void
sbdrop(struct sockbuf *sb, int len)
{
SOCKBUF_LOCK(sb);
sbdrop_locked(sb, len);
SOCKBUF_UNLOCK(sb);
}
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 (plen > 0) {
struct sockbuf *sb = &so->so_snd;
SOCKBUF_LOCK(sb);
sbdrop_locked(sb, plen);
sowwakeup_locked(so);
SOCKBUF_UNLOCK_ASSERT(sb);
}
/* XXX */
if ((toep->flags & TPF_TX_SUSPENDED &&
toep->tx_credits >= MIN_OFLD_TX_CREDITS) ||
toep->tx_credits == toep->txsd_total *
howmany((sizeof(struct fw_ofld_tx_data_wr) + 1), 16)) {
toep->flags &= ~TPF_TX_SUSPENDED;
t4_push_frames(sc, toep);
}
INP_WUNLOCK(inp);
return (0);
}
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);
}
/*
* 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
userfw_sosend(struct socket *so,
int flags,
struct mbuf *m,
struct sockaddr *addr_,
struct mbuf *control,
struct thread *td)
{
int err = 0;
struct userfwpcb *pcb = sotopcb(so);
struct userfw_io_header msg;
int cmd_ready = 0;
unsigned char *data = NULL;
struct mdchain chain;
if (pcb == NULL)
err = ENOTCONN;
if (control != NULL)
err = EINVAL;
SOCKBUF_LOCK(&(so->so_snd));
if (err == 0)
{
sbappendstream_locked(&(so->so_snd), m);
m = NULL;
md_initm(&chain, so->so_snd.sb_mb);
if (SOCKBUF_LEN(so->so_snd) >= sizeof(msg))
{
md_get_mem(&chain, (caddr_t)(&msg), sizeof(msg), MB_MSYSTEM);
if (SOCKBUF_LEN(so->so_snd) >= msg.length)
cmd_ready = 1;
}
}
if (err == 0 && cmd_ready)
{
if (msg.type != T_CONTAINER || (msg.subtype != ST_MESSAGE && msg.subtype != ST_CMDCALL))
{
cmd_ready = 0;
sbdrop_locked(&(so->so_snd), msg.length);
}
}
if (err == 0 && cmd_ready)
{
data = malloc(msg.length, M_USERFW, M_WAITOK);
md_initm(&chain, so->so_snd.sb_mb);
md_get_mem(&chain, data, msg.length, MB_MSYSTEM);
err = userfw_cmd_dispatch(data, so, td);
sbdrop_locked(&(so->so_snd), msg.length);
free(data, M_USERFW);
}
SOCKBUF_UNLOCK(&(so->so_snd));
if (control != NULL)
m_freem(control);
if (m != NULL)
m_freem(m);
return err;
}
