2 * Copyright (c) 2012 Chelsio Communications, Inc.
4 * Written by: Navdeep Parhar <np@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
34 #include <sys/param.h>
35 #include <sys/types.h>
36 #include <sys/kernel.h>
38 #include <sys/module.h>
39 #include <sys/protosw.h>
40 #include <sys/domain.h>
41 #include <sys/socket.h>
42 #include <sys/socketvar.h>
43 #include <sys/sglist.h>
44 #include <netinet/in.h>
45 #include <netinet/in_pcb.h>
46 #include <netinet/ip.h>
47 #include <netinet/tcp_var.h>
49 #include <netinet/tcp_fsm.h>
50 #include <netinet/tcp_seq.h>
51 #include <netinet/toecore.h>
53 #include "common/common.h"
54 #include "common/t4_msg.h"
55 #include "common/t4_regs.h"
56 #include "common/t4_tcb.h"
57 #include "tom/t4_tom_l2t.h"
58 #include "tom/t4_tom.h"
60 VNET_DECLARE(int, tcp_do_autosndbuf);
61 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
62 VNET_DECLARE(int, tcp_autosndbuf_inc);
63 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
64 VNET_DECLARE(int, tcp_autosndbuf_max);
65 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
66 VNET_DECLARE(int, tcp_do_autorcvbuf);
67 #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf)
68 VNET_DECLARE(int, tcp_autorcvbuf_inc);
69 #define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc)
70 VNET_DECLARE(int, tcp_autorcvbuf_max);
71 #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max)
74 send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp)
77 struct fw_flowc_wr *flowc;
78 unsigned int nparams = ftxp ? 8 : 6, flowclen;
79 struct port_info *pi = toep->port;
80 struct adapter *sc = pi->adapter;
81 unsigned int pfvf = G_FW_VIID_PFN(pi->viid) << S_FW_VIID_PFN;
82 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
84 KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
85 ("%s: flowc for tid %u sent already", __func__, toep->tid));
87 CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid);
89 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
91 wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
94 panic("%s: allocation failure.", __func__);
97 memset(flowc, 0, wr->wr_len);
99 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
100 V_FW_FLOWC_WR_NPARAMS(nparams));
101 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
102 V_FW_WR_FLOWID(toep->tid));
104 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
105 flowc->mnemval[0].val = htobe32(pfvf);
106 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
107 flowc->mnemval[1].val = htobe32(pi->tx_chan);
108 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
109 flowc->mnemval[2].val = htobe32(pi->tx_chan);
110 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
111 flowc->mnemval[3].val = htobe32(toep->ofld_rxq->iq.abs_id);
113 uint32_t sndbuf = min(ftxp->snd_space, sc->tt.sndbuf);
115 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDNXT;
116 flowc->mnemval[4].val = htobe32(ftxp->snd_nxt);
117 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_RCVNXT;
118 flowc->mnemval[5].val = htobe32(ftxp->rcv_nxt);
119 flowc->mnemval[6].mnemonic = FW_FLOWC_MNEM_SNDBUF;
120 flowc->mnemval[6].val = htobe32(sndbuf);
121 flowc->mnemval[7].mnemonic = FW_FLOWC_MNEM_MSS;
122 flowc->mnemval[7].val = htobe32(ftxp->mss);
124 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF;
125 flowc->mnemval[4].val = htobe32(512);
126 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS;
127 flowc->mnemval[5].val = htobe32(512);
130 txsd->tx_credits = howmany(flowclen, 16);
132 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
133 ("%s: not enough credits (%d)", __func__, toep->tx_credits));
134 toep->tx_credits -= txsd->tx_credits;
135 if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
139 toep->flags |= TPF_FLOWC_WR_SENT;
144 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
147 struct cpl_abort_req *req;
149 struct inpcb *inp = toep->inp;
150 struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */
152 INP_WLOCK_ASSERT(inp);
154 CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
156 inp->inp_flags & INP_DROPPED ? "inp dropped" :
157 tcpstates[tp->t_state],
158 toep->flags, inp->inp_flags,
159 toep->flags & TPF_ABORT_SHUTDOWN ?
