so->so_sigio);
break;
case SIOCGPGRP:
*(int *)data = -fgetown(&so->so_sigio);
break;
case SIOCATMARK:
/* Unlocked read. */
*(int *)data = (so->so_rcv.sb_state & SBS_RCVATMARK) != 0;
break;
default:
/*
* Interface/routing/protocol specific ioctls: interface and
* routing ioctls should have a different entry since a
* socket is unnecessary.
*/
if (IOCGROUP(cmd) == 'i')
error = ifioctl(so, cmd, data, td);
else if (IOCGROUP(cmd) == 'r') {
CURVNET_SET(so->so_vnet);
error = rtioctl_fib(cmd, data, so->so_fibnum);
CURVNET_RESTORE();
} else {
CURVNET_SET(so->so_vnet);
error = ((*so->so_proto->pr_usrreqs->pru_control)
(so, cmd, data, 0, td));
CURVNET_RESTORE();
}
break;
}
return (error);
}
static int
handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len)
{
uint32_t report = be32toh(ddp_report);
unsigned int db_flag;
struct inpcb *inp = toep->inp;
struct tcpcb *tp;
struct socket *so;
struct sockbuf *sb;
struct mbuf *m;
db_flag = report & F_DDP_BUF_IDX ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE;
if (__predict_false(!(report & F_DDP_INV)))
CXGBE_UNIMPLEMENTED("DDP buffer still valid");
INP_WLOCK(inp);
so = inp_inpcbtosocket(inp);
sb = &so->so_rcv;
if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) {
/*
* XXX: think a bit more.
* tcpcb probably gone, but socket should still be around
* because we always wait for DDP completion in soreceive no
* matter what. Just wake it up and let it clean up.
*/
CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x",
__func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags);
SOCKBUF_LOCK(sb);
goto wakeup;
}
tp = intotcpcb(inp);
len += be32toh(rcv_nxt) - tp->rcv_nxt;
tp->rcv_nxt += len;
tp->t_rcvtime = ticks;
#ifndef USE_DDP_RX_FLOW_CONTROL
KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
tp->rcv_wnd -= len;
#endif
m = get_ddp_mbuf(len);
SOCKBUF_LOCK(sb);
if (report & F_DDP_BUF_COMPLETE)
toep->ddp_score = DDP_HIGH_SCORE;
else
discourage_ddp(toep);
/* receive buffer autosize */
if (sb->sb_flags & SB_AUTOSIZE &&
V_tcp_do_autorcvbuf &&
sb->sb_hiwat < V_tcp_autorcvbuf_max &&
len > (sbspace(sb) / 8 * 7)) {
unsigned int hiwat = sb->sb_hiwat;
unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
V_tcp_autorcvbuf_max);
if (!sbreserve_locked(sb, newsize, so, NULL))
sb->sb_flags &= ~SB_AUTOSIZE;
else
toep->rx_credits += newsize - hiwat;
}
KASSERT(toep->sb_cc >= sbused(sb),
("%s: sb %p has more data (%d) than last time (%d).",
__func__, sb, sbused(sb), toep->sb_cc));
toep->rx_credits += toep->sb_cc - sbused(sb);
#ifdef USE_DDP_RX_FLOW_CONTROL
toep->rx_credits -= len; /* adjust for F_RX_FC_DDP */
#endif
sbappendstream_locked(sb, m, 0);
toep->sb_cc = sbused(sb);
wakeup:
KASSERT(toep->ddp_flags & db_flag,
("%s: DDP buffer not active. toep %p, ddp_flags 0x%x, report 0x%x",
__func__, toep, toep->ddp_flags, report));
toep->ddp_flags &= ~db_flag;
sorwakeup_locked(so);
SOCKBUF_UNLOCK_ASSERT(sb);
INP_WUNLOCK(inp);
return (0);
}
/*
* 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.
*
* drop indicates the number of bytes that should be dropped from the head of
* the send buffer. It is an optimization that lets do_fw4_ack avoid creating
* contention on the send buffer lock (before this change it used to do
* sowwakeup and then t4_push_frames right after that when recovering from tx
* stalls). When drop is set this function MUST drop the bytes and wake up any
* writers.
