2 * Copyright (c) 2012, 2015 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/ip6.h>
49 #include <netinet/tcp_fsm.h>
50 #include <netinet/tcp_seq.h>
51 #include <netinet/tcp_var.h>
52 #include <netinet/toecore.h>
54 #include "common/common.h"
55 #include "common/t4_msg.h"
56 #include "common/t4_regs.h"
57 #include "common/t4_tcb.h"
58 #include "tom/t4_tom_l2t.h"
59 #include "tom/t4_tom.h"
61 VNET_DECLARE(int, tcp_do_autosndbuf);
62 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
63 VNET_DECLARE(int, tcp_autosndbuf_inc);
64 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
65 VNET_DECLARE(int, tcp_autosndbuf_max);
66 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
67 VNET_DECLARE(int, tcp_do_autorcvbuf);
68 #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf)
69 VNET_DECLARE(int, tcp_autorcvbuf_inc);
70 #define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc)
71 VNET_DECLARE(int, tcp_autorcvbuf_max);
72 #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max)
75 send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp)
78 struct fw_flowc_wr *flowc;
79 unsigned int nparams = ftxp ? 8 : 6, flowclen;
80 struct vi_info *vi = toep->vi;
81 struct port_info *pi = vi->pi;
82 struct adapter *sc = pi->adapter;
83 unsigned int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
84 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
86 KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
87 ("%s: flowc for tid %u sent already", __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);
125 "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x",
126 __func__, toep->tid, ftxp->mss, sndbuf, ftxp->snd_nxt,
129 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF;
130 flowc->mnemval[4].val = htobe32(512);
131 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS;
132 flowc->mnemval[5].val = htobe32(512);
134 CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid);
137 txsd->tx_credits = howmany(flowclen, 16);
139 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
140 ("%s: not enough credits (%d)", __func__, toep->tx_credits));
141 toep->tx_credits -= txsd->tx_credits;
142 if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
146 toep->flags |= TPF_FLOWC_WR_SENT;
151 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
154 struct cpl_abort_req *req;
156 struct inpcb *inp = toep->inp;
157 struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */
159 INP_WLOCK_ASSERT(inp);
161 CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
163 inp->inp_flags & INP_DROPPED ? "inp dropped" :
164 tcpstates[tp->t_state],
165 toep->flags, inp->inp_flags,
166 toep->flags & TPF_ABORT_SHUTDOWN ?
167 " (abort already in progress)" : "");
169 if (toep->flags & TPF_ABORT_SHUTDOWN)
170 return; /* abort already in progress */
172 toep->flags |= TPF_ABORT_SHUTDOWN;
174 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
175 ("%s: flowc_wr not sent for tid %d.", __func__, tid));
177 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
180 panic("%s: allocation failure.", __func__);
184 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
185 if (inp->inp_flags & INP_DROPPED)
186 req->rsvd0 = htobe32(snd_nxt);
188 req->rsvd0 = htobe32(tp->snd_nxt);
189 req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
190 req->cmd = CPL_ABORT_SEND_RST;
193 * XXX: What's the correct way to tell that the inp hasn't been detached
194 * from its socket? Should I even be flushing the snd buffer here?
196 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
197 struct socket *so = inp->inp_socket;
199 if (so != NULL) /* because I'm not sure. See comment above */
200 sbflush(&so->so_snd);
203 t4_l2t_send(sc, wr, toep->l2te);
207 * Called when a connection is established to translate the TCP options
208 * reported by HW to FreeBSD's native format.
211 assign_rxopt(struct tcpcb *tp, unsigned int opt)
213 struct toepcb *toep = tp->t_toe;
214 struct inpcb *inp = tp->t_inpcb;
215 struct adapter *sc = td_adapter(toep->td);
218 INP_LOCK_ASSERT(inp);
220 if (inp->inp_inc.inc_flags & INC_ISIPV6)
221 n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
223 n = sizeof(struct ip) + sizeof(struct tcphdr);
224 tp->t_maxseg = sc->params.mtus[G_TCPOPT_MSS(opt)] - n;
226 CTR4(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u)", __func__, toep->tid,
227 G_TCPOPT_MSS(opt), sc->params.mtus[G_TCPOPT_MSS(opt)]);
229 if (G_TCPOPT_TSTAMP(opt)) {
230 tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */
231 tp->ts_recent = 0; /* hmmm */
232 tp->ts_recent_age = tcp_ts_getticks();
235 if (G_TCPOPT_SACK(opt))
236 tp->t_flags |= TF_SACK_PERMIT; /* should already be set */
238 tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */
240 if (G_TCPOPT_WSCALE_OK(opt))
241 tp->t_flags |= TF_RCVD_SCALE;
243 /* Doing window scaling? */
244 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
245 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
246 tp->rcv_scale = tp->request_r_scale;
247 tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
252 * Completes some final bits of initialization for just established connections
253 * and changes their state to TCPS_ESTABLISHED.