160 " (abort already in progress)" : "");
162 if (toep->flags & TPF_ABORT_SHUTDOWN)
163 return; /* abort already in progress */
165 toep->flags |= TPF_ABORT_SHUTDOWN;
167 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
168 ("%s: flowc_wr not sent for tid %d.", __func__, tid));
170 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
173 panic("%s: allocation failure.", __func__);
177 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
178 if (inp->inp_flags & INP_DROPPED)
179 req->rsvd0 = htobe32(snd_nxt);
181 req->rsvd0 = htobe32(tp->snd_nxt);
182 req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
183 req->cmd = CPL_ABORT_SEND_RST;
186 * XXX: What's the correct way to tell that the inp hasn't been detached
187 * from its socket? Should I even be flushing the snd buffer here?
189 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
190 struct socket *so = inp->inp_socket;
192 if (so != NULL) /* because I'm not sure. See comment above */
193 sbflush(&so->so_snd);
196 t4_l2t_send(sc, wr, toep->l2te);
200 * Called when a connection is established to translate the TCP options
201 * reported by HW to FreeBSD's native format.
204 assign_rxopt(struct tcpcb *tp, unsigned int opt)
206 struct toepcb *toep = tp->t_toe;
207 struct adapter *sc = td_adapter(toep->td);
209 INP_LOCK_ASSERT(tp->t_inpcb);
211 tp->t_maxseg = tp->t_maxopd = sc->params.mtus[G_TCPOPT_MSS(opt)] - 40;
213 if (G_TCPOPT_TSTAMP(opt)) {
214 tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */
215 tp->ts_recent = 0; /* hmmm */
216 tp->ts_recent_age = tcp_ts_getticks();
217 tp->t_maxseg -= TCPOLEN_TSTAMP_APPA;
220 if (G_TCPOPT_SACK(opt))
221 tp->t_flags |= TF_SACK_PERMIT; /* should already be set */
223 tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */
225 if (G_TCPOPT_WSCALE_OK(opt))
226 tp->t_flags |= TF_RCVD_SCALE;
228 /* Doing window scaling? */
229 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
230 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
231 tp->rcv_scale = tp->request_r_scale;
232 tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
237 * Completes some final bits of initialization for just established connections
238 * and changes their state to TCPS_ESTABLISHED.
240 * The ISNs are from after the exchange of SYNs. i.e., the true ISN + 1.
243 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
246 struct inpcb *inp = toep->inp;
247 struct socket *so = inp->inp_socket;
248 struct tcpcb *tp = intotcpcb(inp);
250 uint32_t iss = be32toh(snd_isn) - 1; /* true ISS */
251 uint32_t irs = be32toh(rcv_isn) - 1; /* true IRS */
252 uint16_t tcpopt = be16toh(opt);
253 struct flowc_tx_params ftxp;
255 INP_WLOCK_ASSERT(inp);
256 KASSERT(tp->t_state == TCPS_SYN_SENT ||
257 tp->t_state == TCPS_SYN_RECEIVED,
258 ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
260 CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p",
261 __func__, toep->tid, toep, inp);
263 tp->t_state = TCPS_ESTABLISHED;
264 tp->t_starttime = ticks;
265 TCPSTAT_INC(tcps_connects);
269 tp->rcv_wnd = toep->rx_credits << 10;
270 tp->rcv_adv += tp->rcv_wnd;
271 tp->last_ack_sent = tp->rcv_nxt;
274 * If we were unable to send all rx credits via opt0, save the remainder
275 * in rx_credits so that they can be handed over with the next credit
278 SOCKBUF_LOCK(&so->so_rcv);
279 bufsize = select_rcv_wnd(so);
280 SOCKBUF_UNLOCK(&so->so_rcv);
281 toep->rx_credits = bufsize - tp->rcv_wnd;
285 tp->snd_una = iss + 1;
286 tp->snd_nxt = iss + 1;
287 tp->snd_max = iss + 1;
289 assign_rxopt(tp, tcpopt);
291 SOCKBUF_LOCK(&so->so_snd);
292 if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf)
293 bufsize = V_tcp_autosndbuf_max;
295 bufsize = sbspace(&so->so_snd);
296 SOCKBUF_UNLOCK(&so->so_snd);
298 ftxp.snd_nxt = tp->snd_nxt;
299 ftxp.rcv_nxt = tp->rcv_nxt;
300 ftxp.snd_space = bufsize;
301 ftxp.mss = tp->t_maxseg;
302 send_flowc_wr(toep, &ftxp);
308 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
311 struct cpl_rx_data_ack *req;
312 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
314 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
316 wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
321 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
322 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
329 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
331 struct adapter *sc = tod->tod_softc;
332 struct inpcb *inp = tp->t_inpcb;
333 struct socket *so = inp->inp_socket;
334 struct sockbuf *sb = &so->so_rcv;
335 struct toepcb *toep = tp->t_toe;
338 INP_WLOCK_ASSERT(inp);
341 KASSERT(toep->sb_cc >= sb->sb_cc,
342 ("%s: sb %p has more data (%d) than last time (%d).",
343 __func__, sb, sb->sb_cc, toep->sb_cc));
344 toep->rx_credits += toep->sb_cc - sb->sb_cc;
345 toep->sb_cc = sb->sb_cc;
346 credits = toep->rx_credits;
350 (credits + 16384 >= tp->rcv_wnd || credits >= 15 * 1024)) {
352 credits = send_rx_credits(sc, toep, credits);
354 toep->rx_credits -= credits;
356 tp->rcv_wnd += credits;
357 tp->rcv_adv += credits;
362 * Close a connection by sending a CPL_CLOSE_CON_REQ message.