*/
void
t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
{
struct mbuf *sndptr, *m, *sb_sndptr;
struct fw_ofld_tx_data_wr *txwr;
struct wrqe *wr;
u_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, shove, compl, sowwakeup;
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));
KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
toep->ulp_mode == ULP_MODE_TCPDDP ||
toep->ulp_mode == ULP_MODE_RDMA,
("%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;
}
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);
sowwakeup = drop;
if (drop) {
sbdrop_locked(sb, drop);
drop = 0;
}
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;
if (sowwakeup)
sowwakeup_locked(so);
else
SOCKBUF_UNLOCK(sb);
SOCKBUF_UNLOCK_ASSERT(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 (sbused(sb) > sb->sb_hiwat * 5 / 8 &&
toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
compl = 1;
else
compl = 0;
if (sb->sb_flags & SB_AUTOSIZE &&
V_tcp_do_autosndbuf &&
sb->sb_hiwat < V_tcp_autosndbuf_max &&
sbused(sb) >= sb->sb_hiwat * 7 / 8) {
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 = 1; /* room available */
}
if (sowwakeup)
sowwakeup_locked(so);
else
SOCKBUF_UNLOCK(sb);
SOCKBUF_UNLOCK_ASSERT(sb);
/* 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__);
shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
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, plen, credits, shove, 0,
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, plen, credits, shove, 0,
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__));
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)
compl = 1;
if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
toep->tx_nocompl = 0;
toep->plen_nocompl = 0;
}
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;
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);
} 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);
}
static int
do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
{
struct adapter *sc = iq->adapter;
const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
struct toepcb *toep = lookup_tid(sc, tid);
struct inpcb *inp;
struct tcpcb *tp;
struct socket *so;
uint8_t credits = cpl->credits;
struct ofld_tx_sdesc *txsd;
int plen;
#ifdef INVARIANTS
unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
#endif
/*
* Very unusual case: we'd sent a flowc + abort_req for a synq entry and
* now this comes back carrying the credits for the flowc.
*/
if (__predict_false(toep->flags & TPF_SYNQE)) {
KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
("%s: credits for a synq entry %p", __func__, toep));
return (0);
}
inp = toep->inp;
KASSERT(opcode == CPL_FW4_ACK,
("%s: unexpected opcode 0x%x", __func__, opcode));
KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
INP_WLOCK(inp);
if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
INP_WUNLOCK(inp);
return (0);
}
KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
("%s: inp_flags 0x%x", __func__, inp->inp_flags));
tp = intotcpcb(inp);
if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
tcp_seq snd_una = be32toh(cpl->snd_una);
#ifdef INVARIANTS
if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
log(LOG_ERR,
"%s: unexpected seq# %x for TID %u, snd_una %x\n",
__func__, snd_una, toep->tid, tp->snd_una);
}
#endif
if (tp->snd_una != snd_una) {
tp->snd_una = snd_una;
tp->ts_recent_age = tcp_ts_getticks();
}
}
so = inp->inp_socket;
txsd = &toep->txsd[toep->txsd_cidx];
plen = 0;
while (credits) {
KASSERT(credits >= txsd->tx_credits,
("%s: too many (or partial) credits", __func__));
credits -= txsd->tx_credits;
toep->tx_credits += txsd->tx_credits;
plen += txsd->plen;
txsd++;
toep->txsd_avail++;
KASSERT(toep->txsd_avail <= toep->txsd_total,
("%s: txsd avail > total", __func__));
if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
txsd = &toep->txsd[0];
toep->txsd_cidx = 0;
}
}
if (toep->tx_credits == toep->tx_total) {
toep->tx_nocompl = 0;
toep->plen_nocompl = 0;
}
if (toep->flags & TPF_TX_SUSPENDED &&
toep->tx_credits >= toep->tx_total / 4) {
toep->flags &= ~TPF_TX_SUSPENDED;
if (toep->ulp_mode == ULP_MODE_ISCSI)
t4_push_pdus(sc, toep, plen);
else
t4_push_frames(sc, toep, plen);
} else if (plen > 0) {
struct sockbuf *sb = &so->so_snd;
int sbu;
SOCKBUF_LOCK(sb);
sbu = sbused(sb);
if (toep->ulp_mode == ULP_MODE_ISCSI) {
if (__predict_false(sbu > 0)) {
/*
* The data trasmitted before the tid's ULP mode
* changed to ISCSI is still in so_snd.