255 * The ISNs are from after the exchange of SYNs. i.e., the true ISN + 1.
258 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
261 struct inpcb *inp = toep->inp;
262 struct socket *so = inp->inp_socket;
263 struct tcpcb *tp = intotcpcb(inp);
265 uint32_t iss = be32toh(snd_isn) - 1; /* true ISS */
266 uint32_t irs = be32toh(rcv_isn) - 1; /* true IRS */
267 uint16_t tcpopt = be16toh(opt);
268 struct flowc_tx_params ftxp;
270 INP_WLOCK_ASSERT(inp);
271 KASSERT(tp->t_state == TCPS_SYN_SENT ||
272 tp->t_state == TCPS_SYN_RECEIVED,
273 ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
275 CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p",
276 __func__, toep->tid, toep, inp);
278 tp->t_state = TCPS_ESTABLISHED;
279 tp->t_starttime = ticks;
280 TCPSTAT_INC(tcps_connects);
284 tp->rcv_wnd = toep->rx_credits << 10;
285 tp->rcv_adv += tp->rcv_wnd;
286 tp->last_ack_sent = tp->rcv_nxt;
289 * If we were unable to send all rx credits via opt0, save the remainder
290 * in rx_credits so that they can be handed over with the next credit
293 SOCKBUF_LOCK(&so->so_rcv);
294 bufsize = select_rcv_wnd(so);
295 SOCKBUF_UNLOCK(&so->so_rcv);
296 toep->rx_credits = bufsize - tp->rcv_wnd;
300 tp->snd_una = iss + 1;
301 tp->snd_nxt = iss + 1;
302 tp->snd_max = iss + 1;
304 assign_rxopt(tp, tcpopt);
306 SOCKBUF_LOCK(&so->so_snd);
307 if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf)
308 bufsize = V_tcp_autosndbuf_max;
310 bufsize = sbspace(&so->so_snd);
311 SOCKBUF_UNLOCK(&so->so_snd);
313 ftxp.snd_nxt = tp->snd_nxt;
314 ftxp.rcv_nxt = tp->rcv_nxt;
315 ftxp.snd_space = bufsize;
316 ftxp.mss = tp->t_maxseg;
317 send_flowc_wr(toep, &ftxp);
323 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
326 struct cpl_rx_data_ack *req;
327 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
329 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
331 wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
336 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
337 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
344 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
346 struct adapter *sc = tod->tod_softc;
347 struct inpcb *inp = tp->t_inpcb;
348 struct socket *so = inp->inp_socket;
349 struct sockbuf *sb = &so->so_rcv;
350 struct toepcb *toep = tp->t_toe;
353 INP_WLOCK_ASSERT(inp);
356 KASSERT(toep->sb_cc >= sbused(sb),
357 ("%s: sb %p has more data (%d) than last time (%d).",
358 __func__, sb, sbused(sb), toep->sb_cc));
360 toep->rx_credits += toep->sb_cc - sbused(sb);
361 toep->sb_cc = sbused(sb);
363 if (toep->rx_credits > 0 &&
364 (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
365 (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
366 toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
368 credits = send_rx_credits(sc, toep, toep->rx_credits);
369 toep->rx_credits -= credits;
370 tp->rcv_wnd += credits;
371 tp->rcv_adv += credits;
377 * Close a connection by sending a CPL_CLOSE_CON_REQ message.