365 close_conn(struct adapter *sc, struct toepcb *toep)
368 struct cpl_close_con_req *req;
369 unsigned int tid = toep->tid;
371 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
372 toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
374 if (toep->flags & TPF_FIN_SENT)
377 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
378 ("%s: flowc_wr not sent for tid %u.", __func__, tid));
380 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
383 panic("%s: allocation failure.", __func__);
387 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
388 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
389 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
390 V_FW_WR_FLOWID(tid));
391 req->wr.wr_lo = cpu_to_be64(0);
392 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
395 toep->flags |= TPF_FIN_SENT;
396 toep->flags &= ~TPF_SEND_FIN;
397 t4_l2t_send(sc, wr, toep->l2te);
402 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
403 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
405 /* Maximum amount of immediate data we could stuff in a WR */
407 max_imm_payload(int tx_credits)
409 const int n = 2; /* Use only up to 2 desc for imm. data WR */
411 KASSERT(tx_credits >= 0 &&
412 tx_credits <= MAX_OFLD_TX_CREDITS,
413 ("%s: %d credits", __func__, tx_credits));
415 if (tx_credits < MIN_OFLD_TX_CREDITS)
418 if (tx_credits >= (n * EQ_ESIZE) / 16)
419 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
421 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
424 /* Maximum number of SGL entries we could stuff in a WR */
426 max_dsgl_nsegs(int tx_credits)
428 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
429 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
431 KASSERT(tx_credits >= 0 &&
432 tx_credits <= MAX_OFLD_TX_CREDITS,
433 ("%s: %d credits", __func__, tx_credits));
435 if (tx_credits < MIN_OFLD_TX_CREDITS)
438 nseg += 2 * (sge_pair_credits * 16 / 24);
439 if ((sge_pair_credits * 16) % 24 == 16)
446 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
447 unsigned int plen, uint8_t credits, int shove)
449 struct fw_ofld_tx_data_wr *txwr = dst;
451 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
452 V_FW_WR_IMMDLEN(immdlen));
453 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
454 V_FW_WR_LEN16(credits));
455 txwr->tunnel_to_proxy =
456 htobe32(V_FW_OFLD_TX_DATA_WR_ULPMODE(toep->ulp_mode) |
457 V_FW_OFLD_TX_DATA_WR_URGENT(0) | /* XXX */
458 V_FW_OFLD_TX_DATA_WR_SHOVE(shove));
459 txwr->plen = htobe32(plen);
463 * Generate a DSGL from a starting mbuf. The total number of segments and the
464 * maximum segments in any one mbuf are provided.
467 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
470 struct ulptx_sgl *usgl = dst;
473 struct sglist_seg segs[n];
475 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
477 sglist_init(&sg, n, segs);
478 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
479 V_ULPTX_NSGE(nsegs));
482 for (m = start; m != stop; m = m->m_next) {
483 rc = sglist_append(&sg, mtod(m, void *), m->m_len);
484 if (__predict_false(rc != 0))
485 panic("%s: sglist_append %d", __func__, rc);
487 for (j = 0; j < sg.sg_nseg; i++, j++) {
489 usgl->len0 = htobe32(segs[j].ss_len);
490 usgl->addr0 = htobe64(segs[j].ss_paddr);
492 usgl->sge[i / 2].len[i & 1] =
493 htobe32(segs[j].ss_len);
494 usgl->sge[i / 2].addr[i & 1] =
495 htobe64(segs[j].ss_paddr);
504 usgl->sge[i / 2].len[1] = htobe32(0);
505 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
506 __func__, nsegs, start, stop));
510 * Max number of SGL entries an offload tx work request can have. This is 41
511 * (1 + 40) for a full 512B work request.
