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 CURVNET_SET(so->so_vnet);
271 INP_WLOCK_ASSERT(inp);
272 KASSERT(tp->t_state == TCPS_SYN_SENT ||
273 tp->t_state == TCPS_SYN_RECEIVED,
274 ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
276 CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p",
277 __func__, toep->tid, toep, inp);
279 tp->t_state = TCPS_ESTABLISHED;
280 tp->t_starttime = ticks;
281 TCPSTAT_INC(tcps_connects);
285 tp->rcv_wnd = toep->rx_credits << 10;
286 tp->rcv_adv += tp->rcv_wnd;
287 tp->last_ack_sent = tp->rcv_nxt;
290 * If we were unable to send all rx credits via opt0, save the remainder
291 * in rx_credits so that they can be handed over with the next credit
294 SOCKBUF_LOCK(&so->so_rcv);
295 bufsize = select_rcv_wnd(so);
296 SOCKBUF_UNLOCK(&so->so_rcv);
297 toep->rx_credits = bufsize - tp->rcv_wnd;
301 tp->snd_una = iss + 1;
302 tp->snd_nxt = iss + 1;
303 tp->snd_max = iss + 1;
305 assign_rxopt(tp, tcpopt);
307 SOCKBUF_LOCK(&so->so_snd);
308 if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf)
309 bufsize = V_tcp_autosndbuf_max;
311 bufsize = sbspace(&so->so_snd);
312 SOCKBUF_UNLOCK(&so->so_snd);
314 ftxp.snd_nxt = tp->snd_nxt;
315 ftxp.rcv_nxt = tp->rcv_nxt;
316 ftxp.snd_space = bufsize;
317 ftxp.mss = tp->t_maxseg;
318 send_flowc_wr(toep, &ftxp);
325 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
328 struct cpl_rx_data_ack *req;
329 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
331 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
333 wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
338 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
339 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
346 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp)
348 struct adapter *sc = tod->tod_softc;
349 struct inpcb *inp = tp->t_inpcb;
350 struct socket *so = inp->inp_socket;
351 struct sockbuf *sb = &so->so_rcv;
352 struct toepcb *toep = tp->t_toe;
355 INP_WLOCK_ASSERT(inp);
357 SOCKBUF_LOCK_ASSERT(sb);
358 KASSERT(toep->sb_cc >= sbused(sb),
359 ("%s: sb %p has more data (%d) than last time (%d).",
360 __func__, sb, sbused(sb), toep->sb_cc));
362 toep->rx_credits += toep->sb_cc - sbused(sb);
363 toep->sb_cc = sbused(sb);
365 if (toep->rx_credits > 0 &&
366 (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
367 (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
368 toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
370 credits = send_rx_credits(sc, toep, toep->rx_credits);
371 toep->rx_credits -= credits;
372 tp->rcv_wnd += credits;
373 tp->rcv_adv += credits;
378 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
380 struct inpcb *inp = tp->t_inpcb;
381 struct socket *so = inp->inp_socket;
382 struct sockbuf *sb = &so->so_rcv;
385 t4_rcvd_locked(tod, tp);
390 * Close a connection by sending a CPL_CLOSE_CON_REQ message.