380 close_conn(struct adapter *sc, struct toepcb *toep)
383 struct cpl_close_con_req *req;
384 unsigned int tid = toep->tid;
386 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
387 toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
389 if (toep->flags & TPF_FIN_SENT)
392 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
393 ("%s: flowc_wr not sent for tid %u.", __func__, tid));
395 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
398 panic("%s: allocation failure.", __func__);
402 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
403 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
404 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
405 V_FW_WR_FLOWID(tid));
406 req->wr.wr_lo = cpu_to_be64(0);
407 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
410 toep->flags |= TPF_FIN_SENT;
411 toep->flags &= ~TPF_SEND_FIN;
412 t4_l2t_send(sc, wr, toep->l2te);
417 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
418 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
420 /* Maximum amount of immediate data we could stuff in a WR */
422 max_imm_payload(int tx_credits)
424 const int n = 2; /* Use only up to 2 desc for imm. data WR */
426 KASSERT(tx_credits >= 0 &&
427 tx_credits <= MAX_OFLD_TX_CREDITS,
428 ("%s: %d credits", __func__, tx_credits));
430 if (tx_credits < MIN_OFLD_TX_CREDITS)
433 if (tx_credits >= (n * EQ_ESIZE) / 16)
434 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
436 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
439 /* Maximum number of SGL entries we could stuff in a WR */
441 max_dsgl_nsegs(int tx_credits)
443 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
444 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
446 KASSERT(tx_credits >= 0 &&
447 tx_credits <= MAX_OFLD_TX_CREDITS,
448 ("%s: %d credits", __func__, tx_credits));
450 if (tx_credits < MIN_OFLD_TX_CREDITS)
453 nseg += 2 * (sge_pair_credits * 16 / 24);
454 if ((sge_pair_credits * 16) % 24 == 16)
461 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
462 unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign)
464 struct fw_ofld_tx_data_wr *txwr = dst;
466 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
467 V_FW_WR_IMMDLEN(immdlen));
468 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
469 V_FW_WR_LEN16(credits));
470 txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) |
471 V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
472 txwr->plen = htobe32(plen);
475 struct tcpcb *tp = intotcpcb(toep->inp);
477 if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi))
478 txwr->lsodisable_to_flags |=
479 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
481 txwr->lsodisable_to_flags |=
482 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
483 (tp->t_flags & TF_NODELAY ? 0 :
484 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
489 * Generate a DSGL from a starting mbuf. The total number of segments and the
490 * maximum segments in any one mbuf are provided.
493 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
496 struct ulptx_sgl *usgl = dst;
499 struct sglist_seg segs[n];
501 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
503 sglist_init(&sg, n, segs);
504 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
505 V_ULPTX_NSGE(nsegs));
508 for (m = start; m != stop; m = m->m_next) {
509 rc = sglist_append(&sg, mtod(m, void *), m->m_len);
510 if (__predict_false(rc != 0))
511 panic("%s: sglist_append %d", __func__, rc);
513 for (j = 0; j < sg.sg_nseg; i++, j++) {
515 usgl->len0 = htobe32(segs[j].ss_len);
516 usgl->addr0 = htobe64(segs[j].ss_paddr);
518 usgl->sge[i / 2].len[i & 1] =
519 htobe32(segs[j].ss_len);
520 usgl->sge[i / 2].addr[i & 1] =
521 htobe64(segs[j].ss_paddr);
530 usgl->sge[i / 2].len[1] = htobe32(0);
531 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
532 __func__, nsegs, start, stop));
536 * Max number of SGL entries an offload tx work request can have. This is 41
537 * (1 + 40) for a full 512B work request.
538 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
540 #define OFLD_SGL_LEN (41)
543 * Send data and/or a FIN to the peer.
545 * The socket's so_snd buffer consists of a stream of data starting with sb_mb
546 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that
549 * drop indicates the number of bytes that should be dropped from the head of
550 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating
551 * contention on the send buffer lock (before this change it used to do
552 * sowwakeup and then t4_push_frames right after that when recovering from tx
553 * stalls). When drop is set this function MUST drop the bytes and wake up any
557 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
559 struct mbuf *sndptr, *m, *sb_sndptr;
560 struct fw_ofld_tx_data_wr *txwr;
562 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
563 struct inpcb *inp = toep->inp;
564 struct tcpcb *tp = intotcpcb(inp);
565 struct socket *so = inp->inp_socket;
566 struct sockbuf *sb = &so->so_snd;
567 int tx_credits, shove, compl, space, sowwakeup;
568 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
570 INP_WLOCK_ASSERT(inp);
571 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
572 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
574 KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
575 toep->ulp_mode == ULP_MODE_TCPDDP ||
576 toep->ulp_mode == ULP_MODE_RDMA,
577 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
579 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
583 * This function doesn't resume by itself. Someone else must clear the
584 * flag and call this function.