512 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
514 #define OFLD_SGL_LEN (41)
517 * Send data and/or a FIN to the peer.
519 * The socket's so_snd buffer consists of a stream of data starting with sb_mb
520 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that
523 * drop indicates the number of bytes that should be dropped from the head of
524 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating
525 * contention on the send buffer lock (before this change it used to do
526 * sowwakeup and then t4_push_frames right after that when recovering from tx
527 * stalls). When drop is set this function MUST drop the bytes and wake up any
531 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
533 struct mbuf *sndptr, *m, *sb_sndptr;
534 struct fw_ofld_tx_data_wr *txwr;
536 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
537 struct inpcb *inp = toep->inp;
538 struct tcpcb *tp = intotcpcb(inp);
539 struct socket *so = inp->inp_socket;
540 struct sockbuf *sb = &so->so_snd;
541 int tx_credits, shove, compl, space, sowwakeup;
542 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
544 INP_WLOCK_ASSERT(inp);
545 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
546 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
548 KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
549 toep->ulp_mode == ULP_MODE_TCPDDP ||
550 toep->ulp_mode == ULP_MODE_RDMA,
551 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
554 * This function doesn't resume by itself. Someone else must clear the
555 * flag and call this function.
557 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
559 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
564 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
565 max_imm = max_imm_payload(tx_credits);
566 max_nsegs = max_dsgl_nsegs(tx_credits);
571 sbdrop_locked(sb, drop);
574 sb_sndptr = sb->sb_sndptr;
575 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
578 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
579 for (m = sndptr; m != NULL; m = m->m_next) {
580 int n = sglist_count(mtod(m, void *), m->m_len);
585 /* This mbuf sent us _over_ the nsegs limit, back out */
586 if (plen > max_imm && nsegs > max_nsegs) {
590 /* Too few credits */
591 toep->flags |= TPF_TX_SUSPENDED;
593 sowwakeup_locked(so);
596 SOCKBUF_UNLOCK_ASSERT(sb);
602 if (max_nsegs_1mbuf < n)
604 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */
606 /* This mbuf put us right at the max_nsegs limit */
607 if (plen > max_imm && nsegs == max_nsegs) {
613 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
616 if (space <= sb->sb_hiwat * 3 / 8 &&
617 toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
622 if (sb->sb_flags & SB_AUTOSIZE &&
623 V_tcp_do_autosndbuf &&
624 sb->sb_hiwat < V_tcp_autosndbuf_max &&
625 space < sb->sb_hiwat / 8) {
626 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
627 V_tcp_autosndbuf_max);
629 if (!sbreserve_locked(sb, newsize, so, NULL))
630 sb->sb_flags &= ~SB_AUTOSIZE;
632 sowwakeup = 1; /* room available */
635 sowwakeup_locked(so);
638 SOCKBUF_UNLOCK_ASSERT(sb);
640 /* nothing to send */
643 ("%s: nothing to send, but m != NULL", __func__));
647 if (__predict_false(toep->flags & TPF_FIN_SENT))
648 panic("%s: excess tx.", __func__);
650 if (plen <= max_imm) {
652 /* Immediate data tx */
654 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
657 /* XXX: how will we recover from this? */
658 toep->flags |= TPF_TX_SUSPENDED;
662 credits = howmany(wr->wr_len, 16);
663 write_tx_wr(txwr, toep, plen, plen, credits, shove);
664 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
671 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
672 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
673 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
675 /* XXX: how will we recover from this? */
676 toep->flags |= TPF_TX_SUSPENDED;
680 credits = howmany(wr_len, 16);
681 write_tx_wr(txwr, toep, 0, plen, credits, shove);
682 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
685 uint64_t *pad = (uint64_t *)
686 ((uintptr_t)txwr + wr_len);
691 KASSERT(toep->tx_credits >= credits,
692 ("%s: not enough credits", __func__));
694 toep->tx_credits -= credits;
695 toep->tx_nocompl += credits;
696 toep->plen_nocompl += plen;
697 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
698 toep->tx_nocompl >= toep->tx_total / 4)
702 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
703 toep->tx_nocompl = 0;
704 toep->plen_nocompl = 0;
711 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
712 sb->sb_sndptr = sb_sndptr;
715 toep->flags |= TPF_TX_DATA_SENT;