* Incoming credits should account for so_snd
* first.
*/
sbdrop_locked(sb, min(sbu, plen));
plen -= min(sbu, plen);
}
sowwakeup_locked(so); /* unlocks so_snd */
rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
} else {
sbdrop_locked(sb, plen);
sowwakeup_locked(so); /* unlocks so_snd */
}
SOCKBUF_UNLOCK_ASSERT(sb);
}
INP_WUNLOCK(inp);
return (0);
}
static int
do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
{
struct adapter *sc = iq->adapter;
const struct cpl_rx_data *cpl = mtod(m, const void *);
unsigned int tid = GET_TID(cpl);
struct toepcb *toep = lookup_tid(sc, tid);
struct inpcb *inp = toep->inp;
struct tcpcb *tp;
struct socket *so;
struct sockbuf *sb;
int len;
uint32_t ddp_placed = 0;
if (__predict_false(toep->flags & TPF_SYNQE)) {
#ifdef INVARIANTS
struct synq_entry *synqe = (void *)toep;
INP_WLOCK(synqe->lctx->inp);
if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
("%s: listen socket closed but tid %u not aborted.",
__func__, tid));
} else {
/*
* do_pass_accept_req is still running and will
* eventually take care of this tid.
*/
}
INP_WUNLOCK(synqe->lctx->inp);
#endif
CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
toep, toep->flags);
m_freem(m);
return (0);
}
KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
/* strip off CPL header */
m_adj(m, sizeof(*cpl));
len = m->m_pkthdr.len;
INP_WLOCK(inp);
if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
__func__, tid, len, inp->inp_flags);
INP_WUNLOCK(inp);
m_freem(m);
return (0);
}
tp = intotcpcb(inp);
if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
tp->rcv_nxt += len;
if (tp->rcv_wnd < len) {
KASSERT(toep->ulp_mode == ULP_MODE_RDMA,
("%s: negative window size", __func__));
}
tp->rcv_wnd -= len;
tp->t_rcvtime = ticks;
if (toep->ulp_mode == ULP_MODE_TCPDDP)
DDP_LOCK(toep);
so = inp_inpcbtosocket(inp);
sb = &so->so_rcv;
SOCKBUF_LOCK(sb);
if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
__func__, tid, len);
m_freem(m);
SOCKBUF_UNLOCK(sb);
if (toep->ulp_mode == ULP_MODE_TCPDDP)
DDP_UNLOCK(toep);
INP_WUNLOCK(inp);
INP_INFO_RLOCK(&V_tcbinfo);
INP_WLOCK(inp);
tp = tcp_drop(tp, ECONNRESET);
if (tp)
INP_WUNLOCK(inp);
INP_INFO_RUNLOCK(&V_tcbinfo);
return (0);
}
/* receive buffer autosize */
CURVNET_SET(so->so_vnet);
if (sb->sb_flags & SB_AUTOSIZE &&
V_tcp_do_autorcvbuf &&
sb->sb_hiwat < V_tcp_autorcvbuf_max &&
len > (sbspace(sb) / 8 * 7)) {
unsigned int hiwat = sb->sb_hiwat;
unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
V_tcp_autorcvbuf_max);
if (!sbreserve_locked(sb, newsize, so, NULL))
sb->sb_flags &= ~SB_AUTOSIZE;
else
toep->rx_credits += newsize - hiwat;
}
if (toep->ddp_waiting_count != 0 || toep->ddp_active_count != 0)
CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)", __func__,
tid, len);
if (toep->ulp_mode == ULP_MODE_TCPDDP) {
int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
if (changed) {
if (toep->ddp_flags & DDP_SC_REQ)
toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
else {
KASSERT(cpl->ddp_off == 1,
("%s: DDP switched on by itself.",
__func__));
/* Fell out of DDP mode */
toep->ddp_flags &= ~DDP_ON;
CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
__func__);
insert_ddp_data(toep, ddp_placed);
}
}
if (toep->ddp_flags & DDP_ON) {
/*
* CPL_RX_DATA with DDP on can only be an indicate.