586 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
588 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
593 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
594 max_imm = max_imm_payload(tx_credits);
595 max_nsegs = max_dsgl_nsegs(tx_credits);
600 sbdrop_locked(sb, drop);
603 sb_sndptr = sb->sb_sndptr;
604 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
607 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
608 for (m = sndptr; m != NULL; m = m->m_next) {
609 int n = sglist_count(mtod(m, void *), m->m_len);
614 /* This mbuf sent us _over_ the nsegs limit, back out */
615 if (plen > max_imm && nsegs > max_nsegs) {
619 /* Too few credits */
620 toep->flags |= TPF_TX_SUSPENDED;
622 sowwakeup_locked(so);
625 SOCKBUF_UNLOCK_ASSERT(sb);
631 if (max_nsegs_1mbuf < n)
633 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */
635 /* This mbuf put us right at the max_nsegs limit */
636 if (plen > max_imm && nsegs == max_nsegs) {
644 if (space <= sb->sb_hiwat * 3 / 8 &&
645 toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
650 if (sb->sb_flags & SB_AUTOSIZE &&
651 V_tcp_do_autosndbuf &&
652 sb->sb_hiwat < V_tcp_autosndbuf_max &&
653 space < sb->sb_hiwat / 8) {
654 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
655 V_tcp_autosndbuf_max);
657 if (!sbreserve_locked(sb, newsize, so, NULL))
658 sb->sb_flags &= ~SB_AUTOSIZE;
660 sowwakeup = 1; /* room available */
663 sowwakeup_locked(so);
666 SOCKBUF_UNLOCK_ASSERT(sb);
668 /* nothing to send */
671 ("%s: nothing to send, but m != NULL", __func__));
675 if (__predict_false(toep->flags & TPF_FIN_SENT))
676 panic("%s: excess tx.", __func__);
678 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
679 if (plen <= max_imm) {
681 /* Immediate data tx */
683 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
686 /* XXX: how will we recover from this? */
687 toep->flags |= TPF_TX_SUSPENDED;
691 credits = howmany(wr->wr_len, 16);
692 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
694 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
701 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
702 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
703 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
705 /* XXX: how will we recover from this? */
706 toep->flags |= TPF_TX_SUSPENDED;
710 credits = howmany(wr_len, 16);
711 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
713 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
716 uint64_t *pad = (uint64_t *)
717 ((uintptr_t)txwr + wr_len);
722 KASSERT(toep->tx_credits >= credits,
723 ("%s: not enough credits", __func__));
725 toep->tx_credits -= credits;
726 toep->tx_nocompl += credits;
727 toep->plen_nocompl += plen;
728 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
729 toep->tx_nocompl >= toep->tx_total / 4)
732 if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
733 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
734 toep->tx_nocompl = 0;
735 toep->plen_nocompl = 0;
742 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
743 sb->sb_sndptr = sb_sndptr;
746 toep->flags |= TPF_TX_DATA_SENT;
747 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
748 toep->flags |= TPF_TX_SUSPENDED;
750 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
752 txsd->tx_credits = credits;
754 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
756 txsd = &toep->txsd[0];
760 t4_l2t_send(sc, wr, toep->l2te);
763 /* Send a FIN if requested, but only if there's no more data to send */
764 if (m == NULL && toep->flags & TPF_SEND_FIN)
765 close_conn(sc, toep);
769 rqdrop_locked(struct mbufq *q, int plen)
774 m = mbufq_dequeue(q);
776 /* Too many credits. */
780 /* Partial credits. */
781 MPASS(plen >= m->m_pkthdr.len);
783 plen -= m->m_pkthdr.len;
789 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
791 struct mbuf *sndptr, *m;
792 struct fw_ofld_tx_data_wr *txwr;
794 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
795 u_int adjusted_plen, ulp_submode;
796 struct inpcb *inp = toep->inp;
797 struct tcpcb *tp = intotcpcb(inp);
798 int tx_credits, shove;
799 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
800 struct mbufq *pduq = &toep->ulp_pduq;
801 static const u_int ulp_extra_len[] = {0, 4, 4, 8};
803 INP_WLOCK_ASSERT(inp);
804 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
805 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
806 KASSERT(toep->ulp_mode == ULP_MODE_ISCSI,
807 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
809 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
813 * This function doesn't resume by itself. Someone else must clear the
814 * flag and call this function.
816 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
818 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
823 rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
825 while ((sndptr = mbufq_first(pduq)) != NULL) {
826 M_ASSERTPKTHDR(sndptr);
828 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
829 max_imm = max_imm_payload(tx_credits);
830 max_nsegs = max_dsgl_nsegs(tx_credits);
834 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
835 for (m = sndptr; m != NULL; m = m->m_next) {
836 int n = sglist_count(mtod(m, void *), m->m_len);
842 * This mbuf would send us _over_ the nsegs limit.
843 * Suspend tx because the PDU can't be sent out.
845 if (plen > max_imm && nsegs > max_nsegs) {
846 toep->flags |= TPF_TX_SUSPENDED;
850 if (max_nsegs_1mbuf < n)
854 if (__predict_false(toep->flags & TPF_FIN_SENT))
855 panic("%s: excess tx.", __func__);
858 * We have a PDU to send. All of it goes out in one WR so 'm'
859 * is NULL. A PDU's length is always a multiple of 4.