716 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
717 toep->flags |= TPF_TX_SUSPENDED;
719 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
721 txsd->tx_credits = credits;
723 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
725 txsd = &toep->txsd[0];
729 t4_l2t_send(sc, wr, toep->l2te);
732 /* Send a FIN if requested, but only if there's no more data to send */
733 if (m == NULL && toep->flags & TPF_SEND_FIN)
734 close_conn(sc, toep);
738 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
740 struct adapter *sc = tod->tod_softc;
742 struct inpcb *inp = tp->t_inpcb;
744 struct toepcb *toep = tp->t_toe;
746 INP_WLOCK_ASSERT(inp);
747 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
748 ("%s: inp %p dropped.", __func__, inp));
749 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
751 t4_push_frames(sc, toep, 0);
757 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
759 struct adapter *sc = tod->tod_softc;
761 struct inpcb *inp = tp->t_inpcb;
763 struct toepcb *toep = tp->t_toe;
765 INP_WLOCK_ASSERT(inp);
766 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
767 ("%s: inp %p dropped.", __func__, inp));
768 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
770 toep->flags |= TPF_SEND_FIN;
771 if (tp->t_state >= TCPS_ESTABLISHED)
772 t4_push_frames(sc, toep, 0);
778 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
780 struct adapter *sc = tod->tod_softc;
781 #if defined(INVARIANTS)
782 struct inpcb *inp = tp->t_inpcb;
784 struct toepcb *toep = tp->t_toe;
786 INP_WLOCK_ASSERT(inp);
787 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
788 ("%s: inp %p dropped.", __func__, inp));
789 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
792 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
793 ("%s: flowc for tid %u [%s] not sent already",
794 __func__, toep->tid, tcpstates[tp->t_state]));
796 send_reset(sc, toep, 0);
801 * Peer has sent us a FIN.
804 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
806 struct adapter *sc = iq->adapter;
807 const struct cpl_peer_close *cpl = (const void *)(rss + 1);
808 unsigned int tid = GET_TID(cpl);
809 struct toepcb *toep = lookup_tid(sc, tid);
810 struct inpcb *inp = toep->inp;
811 struct tcpcb *tp = NULL;
815 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
818 KASSERT(opcode == CPL_PEER_CLOSE,
819 ("%s: unexpected opcode 0x%x", __func__, opcode));
820 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
822 if (__predict_false(toep->flags & TPF_SYNQE)) {
824 struct synq_entry *synqe = (void *)toep;
826 INP_WLOCK(synqe->lctx->inp);
827 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
828 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
829 ("%s: listen socket closed but tid %u not aborted.",
833 * do_pass_accept_req is still running and will
834 * eventually take care of this tid.
837 INP_WUNLOCK(synqe->lctx->inp);
839 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
844 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
846 INP_INFO_WLOCK(&V_tcbinfo);
850 CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
851 tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
853 if (toep->flags & TPF_ABORT_SHUTDOWN)
856 tp->rcv_nxt++; /* FIN */
858 so = inp->inp_socket;
861 if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
862 m = get_ddp_mbuf(be32toh(cpl->rcv_nxt) - tp->rcv_nxt);
863 tp->rcv_nxt = be32toh(cpl->rcv_nxt);
864 toep->ddp_flags &= ~(DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE);
866 KASSERT(toep->sb_cc >= sb->sb_cc,
867 ("%s: sb %p has more data (%d) than last time (%d).",
868 __func__, sb, sb->sb_cc, toep->sb_cc));
869 toep->rx_credits += toep->sb_cc - sb->sb_cc;
870 #ifdef USE_DDP_RX_FLOW_CONTROL
871 toep->rx_credits -= m->m_len; /* adjust for F_RX_FC_DDP */
873 sbappendstream_locked(sb, m);
874 toep->sb_cc = sb->sb_cc;
876 socantrcvmore_locked(so); /* unlocks the sockbuf */
878 if (toep->ulp_mode != ULP_MODE_RDMA) {
879 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
880 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
881 be32toh(cpl->rcv_nxt)));
884 switch (tp->t_state) {
885 case TCPS_SYN_RECEIVED:
886 tp->t_starttime = ticks;
889 case TCPS_ESTABLISHED:
890 tp->t_state = TCPS_CLOSE_WAIT;
893 case TCPS_FIN_WAIT_1:
894 tp->t_state = TCPS_CLOSING;
897 case TCPS_FIN_WAIT_2:
899 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
900 INP_INFO_WUNLOCK(&V_tcbinfo);
903 final_cpl_received(toep);
907 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
908 __func__, tid, tp->t_state);
912 INP_INFO_WUNLOCK(&V_tcbinfo);
917 * Peer has ACK'd our FIN.