393 close_conn(struct adapter *sc, struct toepcb *toep)
396 struct cpl_close_con_req *req;
397 unsigned int tid = toep->tid;
399 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
400 toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
402 if (toep->flags & TPF_FIN_SENT)
405 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
406 ("%s: flowc_wr not sent for tid %u.", __func__, tid));
408 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
411 panic("%s: allocation failure.", __func__);
415 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
416 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
417 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
418 V_FW_WR_FLOWID(tid));
419 req->wr.wr_lo = cpu_to_be64(0);
420 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
423 toep->flags |= TPF_FIN_SENT;
424 toep->flags &= ~TPF_SEND_FIN;
425 t4_l2t_send(sc, wr, toep->l2te);
430 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
431 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
433 /* Maximum amount of immediate data we could stuff in a WR */
435 max_imm_payload(int tx_credits)
437 const int n = 2; /* Use only up to 2 desc for imm. data WR */
439 KASSERT(tx_credits >= 0 &&
440 tx_credits <= MAX_OFLD_TX_CREDITS,
441 ("%s: %d credits", __func__, tx_credits));
443 if (tx_credits < MIN_OFLD_TX_CREDITS)
446 if (tx_credits >= (n * EQ_ESIZE) / 16)
447 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
449 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
452 /* Maximum number of SGL entries we could stuff in a WR */
454 max_dsgl_nsegs(int tx_credits)
456 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
457 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
459 KASSERT(tx_credits >= 0 &&
460 tx_credits <= MAX_OFLD_TX_CREDITS,
461 ("%s: %d credits", __func__, tx_credits));
463 if (tx_credits < MIN_OFLD_TX_CREDITS)
466 nseg += 2 * (sge_pair_credits * 16 / 24);
467 if ((sge_pair_credits * 16) % 24 == 16)
474 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
475 unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign)
477 struct fw_ofld_tx_data_wr *txwr = dst;
479 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
480 V_FW_WR_IMMDLEN(immdlen));
481 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
482 V_FW_WR_LEN16(credits));
483 txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) |
484 V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
485 txwr->plen = htobe32(plen);
488 struct tcpcb *tp = intotcpcb(toep->inp);
490 if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi))
491 txwr->lsodisable_to_flags |=
492 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
494 txwr->lsodisable_to_flags |=
495 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
496 (tp->t_flags & TF_NODELAY ? 0 :
497 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
502 * Generate a DSGL from a starting mbuf. The total number of segments and the
503 * maximum segments in any one mbuf are provided.
506 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
509 struct ulptx_sgl *usgl = dst;
512 struct sglist_seg segs[n];
514 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
516 sglist_init(&sg, n, segs);
517 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
518 V_ULPTX_NSGE(nsegs));
521 for (m = start; m != stop; m = m->m_next) {
522 rc = sglist_append(&sg, mtod(m, void *), m->m_len);
523 if (__predict_false(rc != 0))
524 panic("%s: sglist_append %d", __func__, rc);
526 for (j = 0; j < sg.sg_nseg; i++, j++) {
528 usgl->len0 = htobe32(segs[j].ss_len);
529 usgl->addr0 = htobe64(segs[j].ss_paddr);
531 usgl->sge[i / 2].len[i & 1] =
532 htobe32(segs[j].ss_len);
533 usgl->sge[i / 2].addr[i & 1] =
534 htobe64(segs[j].ss_paddr);
543 usgl->sge[i / 2].len[1] = htobe32(0);
544 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
545 __func__, nsegs, start, stop));
549 * Max number of SGL entries an offload tx work request can have. This is 41
550 * (1 + 40) for a full 512B work request.
551 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
553 #define OFLD_SGL_LEN (41)
556 * Send data and/or a FIN to the peer.
558 * The socket's so_snd buffer consists of a stream of data starting with sb_mb
559 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that
562 * drop indicates the number of bytes that should be dropped from the head of
563 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating
564 * contention on the send buffer lock (before this change it used to do
565 * sowwakeup and then t4_push_frames right after that when recovering from tx
566 * stalls). When drop is set this function MUST drop the bytes and wake up any
570 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
572 struct mbuf *sndptr, *m, *sb_sndptr;
573 struct fw_ofld_tx_data_wr *txwr;
575 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
576 struct inpcb *inp = toep->inp;
577 struct tcpcb *tp = intotcpcb(inp);
578 struct socket *so = inp->inp_socket;
579 struct sockbuf *sb = &so->so_snd;
580 int tx_credits, shove, compl, space, sowwakeup;
581 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
583 INP_WLOCK_ASSERT(inp);
584 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
585 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
587 KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
588 toep->ulp_mode == ULP_MODE_TCPDDP ||
589 toep->ulp_mode == ULP_MODE_RDMA,
590 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
592 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
596 * This function doesn't resume by itself. Someone else must clear the
597 * flag and call this function.