862 MPASS((plen & 3) == 0);
863 MPASS(sndptr->m_pkthdr.len == plen);
865 shove = !(tp->t_flags & TF_MORETOCOME);
866 ulp_submode = mbuf_ulp_submode(sndptr);
867 MPASS(ulp_submode < nitems(ulp_extra_len));
870 * plen doesn't include header and data digests, which are
871 * generated and inserted in the right places by the TOE, but
872 * they do occupy TCP sequence space and need to be accounted
875 adjusted_plen = plen + ulp_extra_len[ulp_submode];
876 if (plen <= max_imm) {
878 /* Immediate data tx */
880 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
883 /* XXX: how will we recover from this? */
884 toep->flags |= TPF_TX_SUSPENDED;
888 credits = howmany(wr->wr_len, 16);
889 write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
890 shove, ulp_submode, sc->tt.tx_align);
891 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
897 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
898 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
899 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
901 /* XXX: how will we recover from this? */
902 toep->flags |= TPF_TX_SUSPENDED;
906 credits = howmany(wr_len, 16);
907 write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
908 shove, ulp_submode, sc->tt.tx_align);
909 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
912 uint64_t *pad = (uint64_t *)
913 ((uintptr_t)txwr + wr_len);
918 KASSERT(toep->tx_credits >= credits,
919 ("%s: not enough credits", __func__));
921 m = mbufq_dequeue(pduq);
923 mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
925 toep->tx_credits -= credits;
926 toep->tx_nocompl += credits;
927 toep->plen_nocompl += plen;
928 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
929 toep->tx_nocompl >= toep->tx_total / 4) {
930 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
931 toep->tx_nocompl = 0;
932 toep->plen_nocompl = 0;
935 tp->snd_nxt += adjusted_plen;
936 tp->snd_max += adjusted_plen;
938 toep->flags |= TPF_TX_DATA_SENT;
939 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
940 toep->flags |= TPF_TX_SUSPENDED;
942 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
944 txsd->tx_credits = credits;
946 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
948 txsd = &toep->txsd[0];
952 t4_l2t_send(sc, wr, toep->l2te);
955 /* Send a FIN if requested, but only if there are no more PDUs to send */
956 if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
957 close_conn(sc, toep);
961 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
963 struct adapter *sc = tod->tod_softc;
965 struct inpcb *inp = tp->t_inpcb;
967 struct toepcb *toep = tp->t_toe;
969 INP_WLOCK_ASSERT(inp);
970 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
971 ("%s: inp %p dropped.", __func__, inp));
972 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
974 if (toep->ulp_mode == ULP_MODE_ISCSI)
975 t4_push_pdus(sc, toep, 0);
977 t4_push_frames(sc, toep, 0);
983 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
985 struct adapter *sc = tod->tod_softc;
987 struct inpcb *inp = tp->t_inpcb;
989 struct toepcb *toep = tp->t_toe;
991 INP_WLOCK_ASSERT(inp);
992 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
993 ("%s: inp %p dropped.", __func__, inp));
994 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
996 toep->flags |= TPF_SEND_FIN;
997 if (tp->t_state >= TCPS_ESTABLISHED) {
998 if (toep->ulp_mode == ULP_MODE_ISCSI)
999 t4_push_pdus(sc, toep, 0);
1001 t4_push_frames(sc, toep, 0);
1008 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1010 struct adapter *sc = tod->tod_softc;
1011 #if defined(INVARIANTS)
1012 struct inpcb *inp = tp->t_inpcb;
1014 struct toepcb *toep = tp->t_toe;
1016 INP_WLOCK_ASSERT(inp);
1017 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1018 ("%s: inp %p dropped.", __func__, inp));
1019 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1022 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1023 ("%s: flowc for tid %u [%s] not sent already",
1024 __func__, toep->tid, tcpstates[tp->t_state]));
1026 send_reset(sc, toep, 0);
1031 * Peer has sent us a FIN.
1034 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1036 struct adapter *sc = iq->adapter;
1037 const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1038 unsigned int tid = GET_TID(cpl);
1039 struct toepcb *toep = lookup_tid(sc, tid);
1040 struct inpcb *inp = toep->inp;
1041 struct tcpcb *tp = NULL;
1045 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1048 KASSERT(opcode == CPL_PEER_CLOSE,
1049 ("%s: unexpected opcode 0x%x", __func__, opcode));
1050 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1052 if (__predict_false(toep->flags & TPF_SYNQE)) {
1054 struct synq_entry *synqe = (void *)toep;
1056 INP_WLOCK(synqe->lctx->inp);
1057 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1058 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1059 ("%s: listen socket closed but tid %u not aborted.",
1063 * do_pass_accept_req is still running and will
1064 * eventually take care of this tid.