920 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
923 struct adapter *sc = iq->adapter;
924 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
925 unsigned int tid = GET_TID(cpl);
926 struct toepcb *toep = lookup_tid(sc, tid);
927 struct inpcb *inp = toep->inp;
928 struct tcpcb *tp = NULL;
929 struct socket *so = NULL;
931 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
934 KASSERT(opcode == CPL_CLOSE_CON_RPL,
935 ("%s: unexpected opcode 0x%x", __func__, opcode));
936 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
937 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
939 INP_INFO_WLOCK(&V_tcbinfo);
943 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
944 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
946 if (toep->flags & TPF_ABORT_SHUTDOWN)
949 so = inp->inp_socket;
950 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */
952 switch (tp->t_state) {
953 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */
956 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
957 INP_INFO_WUNLOCK(&V_tcbinfo);
960 final_cpl_received(toep); /* no more CPLs expected */
968 case TCPS_FIN_WAIT_1:
969 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
970 soisdisconnected(so);
971 tp->t_state = TCPS_FIN_WAIT_2;
976 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
977 __func__, tid, tcpstates[tp->t_state]);
981 INP_INFO_WUNLOCK(&V_tcbinfo);
986 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
990 struct cpl_abort_rpl *cpl;
992 wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
995 panic("%s: allocation failure.", __func__);
999 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1000 cpl->cmd = rst_status;
1006 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1008 switch (abort_reason) {
1009 case CPL_ERR_BAD_SYN:
1010 case CPL_ERR_CONN_RESET:
1011 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1012 case CPL_ERR_XMIT_TIMEDOUT:
1013 case CPL_ERR_PERSIST_TIMEDOUT:
1014 case CPL_ERR_FINWAIT2_TIMEDOUT:
1015 case CPL_ERR_KEEPALIVE_TIMEDOUT:
1023 * TCP RST from the peer, timeout, or some other such critical error.
1026 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1028 struct adapter *sc = iq->adapter;
1029 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1030 unsigned int tid = GET_TID(cpl);
1031 struct toepcb *toep = lookup_tid(sc, tid);
1032 struct sge_wrq *ofld_txq = toep->ofld_txq;
1036 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1039 KASSERT(opcode == CPL_ABORT_REQ_RSS,
1040 ("%s: unexpected opcode 0x%x", __func__, opcode));
1041 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1043 if (toep->flags & TPF_SYNQE)
1044 return (do_abort_req_synqe(iq, rss, m));
1046 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1048 if (negative_advice(cpl->status)) {
1049 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1050 __func__, cpl->status, tid, toep->flags);
1051 return (0); /* Ignore negative advice */
1055 INP_INFO_WLOCK(&V_tcbinfo); /* for tcp_close */
1058 tp = intotcpcb(inp);
1061 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1062 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1063 inp->inp_flags, cpl->status);
1066 * If we'd initiated an abort earlier the reply to it is responsible for
1067 * cleaning up resources. Otherwise we tear everything down right here
1068 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
1070 if (toep->flags & TPF_ABORT_SHUTDOWN) {
1074 toep->flags |= TPF_ABORT_SHUTDOWN;
1076 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1077 struct socket *so = inp->inp_socket;
1080 so_error_set(so, abort_status_to_errno(tp,
1084 INP_WLOCK(inp); /* re-acquire */
1087 final_cpl_received(toep);
1089 INP_INFO_WUNLOCK(&V_tcbinfo);
1090 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1095 * Reply to the CPL_ABORT_REQ (send_reset)
1098 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1100 struct adapter *sc = iq->adapter;
1101 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1102 unsigned int tid = GET_TID(cpl);
1103 struct toepcb *toep = lookup_tid(sc, tid);