599 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
601 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
606 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
607 max_imm = max_imm_payload(tx_credits);
608 max_nsegs = max_dsgl_nsegs(tx_credits);
613 sbdrop_locked(sb, drop);
616 sb_sndptr = sb->sb_sndptr;
617 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
620 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
621 for (m = sndptr; m != NULL; m = m->m_next) {
622 int n = sglist_count(mtod(m, void *), m->m_len);
627 /* This mbuf sent us _over_ the nsegs limit, back out */
628 if (plen > max_imm && nsegs > max_nsegs) {
632 /* Too few credits */
633 toep->flags |= TPF_TX_SUSPENDED;
635 sowwakeup_locked(so);
638 SOCKBUF_UNLOCK_ASSERT(sb);
644 if (max_nsegs_1mbuf < n)
646 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */
648 /* This mbuf put us right at the max_nsegs limit */
649 if (plen > max_imm && nsegs == max_nsegs) {
657 if (space <= sb->sb_hiwat * 3 / 8 &&
658 toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
663 if (sb->sb_flags & SB_AUTOSIZE &&
664 V_tcp_do_autosndbuf &&
665 sb->sb_hiwat < V_tcp_autosndbuf_max &&
666 space < sb->sb_hiwat / 8) {
667 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
668 V_tcp_autosndbuf_max);
670 if (!sbreserve_locked(sb, newsize, so, NULL))
671 sb->sb_flags &= ~SB_AUTOSIZE;
673 sowwakeup = 1; /* room available */
676 sowwakeup_locked(so);
679 SOCKBUF_UNLOCK_ASSERT(sb);
681 /* nothing to send */
684 ("%s: nothing to send, but m != NULL", __func__));
688 if (__predict_false(toep->flags & TPF_FIN_SENT))
689 panic("%s: excess tx.", __func__);
691 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
692 if (plen <= max_imm) {
694 /* Immediate data tx */
696 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
699 /* XXX: how will we recover from this? */
700 toep->flags |= TPF_TX_SUSPENDED;
704 credits = howmany(wr->wr_len, 16);
705 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
707 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
714 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
715 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
716 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
718 /* XXX: how will we recover from this? */
719 toep->flags |= TPF_TX_SUSPENDED;
723 credits = howmany(wr_len, 16);
724 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
726 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
729 uint64_t *pad = (uint64_t *)
730 ((uintptr_t)txwr + wr_len);
735 KASSERT(toep->tx_credits >= credits,
736 ("%s: not enough credits", __func__));
738 toep->tx_credits -= credits;
739 toep->tx_nocompl += credits;
740 toep->plen_nocompl += plen;
741 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
742 toep->tx_nocompl >= toep->tx_total / 4)
745 if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
746 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
747 toep->tx_nocompl = 0;
748 toep->plen_nocompl = 0;
755 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
756 sb->sb_sndptr = sb_sndptr;
759 toep->flags |= TPF_TX_DATA_SENT;
760 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
761 toep->flags |= TPF_TX_SUSPENDED;
763 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
765 txsd->tx_credits = credits;
767 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
769 txsd = &toep->txsd[0];
773 t4_l2t_send(sc, wr, toep->l2te);
776 /* Send a FIN if requested, but only if there's no more data to send */
777 if (m == NULL && toep->flags & TPF_SEND_FIN)
778 close_conn(sc, toep);
782 rqdrop_locked(struct mbufq *q, int plen)
787 m = mbufq_dequeue(q);
789 /* Too many credits. */
793 /* Partial credits. */
794 MPASS(plen >= m->m_pkthdr.len);
796 plen -= m->m_pkthdr.len;
802 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
804 struct mbuf *sndptr, *m;
805 struct fw_ofld_tx_data_wr *txwr;
807 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
808 u_int adjusted_plen, ulp_submode;
809 struct inpcb *inp = toep->inp;
810 struct tcpcb *tp = intotcpcb(inp);
811 int tx_credits, shove;
812 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
813 struct mbufq *pduq = &toep->ulp_pduq;
814 static const u_int ulp_extra_len[] = {0, 4, 4, 8};
816 INP_WLOCK_ASSERT(inp);
817 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
818 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
819 KASSERT(toep->ulp_mode == ULP_MODE_ISCSI,
820 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
822 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
826 * This function doesn't resume by itself. Someone else must clear the
827 * flag and call this function.
829 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
831 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
836 rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
838 while ((sndptr = mbufq_first(pduq)) != NULL) {
839 M_ASSERTPKTHDR(sndptr);
841 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
842 max_imm = max_imm_payload(tx_credits);
843 max_nsegs = max_dsgl_nsegs(tx_credits);
847 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
848 for (m = sndptr; m != NULL; m = m->m_next) {
849 int n = sglist_count(mtod(m, void *), m->m_len);
855 * This mbuf would send us _over_ the nsegs limit.