1067 INP_WUNLOCK(synqe->lctx->inp);
1069 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1074 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1076 INP_INFO_RLOCK(&V_tcbinfo);
1078 tp = intotcpcb(inp);
1080 CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1081 tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1083 if (toep->flags & TPF_ABORT_SHUTDOWN)
1086 tp->rcv_nxt++; /* FIN */
1088 so = inp->inp_socket;
1091 if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
1092 handle_ddp_close(toep, tp, sb, cpl->rcv_nxt);
1094 socantrcvmore_locked(so); /* unlocks the sockbuf */
1096 if (toep->ulp_mode != ULP_MODE_RDMA) {
1097 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1098 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1099 be32toh(cpl->rcv_nxt)));
1102 switch (tp->t_state) {
1103 case TCPS_SYN_RECEIVED:
1104 tp->t_starttime = ticks;
1107 case TCPS_ESTABLISHED:
1108 tp->t_state = TCPS_CLOSE_WAIT;
1111 case TCPS_FIN_WAIT_1:
1112 tp->t_state = TCPS_CLOSING;
1115 case TCPS_FIN_WAIT_2:
1117 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1118 INP_INFO_RUNLOCK(&V_tcbinfo);
1121 final_cpl_received(toep);
1125 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1126 __func__, tid, tp->t_state);
1130 INP_INFO_RUNLOCK(&V_tcbinfo);
1135 * Peer has ACK'd our FIN.
1138 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1141 struct adapter *sc = iq->adapter;
1142 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1143 unsigned int tid = GET_TID(cpl);
1144 struct toepcb *toep = lookup_tid(sc, tid);
1145 struct inpcb *inp = toep->inp;
1146 struct tcpcb *tp = NULL;
1147 struct socket *so = NULL;
1149 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1152 KASSERT(opcode == CPL_CLOSE_CON_RPL,
1153 ("%s: unexpected opcode 0x%x", __func__, opcode));
1154 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1155 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1157 INP_INFO_RLOCK(&V_tcbinfo);
1159 tp = intotcpcb(inp);
1161 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1162 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1164 if (toep->flags & TPF_ABORT_SHUTDOWN)
1167 so = inp->inp_socket;
1168 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */
1170 switch (tp->t_state) {
1171 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */
1174 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1175 INP_INFO_RUNLOCK(&V_tcbinfo);
1178 final_cpl_received(toep); /* no more CPLs expected */
1186 case TCPS_FIN_WAIT_1:
1187 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1188 soisdisconnected(so);
1189 tp->t_state = TCPS_FIN_WAIT_2;
1194 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1195 __func__, tid, tcpstates[tp->t_state]);
1199 INP_INFO_RUNLOCK(&V_tcbinfo);
1204 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1208 struct cpl_abort_rpl *cpl;
1210 wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1213 panic("%s: allocation failure.", __func__);
1217 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1218 cpl->cmd = rst_status;
1224 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1226 switch (abort_reason) {
1227 case CPL_ERR_BAD_SYN:
1228 case CPL_ERR_CONN_RESET:
1229 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1230 case CPL_ERR_XMIT_TIMEDOUT:
1231 case CPL_ERR_PERSIST_TIMEDOUT:
1232 case CPL_ERR_FINWAIT2_TIMEDOUT:
1233 case CPL_ERR_KEEPALIVE_TIMEDOUT:
1241 * TCP RST from the peer, timeout, or some other such critical error.
1244 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1246 struct adapter *sc = iq->adapter;
1247 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1248 unsigned int tid = GET_TID(cpl);
1249 struct toepcb *toep = lookup_tid(sc, tid);
1250 struct sge_wrq *ofld_txq = toep->ofld_txq;
1254 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1257 KASSERT(opcode == CPL_ABORT_REQ_RSS,
1258 ("%s: unexpected opcode 0x%x", __func__, opcode));
1259 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1261 if (toep->flags & TPF_SYNQE)
1262 return (do_abort_req_synqe(iq, rss, m));
1264 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1266 if (negative_advice(cpl->status)) {
1267 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1268 __func__, cpl->status, tid, toep->flags);
1269 return (0); /* Ignore negative advice */
1273 INP_INFO_RLOCK(&V_tcbinfo); /* for tcp_close */
1276 tp = intotcpcb(inp);
1279 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1280 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1281 inp->inp_flags, cpl->status);
1284 * If we'd initiated an abort earlier the reply to it is responsible for
1285 * cleaning up resources. Otherwise we tear everything down right here
1286 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
1288 if (toep->flags & TPF_ABORT_SHUTDOWN) {
1292 toep->flags |= TPF_ABORT_SHUTDOWN;
1294 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1295 struct socket *so = inp->inp_socket;
1298 so_error_set(so, abort_status_to_errno(tp,
1302 INP_WLOCK(inp); /* re-acquire */
1305 final_cpl_received(toep);
1307 INP_INFO_RUNLOCK(&V_tcbinfo);
1308 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1313 * Reply to the CPL_ABORT_REQ (send_reset)
1316 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1318 struct adapter *sc = iq->adapter;
1319 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1320 unsigned int tid = GET_TID(cpl);
1321 struct toepcb *toep = lookup_tid(sc, tid);
1322 struct inpcb *inp = toep->inp;
1324 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1327 KASSERT(opcode == CPL_ABORT_RPL_RSS,
1328 ("%s: unexpected opcode 0x%x", __func__, opcode));
1329 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1331 if (toep->flags & TPF_SYNQE)
1332 return (do_abort_rpl_synqe(iq, rss, m));
1334 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1336 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1337 __func__, tid, toep, inp, cpl->status);
1339 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1340 ("%s: wasn't expecting abort reply", __func__));
1343 final_cpl_received(toep);
1349 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1351 struct adapter *sc = iq->adapter;
1352 const struct cpl_rx_data *cpl = mtod(m, const void *);
1353 unsigned int tid = GET_TID(cpl);
1354 struct toepcb *toep = lookup_tid(sc, tid);
1355 struct inpcb *inp = toep->inp;
1360 uint32_t ddp_placed = 0;
1362 if (__predict_false(toep->flags & TPF_SYNQE)) {
1364 struct synq_entry *synqe = (void *)toep;
1366 INP_WLOCK(synqe->lctx->inp);
1367 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1368 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1369 ("%s: listen socket closed but tid %u not aborted.",
1373 * do_pass_accept_req is still running and will
1374 * eventually take care of this tid.