1104 struct inpcb *inp = toep->inp;
1106 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1109 KASSERT(opcode == CPL_ABORT_RPL_RSS,
1110 ("%s: unexpected opcode 0x%x", __func__, opcode));
1111 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1113 if (toep->flags & TPF_SYNQE)
1114 return (do_abort_rpl_synqe(iq, rss, m));
1116 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1118 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1119 __func__, tid, toep, inp, cpl->status);
1121 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1122 ("%s: wasn't expecting abort reply", __func__));
1125 final_cpl_received(toep);
1131 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1133 struct adapter *sc = iq->adapter;
1134 const struct cpl_rx_data *cpl = mtod(m, const void *);
1135 unsigned int tid = GET_TID(cpl);
1136 struct toepcb *toep = lookup_tid(sc, tid);
1137 struct inpcb *inp = toep->inp;
1142 uint32_t ddp_placed = 0;
1144 if (__predict_false(toep->flags & TPF_SYNQE)) {
1146 struct synq_entry *synqe = (void *)toep;
1148 INP_WLOCK(synqe->lctx->inp);
1149 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1150 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1151 ("%s: listen socket closed but tid %u not aborted.",
1155 * do_pass_accept_req is still running and will
1156 * eventually take care of this tid.
1159 INP_WUNLOCK(synqe->lctx->inp);
1161 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1167 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1169 /* strip off CPL header */
1170 m_adj(m, sizeof(*cpl));
1171 len = m->m_pkthdr.len;
1174 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1175 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1176 __func__, tid, len, inp->inp_flags);
1182 tp = intotcpcb(inp);
1184 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1185 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1188 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
1190 tp->t_rcvtime = ticks;
1192 so = inp_inpcbtosocket(inp);
1196 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1197 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1198 __func__, tid, len);
1203 INP_INFO_WLOCK(&V_tcbinfo);
1205 tp = tcp_drop(tp, ECONNRESET);
1208 INP_INFO_WUNLOCK(&V_tcbinfo);
1213 /* receive buffer autosize */
1214 if (sb->sb_flags & SB_AUTOSIZE &&
1215 V_tcp_do_autorcvbuf &&
1216 sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1217 len > (sbspace(sb) / 8 * 7)) {
1218 unsigned int hiwat = sb->sb_hiwat;
1219 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1220 V_tcp_autorcvbuf_max);
1222 if (!sbreserve_locked(sb, newsize, so, NULL))
1223 sb->sb_flags &= ~SB_AUTOSIZE;
1225 toep->rx_credits += newsize - hiwat;
1228 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1229 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1232 if (toep->ddp_flags & DDP_SC_REQ)
1233 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1235 KASSERT(cpl->ddp_off == 1,
1236 ("%s: DDP switched on by itself.",
1239 /* Fell out of DDP mode */
1240 toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1244 insert_ddp_data(toep, ddp_placed);
1248 if ((toep->ddp_flags & DDP_OK) == 0 &&
1249 time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1250 toep->ddp_score = DDP_LOW_SCORE;
1251 toep->ddp_flags |= DDP_OK;
1252 CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1253 __func__, tid, time_uptime);
1256 if (toep->ddp_flags & DDP_ON) {
1259 * CPL_RX_DATA with DDP on can only be an indicate. Ask
1260 * soreceive to post a buffer or disable DDP. The
1261 * payload that arrived in this indicate is appended to
1262 * the socket buffer as usual.
1267 "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1268 __func__, tid, toep->flags, be32toh(cpl->seq), len);
1270 sb->sb_flags |= SB_DDP_INDICATE;
1271 } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1272 tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1275 * DDP allowed but isn't on (and a request to switch it
1276 * on isn't pending either), and conditions are ripe for
1277 * it to work. Switch it on.