856 * Suspend tx because the PDU can't be sent out.
858 if (plen > max_imm && nsegs > max_nsegs) {
859 toep->flags |= TPF_TX_SUSPENDED;
863 if (max_nsegs_1mbuf < n)
867 if (__predict_false(toep->flags & TPF_FIN_SENT))
868 panic("%s: excess tx.", __func__);
871 * We have a PDU to send. All of it goes out in one WR so 'm'
872 * is NULL. A PDU's length is always a multiple of 4.
875 MPASS((plen & 3) == 0);
876 MPASS(sndptr->m_pkthdr.len == plen);
878 shove = !(tp->t_flags & TF_MORETOCOME);
879 ulp_submode = mbuf_ulp_submode(sndptr);
880 MPASS(ulp_submode < nitems(ulp_extra_len));
883 * plen doesn't include header and data digests, which are
884 * generated and inserted in the right places by the TOE, but
885 * they do occupy TCP sequence space and need to be accounted
888 adjusted_plen = plen + ulp_extra_len[ulp_submode];
889 if (plen <= max_imm) {
891 /* Immediate data tx */
893 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
896 /* XXX: how will we recover from this? */
897 toep->flags |= TPF_TX_SUSPENDED;
901 credits = howmany(wr->wr_len, 16);
902 write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
903 shove, ulp_submode, sc->tt.tx_align);
904 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
910 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
911 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
912 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
914 /* XXX: how will we recover from this? */
915 toep->flags |= TPF_TX_SUSPENDED;
919 credits = howmany(wr_len, 16);
920 write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
921 shove, ulp_submode, sc->tt.tx_align);
922 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
925 uint64_t *pad = (uint64_t *)
926 ((uintptr_t)txwr + wr_len);
931 KASSERT(toep->tx_credits >= credits,
932 ("%s: not enough credits", __func__));
934 m = mbufq_dequeue(pduq);
936 mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
938 toep->tx_credits -= credits;
939 toep->tx_nocompl += credits;
940 toep->plen_nocompl += plen;
941 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
942 toep->tx_nocompl >= toep->tx_total / 4) {
943 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
944 toep->tx_nocompl = 0;
945 toep->plen_nocompl = 0;
948 tp->snd_nxt += adjusted_plen;
949 tp->snd_max += adjusted_plen;
951 toep->flags |= TPF_TX_DATA_SENT;
952 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
953 toep->flags |= TPF_TX_SUSPENDED;
955 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
957 txsd->tx_credits = credits;
959 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
961 txsd = &toep->txsd[0];
965 t4_l2t_send(sc, wr, toep->l2te);
968 /* Send a FIN if requested, but only if there are no more PDUs to send */
969 if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
970 close_conn(sc, toep);
974 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
976 struct adapter *sc = tod->tod_softc;
978 struct inpcb *inp = tp->t_inpcb;
980 struct toepcb *toep = tp->t_toe;
982 INP_WLOCK_ASSERT(inp);
983 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
984 ("%s: inp %p dropped.", __func__, inp));
985 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
987 if (toep->ulp_mode == ULP_MODE_ISCSI)
988 t4_push_pdus(sc, toep, 0);
990 t4_push_frames(sc, toep, 0);
996 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
998 struct adapter *sc = tod->tod_softc;
1000 struct inpcb *inp = tp->t_inpcb;
1002 struct toepcb *toep = tp->t_toe;
1004 INP_WLOCK_ASSERT(inp);
1005 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1006 ("%s: inp %p dropped.", __func__, inp));
1007 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1009 toep->flags |= TPF_SEND_FIN;
1010 if (tp->t_state >= TCPS_ESTABLISHED) {
1011 if (toep->ulp_mode == ULP_MODE_ISCSI)
1012 t4_push_pdus(sc, toep, 0);
1014 t4_push_frames(sc, toep, 0);
1021 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1023 struct adapter *sc = tod->tod_softc;
1024 #if defined(INVARIANTS)
1025 struct inpcb *inp = tp->t_inpcb;
1027 struct toepcb *toep = tp->t_toe;
1029 INP_WLOCK_ASSERT(inp);
1030 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1031 ("%s: inp %p dropped.", __func__, inp));
1032 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1035 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1036 ("%s: flowc for tid %u [%s] not sent already",
1037 __func__, toep->tid, tcpstates[tp->t_state]));
1039 send_reset(sc, toep, 0);
1044 * Peer has sent us a FIN.