1377 INP_WUNLOCK(synqe->lctx->inp);
1379 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1385 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1387 /* strip off CPL header */
1388 m_adj(m, sizeof(*cpl));
1389 len = m->m_pkthdr.len;
1392 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1393 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1394 __func__, tid, len, inp->inp_flags);
1400 tp = intotcpcb(inp);
1402 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1403 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1406 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
1408 tp->t_rcvtime = ticks;
1410 so = inp_inpcbtosocket(inp);
1414 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1415 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1416 __func__, tid, len);
1421 INP_INFO_RLOCK(&V_tcbinfo);
1423 tp = tcp_drop(tp, ECONNRESET);
1426 INP_INFO_RUNLOCK(&V_tcbinfo);
1431 /* receive buffer autosize */
1432 if (sb->sb_flags & SB_AUTOSIZE &&
1433 V_tcp_do_autorcvbuf &&
1434 sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1435 len > (sbspace(sb) / 8 * 7)) {
1436 unsigned int hiwat = sb->sb_hiwat;
1437 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1438 V_tcp_autorcvbuf_max);
1440 if (!sbreserve_locked(sb, newsize, so, NULL))
1441 sb->sb_flags &= ~SB_AUTOSIZE;
1443 toep->rx_credits += newsize - hiwat;
1446 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1447 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1450 if (toep->ddp_flags & DDP_SC_REQ)
1451 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1453 KASSERT(cpl->ddp_off == 1,
1454 ("%s: DDP switched on by itself.",
1457 /* Fell out of DDP mode */
1458 toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1462 insert_ddp_data(toep, ddp_placed);
1466 if ((toep->ddp_flags & DDP_OK) == 0 &&
1467 time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1468 toep->ddp_score = DDP_LOW_SCORE;
1469 toep->ddp_flags |= DDP_OK;
1470 CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1471 __func__, tid, time_uptime);
1474 if (toep->ddp_flags & DDP_ON) {
1477 * CPL_RX_DATA with DDP on can only be an indicate. Ask
1478 * soreceive to post a buffer or disable DDP. The
1479 * payload that arrived in this indicate is appended to
1480 * the socket buffer as usual.
1485 "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1486 __func__, tid, toep->flags, be32toh(cpl->seq), len);
1488 sb->sb_flags |= SB_DDP_INDICATE;
1489 } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1490 tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1493 * DDP allowed but isn't on (and a request to switch it
1494 * on isn't pending either), and conditions are ripe for
1495 * it to work. Switch it on.
1498 enable_ddp(sc, toep);
1502 KASSERT(toep->sb_cc >= sbused(sb),
1503 ("%s: sb %p has more data (%d) than last time (%d).",
1504 __func__, sb, sbused(sb), toep->sb_cc));
1505 toep->rx_credits += toep->sb_cc - sbused(sb);
1506 sbappendstream_locked(sb, m, 0);
1507 toep->sb_cc = sbused(sb);
1508 if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1511 credits = send_rx_credits(sc, toep, toep->rx_credits);
1512 toep->rx_credits -= credits;
1513 tp->rcv_wnd += credits;
1514 tp->rcv_adv += credits;
1516 sorwakeup_locked(so);
1517 SOCKBUF_UNLOCK_ASSERT(sb);
1523 #define S_CPL_FW4_ACK_OPCODE 24
1524 #define M_CPL_FW4_ACK_OPCODE 0xff
1525 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1526 #define G_CPL_FW4_ACK_OPCODE(x) \
1527 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1529 #define S_CPL_FW4_ACK_FLOWID 0
1530 #define M_CPL_FW4_ACK_FLOWID 0xffffff
1531 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1532 #define G_CPL_FW4_ACK_FLOWID(x) \
1533 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1535 #define S_CPL_FW4_ACK_CR 24
1536 #define M_CPL_FW4_ACK_CR 0xff
1537 #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR)
1538 #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1540 #define S_CPL_FW4_ACK_SEQVAL 0
1541 #define M_CPL_FW4_ACK_SEQVAL 0x1
1542 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1543 #define G_CPL_FW4_ACK_SEQVAL(x) \
1544 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1545 #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U)
1548 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1550 struct adapter *sc = iq->adapter;
1551 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1552 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1553 struct toepcb *toep = lookup_tid(sc, tid);
1557 uint8_t credits = cpl->credits;
1558 struct ofld_tx_sdesc *txsd;
1561 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1565 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1566 * now this comes back carrying the credits for the flowc.