* Start posting queued AIO requests via DDP. The
* payload that arrived in this indicate is appended
* to the socket buffer as usual.
*/
handle_ddp_indicate(toep);
}
}
KASSERT(toep->sb_cc >= sbused(sb),
("%s: sb %p has more data (%d) than last time (%d).",
__func__, sb, sbused(sb), toep->sb_cc));
toep->rx_credits += toep->sb_cc - sbused(sb);
sbappendstream_locked(sb, m, 0);
toep->sb_cc = sbused(sb);
if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
int credits;
credits = send_rx_credits(sc, toep, toep->rx_credits);
toep->rx_credits -= credits;
tp->rcv_wnd += credits;
tp->rcv_adv += credits;
}
if (toep->ddp_waiting_count > 0 && sbavail(sb) != 0) {
CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__,
tid);
ddp_queue_toep(toep);
}
sorwakeup_locked(so);
SOCKBUF_UNLOCK_ASSERT(sb);
if (toep->ulp_mode == ULP_MODE_TCPDDP)
DDP_UNLOCK(toep);
INP_WUNLOCK(inp);
CURVNET_RESTORE();
return (0);
}
static int
handle_ddp(struct socket *so, struct uio *uio, int flags, int error)
{
struct sockbuf *sb = &so->so_rcv;
struct tcpcb *tp = so_sototcpcb(so);
struct toepcb *toep = tp->t_toe;
struct adapter *sc = td_adapter(toep->td);
vm_page_t *pages;
int npages, db_idx, rc, buf_flag;
struct ddp_buffer *db;
struct wrqe *wr;
uint64_t ddp_flags;
SOCKBUF_LOCK_ASSERT(sb);
#if 0
if (sbused(sb) + sc->tt.ddp_thres > uio->uio_resid) {
CTR4(KTR_CXGBE, "%s: sb_cc %d, threshold %d, resid %d",
__func__, sbused(sb), sc->tt.ddp_thres, uio->uio_resid);
}
#endif
/* XXX: too eager to disable DDP, could handle NBIO better than this. */
if (sbused(sb) >= uio->uio_resid || uio->uio_resid < sc->tt.ddp_thres ||
uio->uio_resid > MAX_DDP_BUFFER_SIZE || uio->uio_iovcnt > 1 ||
so->so_state & SS_NBIO || flags & (MSG_DONTWAIT | MSG_NBIO) ||
error || so->so_error || sb->sb_state & SBS_CANTRCVMORE)
goto no_ddp;
/*
* Fault in and then hold the pages of the uio buffers. We'll wire them
* a bit later if everything else works out.
*/
SOCKBUF_UNLOCK(sb);
if (hold_uio(uio, &pages, &npages) != 0) {
SOCKBUF_LOCK(sb);
goto no_ddp;
}
SOCKBUF_LOCK(sb);
if (__predict_false(so->so_error || sb->sb_state & SBS_CANTRCVMORE)) {
vm_page_unhold_pages(pages, npages);
free(pages, M_CXGBE);
goto no_ddp;
}
/*
* Figure out which one of the two DDP buffers to use this time.
*/
db_idx = select_ddp_buffer(sc, toep, pages, npages,
(uintptr_t)uio->uio_iov->iov_base & PAGE_MASK, uio->uio_resid);
pages = NULL; /* handed off to select_ddp_buffer */
if (db_idx < 0)
goto no_ddp;
db = toep->db[db_idx];
buf_flag = db_idx == 0 ? DDP_BUF0_ACTIVE : DDP_BUF1_ACTIVE;
/*
* Build the compound work request that tells the chip where to DMA the
* payload.
*/
ddp_flags = select_ddp_flags(so, flags, db_idx);
wr = mk_update_tcb_for_ddp(sc, toep, db_idx, sbused(sb), ddp_flags);
if (wr == NULL) {
/*
* Just unhold the pages. The DDP buffer's software state is
* left as-is in the toep. The page pods were written
* successfully and we may have an opportunity to use it in the
* future.