1047 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1049 struct adapter *sc = iq->adapter;
1050 const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1051 unsigned int tid = GET_TID(cpl);
1052 struct toepcb *toep = lookup_tid(sc, tid);
1053 struct inpcb *inp = toep->inp;
1054 struct tcpcb *tp = NULL;
1057 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1060 KASSERT(opcode == CPL_PEER_CLOSE,
1061 ("%s: unexpected opcode 0x%x", __func__, opcode));
1062 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1064 if (__predict_false(toep->flags & TPF_SYNQE)) {
1066 struct synq_entry *synqe = (void *)toep;
1068 INP_WLOCK(synqe->lctx->inp);
1069 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1070 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1071 ("%s: listen socket closed but tid %u not aborted.",
1075 * do_pass_accept_req is still running and will
1076 * eventually take care of this tid.
1079 INP_WUNLOCK(synqe->lctx->inp);
1081 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1086 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1088 INP_INFO_RLOCK(&V_tcbinfo);
1090 tp = intotcpcb(inp);
1092 CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1093 tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1095 if (toep->flags & TPF_ABORT_SHUTDOWN)
1098 tp->rcv_nxt++; /* FIN */
1100 so = inp->inp_socket;
1101 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1103 if (__predict_false(toep->ddp_flags &
1104 (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)))
1105 handle_ddp_close(toep, tp, cpl->rcv_nxt);
1110 if (toep->ulp_mode != ULP_MODE_RDMA) {
1111 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1112 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1113 be32toh(cpl->rcv_nxt)));
1116 switch (tp->t_state) {
1117 case TCPS_SYN_RECEIVED:
1118 tp->t_starttime = ticks;
1121 case TCPS_ESTABLISHED:
1122 tp->t_state = TCPS_CLOSE_WAIT;
1125 case TCPS_FIN_WAIT_1:
1126 tp->t_state = TCPS_CLOSING;
1129 case TCPS_FIN_WAIT_2:
1131 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1132 INP_INFO_RUNLOCK(&V_tcbinfo);
1135 final_cpl_received(toep);
1139 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1140 __func__, tid, tp->t_state);
1144 INP_INFO_RUNLOCK(&V_tcbinfo);
1149 * Peer has ACK'd our FIN.
1152 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1155 struct adapter *sc = iq->adapter;
1156 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1157 unsigned int tid = GET_TID(cpl);
1158 struct toepcb *toep = lookup_tid(sc, tid);
1159 struct inpcb *inp = toep->inp;
1160 struct tcpcb *tp = NULL;
1161 struct socket *so = NULL;
1163 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1166 KASSERT(opcode == CPL_CLOSE_CON_RPL,
1167 ("%s: unexpected opcode 0x%x", __func__, opcode));
1168 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1169 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1171 INP_INFO_RLOCK(&V_tcbinfo);
1173 tp = intotcpcb(inp);
1175 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1176 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1178 if (toep->flags & TPF_ABORT_SHUTDOWN)
1181 so = inp->inp_socket;
1182 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */
1184 switch (tp->t_state) {
1185 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */
1188 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1189 INP_INFO_RUNLOCK(&V_tcbinfo);
1192 final_cpl_received(toep); /* no more CPLs expected */
1200 case TCPS_FIN_WAIT_1:
1201 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1202 soisdisconnected(so);
1203 tp->t_state = TCPS_FIN_WAIT_2;
1208 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1209 __func__, tid, tcpstates[tp->t_state]);
1213 INP_INFO_RUNLOCK(&V_tcbinfo);
1218 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1222 struct cpl_abort_rpl *cpl;
1224 wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1227 panic("%s: allocation failure.", __func__);
1231 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1232 cpl->cmd = rst_status;
1238 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1240 switch (abort_reason) {
1241 case CPL_ERR_BAD_SYN:
1242 case CPL_ERR_CONN_RESET:
1243 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1244 case CPL_ERR_XMIT_TIMEDOUT:
1245 case CPL_ERR_PERSIST_TIMEDOUT:
1246 case CPL_ERR_FINWAIT2_TIMEDOUT:
1247 case CPL_ERR_KEEPALIVE_TIMEDOUT:
1255 * TCP RST from the peer, timeout, or some other such critical error.