1280 enable_ddp(sc, toep);
1284 KASSERT(toep->sb_cc >= sb->sb_cc,
1285 ("%s: sb %p has more data (%d) than last time (%d).",
1286 __func__, sb, sb->sb_cc, toep->sb_cc));
1287 toep->rx_credits += toep->sb_cc - sb->sb_cc;
1288 sbappendstream_locked(sb, m);
1289 toep->sb_cc = sb->sb_cc;
1290 sorwakeup_locked(so);
1291 SOCKBUF_UNLOCK_ASSERT(sb);
1297 #define S_CPL_FW4_ACK_OPCODE 24
1298 #define M_CPL_FW4_ACK_OPCODE 0xff
1299 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1300 #define G_CPL_FW4_ACK_OPCODE(x) \
1301 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1303 #define S_CPL_FW4_ACK_FLOWID 0
1304 #define M_CPL_FW4_ACK_FLOWID 0xffffff
1305 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1306 #define G_CPL_FW4_ACK_FLOWID(x) \
1307 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1309 #define S_CPL_FW4_ACK_CR 24
1310 #define M_CPL_FW4_ACK_CR 0xff
1311 #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR)
1312 #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1314 #define S_CPL_FW4_ACK_SEQVAL 0
1315 #define M_CPL_FW4_ACK_SEQVAL 0x1
1316 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1317 #define G_CPL_FW4_ACK_SEQVAL(x) \
1318 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1319 #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U)
1322 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1324 struct adapter *sc = iq->adapter;
1325 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1326 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1327 struct toepcb *toep = lookup_tid(sc, tid);
1331 uint8_t credits = cpl->credits;
1332 struct ofld_tx_sdesc *txsd;
1335 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1339 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1340 * now this comes back carrying the credits for the flowc.
1342 if (__predict_false(toep->flags & TPF_SYNQE)) {
1343 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1344 ("%s: credits for a synq entry %p", __func__, toep));
1350 KASSERT(opcode == CPL_FW4_ACK,
1351 ("%s: unexpected opcode 0x%x", __func__, opcode));
1352 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1353 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1357 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1362 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1363 ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1365 tp = intotcpcb(inp);
1367 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1368 tcp_seq snd_una = be32toh(cpl->snd_una);
1371 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1373 "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1374 __func__, snd_una, toep->tid, tp->snd_una);
1378 if (tp->snd_una != snd_una) {
1379 tp->snd_una = snd_una;
1380 tp->ts_recent_age = tcp_ts_getticks();
1384 so = inp->inp_socket;
1385 txsd = &toep->txsd[toep->txsd_cidx];
1388 KASSERT(credits >= txsd->tx_credits,
1389 ("%s: too many (or partial) credits", __func__));
1390 credits -= txsd->tx_credits;
1391 toep->tx_credits += txsd->tx_credits;
1395 KASSERT(toep->txsd_avail <= toep->txsd_total,
1396 ("%s: txsd avail > total", __func__));
1397 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1398 txsd = &toep->txsd[0];
1399 toep->txsd_cidx = 0;
1403 if (toep->tx_credits == toep->tx_total) {
1404 toep->tx_nocompl = 0;
1405 toep->plen_nocompl = 0;
1408 if (toep->flags & TPF_TX_SUSPENDED &&
1409 toep->tx_credits >= toep->tx_total / 4) {
1410 toep->flags &= ~TPF_TX_SUSPENDED;
1411 t4_push_frames(sc, toep, plen);
1412 } else if (plen > 0) {
1413 struct sockbuf *sb = &so->so_snd;
1416 sbdrop_locked(sb, plen);
1417 sowwakeup_locked(so);
1418 SOCKBUF_UNLOCK_ASSERT(sb);
1427 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1429 struct adapter *sc = iq->adapter;
1430 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1431 unsigned int tid = GET_TID(cpl);
1433 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1436 KASSERT(opcode == CPL_SET_TCB_RPL,
1437 ("%s: unexpected opcode 0x%x", __func__, opcode));
1438 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1440 if (tid >= sc->tids.ftid_base &&
1441 tid < sc->tids.ftid_base + sc->tids.nftids)
1442 return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1444 CXGBE_UNIMPLEMENTED(__func__);
1448 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1449 uint16_t word, uint64_t mask, uint64_t val)
1452 struct cpl_set_tcb_field *req;
1454 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1457 panic("%s: allocation failure.", __func__);
1461 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1462 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1463 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1464 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1465 req->mask = htobe64(mask);
1466 req->val = htobe64(val);
1472 t4_init_cpl_io_handlers(struct adapter *sc)
1475 t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1476 t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1477 t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1478 t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1479 t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1480 t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1481 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1485 t4_uninit_cpl_io_handlers(struct adapter *sc)
1488 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);