*/
vm_page_unhold_pages(db->pages, db->npages);
goto no_ddp;
}
/* Wire (and then unhold) the pages, and give the chip the go-ahead. */
wire_ddp_buffer(db);
t4_wrq_tx(sc, wr);
sb->sb_flags &= ~SB_DDP_INDICATE;
toep->ddp_flags |= buf_flag;
/*
* Wait for the DDP operation to complete and then unwire the pages.
* The return code from the sbwait will be the final return code of this
* function. But we do need to wait for DDP no matter what.
*/
rc = sbwait(sb);
while (toep->ddp_flags & buf_flag) {
/* XXXGL: shouldn't here be sbwait() call? */
sb->sb_flags |= SB_WAIT;
msleep(&sb->sb_acc, &sb->sb_mtx, PSOCK , "sbwait", 0);
}
unwire_ddp_buffer(db);
return (rc);
no_ddp:
disable_ddp(sc, toep);
discourage_ddp(toep);
sb->sb_flags &= ~SB_DDP_INDICATE;
return (0);
}
static int
do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
{
struct adapter *sc = iq->adapter;
const struct cpl_rx_data *cpl = mtod(m, const void *);
unsigned int tid = GET_TID(cpl);
struct toepcb *toep = lookup_tid(sc, tid);
struct inpcb *inp = toep->inp;
struct tcpcb *tp;
struct socket *so;
struct sockbuf *sb;
int len;
uint32_t ddp_placed = 0;
if (__predict_false(toep->flags & TPF_SYNQE)) {
#ifdef INVARIANTS
struct synq_entry *synqe = (void *)toep;
INP_WLOCK(synqe->lctx->inp);
if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
("%s: listen socket closed but tid %u not aborted.",
__func__, tid));
} else {
/*
* do_pass_accept_req is still running and will
* eventually take care of this tid.
*/
}
INP_WUNLOCK(synqe->lctx->inp);
#endif
CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
toep, toep->flags);
m_freem(m);
return (0);
}
KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
/* strip off CPL header */
m_adj(m, sizeof(*cpl));
len = m->m_pkthdr.len;
INP_WLOCK(inp);
if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
__func__, tid, len, inp->inp_flags);
INP_WUNLOCK(inp);
m_freem(m);
return (0);
}
tp = intotcpcb(inp);
if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
tp->rcv_nxt += len;
KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
tp->rcv_wnd -= len;
tp->t_rcvtime = ticks;
so = inp_inpcbtosocket(inp);
sb = &so->so_rcv;
SOCKBUF_LOCK(sb);
if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
__func__, tid, len);
m_freem(m);
SOCKBUF_UNLOCK(sb);
INP_WUNLOCK(inp);
INP_INFO_RLOCK(&V_tcbinfo);
INP_WLOCK(inp);
tp = tcp_drop(tp, ECONNRESET);
if (tp)
INP_WUNLOCK(inp);
INP_INFO_RUNLOCK(&V_tcbinfo);
return (0);
}
/* receive buffer autosize */
if (sb->sb_flags & SB_AUTOSIZE &&
V_tcp_do_autorcvbuf &&
sb->sb_hiwat < V_tcp_autorcvbuf_max &&
len > (sbspace(sb) / 8 * 7)) {
unsigned int hiwat = sb->sb_hiwat;
unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
V_tcp_autorcvbuf_max);
if (!sbreserve_locked(sb, newsize, so, NULL))
sb->sb_flags &= ~SB_AUTOSIZE;
else
toep->rx_credits += newsize - hiwat;
}
if (toep->ulp_mode == ULP_MODE_TCPDDP) {
int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
if (changed) {
if (toep->ddp_flags & DDP_SC_REQ)
toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
else {
KASSERT(cpl->ddp_off == 1,
("%s: DDP switched on by itself.",
__func__));
/* Fell out of DDP mode */
toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
DDP_BUF1_ACTIVE);
if (ddp_placed)
insert_ddp_data(toep, ddp_placed);
}
}
if ((toep->ddp_flags & DDP_OK) == 0 &&
time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
toep->ddp_score = DDP_LOW_SCORE;
toep->ddp_flags |= DDP_OK;
CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
__func__, tid, time_uptime);
}
if (toep->ddp_flags & DDP_ON) {
/*
* CPL_RX_DATA with DDP on can only be an indicate. Ask
* soreceive to post a buffer or disable DDP. The
* payload that arrived in this indicate is appended to
* the socket buffer as usual.