1568 if (__predict_false(toep->flags & TPF_SYNQE)) {
1569 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1570 ("%s: credits for a synq entry %p", __func__, toep));
1576 KASSERT(opcode == CPL_FW4_ACK,
1577 ("%s: unexpected opcode 0x%x", __func__, opcode));
1578 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1579 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1583 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1588 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1589 ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1591 tp = intotcpcb(inp);
1593 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1594 tcp_seq snd_una = be32toh(cpl->snd_una);
1597 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1599 "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1600 __func__, snd_una, toep->tid, tp->snd_una);
1604 if (tp->snd_una != snd_una) {
1605 tp->snd_una = snd_una;
1606 tp->ts_recent_age = tcp_ts_getticks();
1610 so = inp->inp_socket;
1611 txsd = &toep->txsd[toep->txsd_cidx];
1614 KASSERT(credits >= txsd->tx_credits,
1615 ("%s: too many (or partial) credits", __func__));
1616 credits -= txsd->tx_credits;
1617 toep->tx_credits += txsd->tx_credits;
1621 KASSERT(toep->txsd_avail <= toep->txsd_total,
1622 ("%s: txsd avail > total", __func__));
1623 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1624 txsd = &toep->txsd[0];
1625 toep->txsd_cidx = 0;
1629 if (toep->tx_credits == toep->tx_total) {
1630 toep->tx_nocompl = 0;
1631 toep->plen_nocompl = 0;
1634 if (toep->flags & TPF_TX_SUSPENDED &&
1635 toep->tx_credits >= toep->tx_total / 4) {
1636 toep->flags &= ~TPF_TX_SUSPENDED;
1637 if (toep->ulp_mode == ULP_MODE_ISCSI)
1638 t4_push_pdus(sc, toep, plen);
1640 t4_push_frames(sc, toep, plen);
1641 } else if (plen > 0) {
1642 struct sockbuf *sb = &so->so_snd;
1647 if (toep->ulp_mode == ULP_MODE_ISCSI) {
1649 if (__predict_false(sbu > 0)) {
1651 * The data trasmitted before the tid's ULP mode
1652 * changed to ISCSI is still in so_snd.
1653 * Incoming credits should account for so_snd
1656 sbdrop_locked(sb, min(sbu, plen));
1657 plen -= min(sbu, plen);
1659 sowwakeup_locked(so); /* unlocks so_snd */
1660 rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1662 sbdrop_locked(sb, plen);
1663 sowwakeup_locked(so); /* unlocks so_snd */
1665 SOCKBUF_UNLOCK_ASSERT(sb);
1674 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1676 struct adapter *sc = iq->adapter;
1677 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1678 unsigned int tid = GET_TID(cpl);
1680 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1683 KASSERT(opcode == CPL_SET_TCB_RPL,
1684 ("%s: unexpected opcode 0x%x", __func__, opcode));
1685 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1687 if (is_ftid(sc, tid))
1688 return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1691 * TOM and/or other ULPs don't request replies for CPL_SET_TCB or
1692 * CPL_SET_TCB_FIELD requests. This can easily change and when it does
1693 * the dispatch code will go here.
1696 panic("%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p", __func__,
1699 log(LOG_ERR, "%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p\n",
1707 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1708 uint16_t word, uint64_t mask, uint64_t val)
1711 struct cpl_set_tcb_field *req;
1713 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1716 panic("%s: allocation failure.", __func__);
1720 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1721 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1722 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1723 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1724 req->mask = htobe64(mask);
1725 req->val = htobe64(val);
1731 t4_init_cpl_io_handlers(struct adapter *sc)
1734 t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1735 t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1736 t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1737 t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1738 t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1739 t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1740 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1744 t4_uninit_cpl_io_handlers(struct adapter *sc)
1747 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);