1258 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1260 struct adapter *sc = iq->adapter;
1261 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1262 unsigned int tid = GET_TID(cpl);
1263 struct toepcb *toep = lookup_tid(sc, tid);
1264 struct sge_wrq *ofld_txq = toep->ofld_txq;
1268 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1271 KASSERT(opcode == CPL_ABORT_REQ_RSS,
1272 ("%s: unexpected opcode 0x%x", __func__, opcode));
1273 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1275 if (toep->flags & TPF_SYNQE)
1276 return (do_abort_req_synqe(iq, rss, m));
1278 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1280 if (negative_advice(cpl->status)) {
1281 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1282 __func__, cpl->status, tid, toep->flags);
1283 return (0); /* Ignore negative advice */
1287 INP_INFO_RLOCK(&V_tcbinfo); /* for tcp_close */
1290 tp = intotcpcb(inp);
1293 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1294 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1295 inp->inp_flags, cpl->status);
1298 * If we'd initiated an abort earlier the reply to it is responsible for
1299 * cleaning up resources. Otherwise we tear everything down right here
1300 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
1302 if (toep->flags & TPF_ABORT_SHUTDOWN) {
1306 toep->flags |= TPF_ABORT_SHUTDOWN;
1308 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1309 struct socket *so = inp->inp_socket;
1312 so_error_set(so, abort_status_to_errno(tp,
1316 INP_WLOCK(inp); /* re-acquire */
1319 final_cpl_received(toep);
1321 INP_INFO_RUNLOCK(&V_tcbinfo);
1322 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1327 * Reply to the CPL_ABORT_REQ (send_reset)
1330 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1332 struct adapter *sc = iq->adapter;
1333 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1334 unsigned int tid = GET_TID(cpl);
1335 struct toepcb *toep = lookup_tid(sc, tid);
1336 struct inpcb *inp = toep->inp;
1338 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1341 KASSERT(opcode == CPL_ABORT_RPL_RSS,
1342 ("%s: unexpected opcode 0x%x", __func__, opcode));
1343 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1345 if (toep->flags & TPF_SYNQE)
1346 return (do_abort_rpl_synqe(iq, rss, m));
1348 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1350 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1351 __func__, tid, toep, inp, cpl->status);
1353 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1354 ("%s: wasn't expecting abort reply", __func__));
1357 final_cpl_received(toep);
1363 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1365 struct adapter *sc = iq->adapter;
1366 const struct cpl_rx_data *cpl = mtod(m, const void *);
1367 unsigned int tid = GET_TID(cpl);
1368 struct toepcb *toep = lookup_tid(sc, tid);
1369 struct inpcb *inp = toep->inp;
1374 uint32_t ddp_placed = 0;
1376 if (__predict_false(toep->flags & TPF_SYNQE)) {
1378 struct synq_entry *synqe = (void *)toep;
1380 INP_WLOCK(synqe->lctx->inp);
1381 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1382 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1383 ("%s: listen socket closed but tid %u not aborted.",
1387 * do_pass_accept_req is still running and will
1388 * eventually take care of this tid.