*/
#if 0
CTR5(KTR_CXGBE,
"%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
__func__, tid, toep->flags, be32toh(cpl->seq), len);
#endif
sb->sb_flags |= SB_DDP_INDICATE;
} else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
/*
* DDP allowed but isn't on (and a request to switch it
* on isn't pending either), and conditions are ripe for
* it to work. Switch it on.
*/
enable_ddp(sc, toep);
}
}
KASSERT(toep->sb_cc >= sbused(sb),
("%s: sb %p has more data (%d) than last time (%d).",
__func__, sb, sbused(sb), toep->sb_cc));
toep->rx_credits += toep->sb_cc - sbused(sb);
sbappendstream_locked(sb, m, 0);
toep->sb_cc = sbused(sb);
if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
int credits;
credits = send_rx_credits(sc, toep, toep->rx_credits);
toep->rx_credits -= credits;
tp->rcv_wnd += credits;
tp->rcv_adv += credits;
}
sorwakeup_locked(so);
SOCKBUF_UNLOCK_ASSERT(sb);
INP_WUNLOCK(inp);
return (0);
}
static int
soo_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp)
{
struct sockaddr *sa;
struct inpcb *inpcb;
struct unpcb *unpcb;
struct socket *so;
int error;
kif->kf_type = KF_TYPE_SOCKET;
so = fp->f_data;
kif->kf_un.kf_sock.kf_sock_domain0 =
so->so_proto->pr_domain->dom_family;
kif->kf_un.kf_sock.kf_sock_type0 = so->so_type;
kif->kf_un.kf_sock.kf_sock_protocol0 = so->so_proto->pr_protocol;
kif->kf_un.kf_sock.kf_sock_pcb = (uintptr_t)so->so_pcb;
switch (kif->kf_un.kf_sock.kf_sock_domain0) {
case AF_INET:
case AF_INET6:
if (kif->kf_un.kf_sock.kf_sock_protocol0 == IPPROTO_TCP) {
if (so->so_pcb != NULL) {
inpcb = (struct inpcb *)(so->so_pcb);
kif->kf_un.kf_sock.kf_sock_inpcb =
(uintptr_t)inpcb->inp_ppcb;
kif->kf_un.kf_sock.kf_sock_sendq =
sbused(&so->so_snd);
kif->kf_un.kf_sock.kf_sock_recvq =
sbused(&so->so_rcv);
}
}
break;
case AF_UNIX:
if (so->so_pcb != NULL) {
unpcb = (struct unpcb *)(so->so_pcb);
if (unpcb->unp_conn) {
kif->kf_un.kf_sock.kf_sock_unpconn =
(uintptr_t)unpcb->unp_conn;
kif->kf_un.kf_sock.kf_sock_rcv_sb_state =
so->so_rcv.sb_state;
kif->kf_un.kf_sock.kf_sock_snd_sb_state =
so->so_snd.sb_state;
kif->kf_un.kf_sock.kf_sock_sendq =
sbused(&so->so_snd);
kif->kf_un.kf_sock.kf_sock_recvq =
sbused(&so->so_rcv);
}
}
break;
}
error = so->so_proto->pr_usrreqs->pru_sockaddr(so, &sa);
if (error == 0 &&
sa->sa_len <= sizeof(kif->kf_un.kf_sock.kf_sa_local)) {
bcopy(sa, &kif->kf_un.kf_sock.kf_sa_local, sa->sa_len);
free(sa, M_SONAME);
}
error = so->so_proto->pr_usrreqs->pru_peeraddr(so, &sa);
if (error == 0 &&
sa->sa_len <= sizeof(kif->kf_un.kf_sock.kf_sa_peer)) {
bcopy(sa, &kif->kf_un.kf_sock.kf_sa_peer, sa->sa_len);
free(sa, M_SONAME);
}
strncpy(kif->kf_path, so->so_proto->pr_domain->dom_name,
sizeof(kif->kf_path));
return (0);
}