1391 INP_WUNLOCK(synqe->lctx->inp);
1393 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1399 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1401 /* strip off CPL header */
1402 m_adj(m, sizeof(*cpl));
1403 len = m->m_pkthdr.len;
1406 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1407 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1408 __func__, tid, len, inp->inp_flags);
1414 tp = intotcpcb(inp);
1416 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1417 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1420 if (tp->rcv_wnd < len) {
1421 KASSERT(toep->ulp_mode != ULP_MODE_RDMA,
1422 ("%s: negative window size", __func__));
1426 tp->t_rcvtime = ticks;
1428 if (toep->ulp_mode == ULP_MODE_TCPDDP)
1430 so = inp_inpcbtosocket(inp);
1434 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1435 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1436 __func__, tid, len);
1439 if (toep->ulp_mode == ULP_MODE_TCPDDP)
1443 INP_INFO_RLOCK(&V_tcbinfo);
1445 tp = tcp_drop(tp, ECONNRESET);
1448 INP_INFO_RUNLOCK(&V_tcbinfo);
1453 /* receive buffer autosize */
1454 CURVNET_SET(so->so_vnet);
1455 if (sb->sb_flags & SB_AUTOSIZE &&
1456 V_tcp_do_autorcvbuf &&
1457 sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1458 len > (sbspace(sb) / 8 * 7)) {
1459 unsigned int hiwat = sb->sb_hiwat;
1460 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1461 V_tcp_autorcvbuf_max);
1463 if (!sbreserve_locked(sb, newsize, so, NULL))
1464 sb->sb_flags &= ~SB_AUTOSIZE;
1466 toep->rx_credits += newsize - hiwat;
1469 if (toep->ddp_waiting_count != 0 || toep->ddp_active_count != 0)
1470 CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)", __func__,
1473 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1474 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1477 if (toep->ddp_flags & DDP_SC_REQ)
1478 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1480 KASSERT(cpl->ddp_off == 1,
1481 ("%s: DDP switched on by itself.",
1484 /* Fell out of DDP mode */
1485 toep->ddp_flags &= ~DDP_ON;
1486 CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
1489 insert_ddp_data(toep, ddp_placed);
1493 if (toep->ddp_flags & DDP_ON) {
1495 * CPL_RX_DATA with DDP on can only be an indicate.
1496 * Start posting queued AIO requests via DDP. The
1497 * payload that arrived in this indicate is appended
1498 * to the socket buffer as usual.
1500 handle_ddp_indicate(toep);
1504 KASSERT(toep->sb_cc >= sbused(sb),
1505 ("%s: sb %p has more data (%d) than last time (%d).",
1506 __func__, sb, sbused(sb), toep->sb_cc));
1507 toep->rx_credits += toep->sb_cc - sbused(sb);
1508 sbappendstream_locked(sb, m, 0);
1509 toep->sb_cc = sbused(sb);
1510 if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1513 credits = send_rx_credits(sc, toep, toep->rx_credits);
1514 toep->rx_credits -= credits;
1515 tp->rcv_wnd += credits;
1516 tp->rcv_adv += credits;
1519 if (toep->ddp_waiting_count > 0 && sbavail(sb) != 0) {
1520 CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__,
1522 ddp_queue_toep(toep);
1524 sorwakeup_locked(so);
1525 SOCKBUF_UNLOCK_ASSERT(sb);
1526 if (toep->ulp_mode == ULP_MODE_TCPDDP)
1534 #define S_CPL_FW4_ACK_OPCODE 24
1535 #define M_CPL_FW4_ACK_OPCODE 0xff
1536 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1537 #define G_CPL_FW4_ACK_OPCODE(x) \
1538 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1540 #define S_CPL_FW4_ACK_FLOWID 0
1541 #define M_CPL_FW4_ACK_FLOWID 0xffffff
1542 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1543 #define G_CPL_FW4_ACK_FLOWID(x) \
1544 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1546 #define S_CPL_FW4_ACK_CR 24
1547 #define M_CPL_FW4_ACK_CR 0xff
1548 #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR)
1549 #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1551 #define S_CPL_FW4_ACK_SEQVAL 0
1552 #define M_CPL_FW4_ACK_SEQVAL 0x1
1553 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1554 #define G_CPL_FW4_ACK_SEQVAL(x) \
1555 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1556 #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U)
1559 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1561 struct adapter *sc = iq->adapter;
1562 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1563 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1564 struct toepcb *toep = lookup_tid(sc, tid);
1568 uint8_t credits = cpl->credits;
1569 struct ofld_tx_sdesc *txsd;
1572 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1576 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1577 * now this comes back carrying the credits for the flowc.
1579 if (__predict_false(toep->flags & TPF_SYNQE)) {
1580 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1581 ("%s: credits for a synq entry %p", __func__, toep));
1587 KASSERT(opcode == CPL_FW4_ACK,
1588 ("%s: unexpected opcode 0x%x", __func__, opcode));
1589 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1590 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1594 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1599 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1600 ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1602 tp = intotcpcb(inp);
1604 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1605 tcp_seq snd_una = be32toh(cpl->snd_una);
1608 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1610 "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1611 __func__, snd_una, toep->tid, tp->snd_una);
1615 if (tp->snd_una != snd_una) {
1616 tp->snd_una = snd_una;
1617 tp->ts_recent_age = tcp_ts_getticks();
1621 so = inp->inp_socket;
1622 txsd = &toep->txsd[toep->txsd_cidx];
1625 KASSERT(credits >= txsd->tx_credits,
1626 ("%s: too many (or partial) credits", __func__));
1627 credits -= txsd->tx_credits;
1628 toep->tx_credits += txsd->tx_credits;
1632 KASSERT(toep->txsd_avail <= toep->txsd_total,
1633 ("%s: txsd avail > total", __func__));
1634 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1635 txsd = &toep->txsd[0];
1636 toep->txsd_cidx = 0;
1640 if (toep->tx_credits == toep->tx_total) {
1641 toep->tx_nocompl = 0;
1642 toep->plen_nocompl = 0;
1645 if (toep->flags & TPF_TX_SUSPENDED &&
1646 toep->tx_credits >= toep->tx_total / 4) {
1647 toep->flags &= ~TPF_TX_SUSPENDED;
1648 if (toep->ulp_mode == ULP_MODE_ISCSI)
1649 t4_push_pdus(sc, toep, plen);
1651 t4_push_frames(sc, toep, plen);
1652 } else if (plen > 0) {
1653 struct sockbuf *sb = &so->so_snd;
1658 if (toep->ulp_mode == ULP_MODE_ISCSI) {
1660 if (__predict_false(sbu > 0)) {
1662 * The data trasmitted before the tid's ULP mode
1663 * changed to ISCSI is still in so_snd.
1664 * Incoming credits should account for so_snd
1667 sbdrop_locked(sb, min(sbu, plen));
1668 plen -= min(sbu, plen);
1670 sowwakeup_locked(so); /* unlocks so_snd */
1671 rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1673 sbdrop_locked(sb, plen);
1674 sowwakeup_locked(so); /* unlocks so_snd */
1676 SOCKBUF_UNLOCK_ASSERT(sb);
1685 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1687 struct adapter *sc = iq->adapter;
1688 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1689 unsigned int tid = GET_TID(cpl);
1690 struct toepcb *toep;
1692 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1695 KASSERT(opcode == CPL_SET_TCB_RPL,
1696 ("%s: unexpected opcode 0x%x", __func__, opcode));
1697 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1699 if (is_ftid(sc, tid))
1700 return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1702 toep = lookup_tid(sc, tid);
1703 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1704 handle_ddp_tcb_rpl(toep, cpl);
1709 * TOM and/or other ULPs don't request replies for CPL_SET_TCB or
1710 * CPL_SET_TCB_FIELD requests. This can easily change and when it does
1711 * the dispatch code will go here.
1714 panic("%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p", __func__,
1717 log(LOG_ERR, "%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p\n",
1725 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1726 uint16_t word, uint64_t mask, uint64_t val)
1729 struct cpl_set_tcb_field *req;
1731 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1734 panic("%s: allocation failure.", __func__);
1738 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1739 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1740 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1741 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1742 req->mask = htobe64(mask);
1743 req->val = htobe64(val);
1749 t4_set_tcb_field_rpl(struct adapter *sc, struct toepcb *toep, int ctrl,
1750 uint16_t word, uint64_t mask, uint64_t val, uint8_t cookie)
1753 struct cpl_set_tcb_field *req;
1755 KASSERT((cookie & ~M_COOKIE) == 0, ("%s: invalid cookie %#x", __func__,
1757 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1760 panic("%s: allocation failure.", __func__);
1764 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1765 req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1766 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie));
1767 req->mask = htobe64(mask);
1768 req->val = htobe64(val);
1774 t4_init_cpl_io_handlers(struct adapter *sc)
1777 t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1778 t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1779 t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1780 t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1781 t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1782 t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1783 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1787 t4_uninit_cpl_io_handlers(struct adapter *sc)
1790 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);