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>
48 #include <netinet/tcp_var.h>
50 #include <netinet/tcp_fsm.h>
51 #include <netinet/tcp_seq.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 = tp->t_maxopd = 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();
233 tp->t_maxseg -= TCPOLEN_TSTAMP_APPA;
236 if (G_TCPOPT_SACK(opt))
237 tp->t_flags |= TF_SACK_PERMIT; /* should already be set */
239 tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */
241 if (G_TCPOPT_WSCALE_OK(opt))
242 tp->t_flags |= TF_RCVD_SCALE;
244 /* Doing window scaling? */
245 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
246 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
247 tp->rcv_scale = tp->request_r_scale;
248 tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
253 * Completes some final bits of initialization for just established connections
254 * and changes their state to TCPS_ESTABLISHED.
256 * The ISNs are from after the exchange of SYNs. i.e., the true ISN + 1.
259 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
262 struct inpcb *inp = toep->inp;
263 struct socket *so = inp->inp_socket;
264 struct tcpcb *tp = intotcpcb(inp);
266 uint32_t iss = be32toh(snd_isn) - 1; /* true ISS */
267 uint32_t irs = be32toh(rcv_isn) - 1; /* true IRS */
268 uint16_t tcpopt = be16toh(opt);
269 struct flowc_tx_params ftxp;
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);
324 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
327 struct cpl_rx_data_ack *req;
328 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
330 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
332 wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
337 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
338 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
345 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
347 struct adapter *sc = tod->tod_softc;
348 struct inpcb *inp = tp->t_inpcb;
349 struct socket *so = inp->inp_socket;
350 struct sockbuf *sb = &so->so_rcv;
351 struct toepcb *toep = tp->t_toe;
354 INP_WLOCK_ASSERT(inp);
357 KASSERT(toep->sb_cc >= sbused(sb),
358 ("%s: sb %p has more data (%d) than last time (%d).",
359 __func__, sb, sbused(sb), toep->sb_cc));
361 toep->rx_credits += toep->sb_cc - sbused(sb);
362 toep->sb_cc = sbused(sb);
364 if (toep->rx_credits > 0 &&
365 (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
366 (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
367 toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
369 credits = send_rx_credits(sc, toep, toep->rx_credits);
370 toep->rx_credits -= credits;
371 tp->rcv_wnd += credits;
372 tp->rcv_adv += credits;
378 * Close a connection by sending a CPL_CLOSE_CON_REQ message.
381 close_conn(struct adapter *sc, struct toepcb *toep)
384 struct cpl_close_con_req *req;
385 unsigned int tid = toep->tid;
387 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
388 toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
390 if (toep->flags & TPF_FIN_SENT)
393 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
394 ("%s: flowc_wr not sent for tid %u.", __func__, tid));
396 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
399 panic("%s: allocation failure.", __func__);
403 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
404 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
405 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
406 V_FW_WR_FLOWID(tid));
407 req->wr.wr_lo = cpu_to_be64(0);
408 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
411 toep->flags |= TPF_FIN_SENT;
412 toep->flags &= ~TPF_SEND_FIN;
413 t4_l2t_send(sc, wr, toep->l2te);
418 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
419 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
421 /* Maximum amount of immediate data we could stuff in a WR */
423 max_imm_payload(int tx_credits)
425 const int n = 2; /* Use only up to 2 desc for imm. data WR */
427 KASSERT(tx_credits >= 0 &&
428 tx_credits <= MAX_OFLD_TX_CREDITS,
429 ("%s: %d credits", __func__, tx_credits));
431 if (tx_credits < MIN_OFLD_TX_CREDITS)
434 if (tx_credits >= (n * EQ_ESIZE) / 16)
435 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
437 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
440 /* Maximum number of SGL entries we could stuff in a WR */
442 max_dsgl_nsegs(int tx_credits)
444 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
445 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
447 KASSERT(tx_credits >= 0 &&
448 tx_credits <= MAX_OFLD_TX_CREDITS,
449 ("%s: %d credits", __func__, tx_credits));
451 if (tx_credits < MIN_OFLD_TX_CREDITS)
454 nseg += 2 * (sge_pair_credits * 16 / 24);
455 if ((sge_pair_credits * 16) % 24 == 16)
462 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
463 unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign)
465 struct fw_ofld_tx_data_wr *txwr = dst;
467 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
468 V_FW_WR_IMMDLEN(immdlen));
469 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
470 V_FW_WR_LEN16(credits));
471 txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) |
472 V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
473 txwr->plen = htobe32(plen);
476 struct tcpcb *tp = intotcpcb(toep->inp);
478 if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi))
479 txwr->lsodisable_to_flags |=
480 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
482 txwr->lsodisable_to_flags |=
483 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
484 (tp->t_flags & TF_NODELAY ? 0 :
485 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
490 * Generate a DSGL from a starting mbuf. The total number of segments and the
491 * maximum segments in any one mbuf are provided.
494 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
497 struct ulptx_sgl *usgl = dst;
500 struct sglist_seg segs[n];
502 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
504 sglist_init(&sg, n, segs);
505 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
506 V_ULPTX_NSGE(nsegs));
509 for (m = start; m != stop; m = m->m_next) {
510 rc = sglist_append(&sg, mtod(m, void *), m->m_len);
511 if (__predict_false(rc != 0))
512 panic("%s: sglist_append %d", __func__, rc);
514 for (j = 0; j < sg.sg_nseg; i++, j++) {
516 usgl->len0 = htobe32(segs[j].ss_len);
517 usgl->addr0 = htobe64(segs[j].ss_paddr);
519 usgl->sge[i / 2].len[i & 1] =
520 htobe32(segs[j].ss_len);
521 usgl->sge[i / 2].addr[i & 1] =
522 htobe64(segs[j].ss_paddr);
531 usgl->sge[i / 2].len[1] = htobe32(0);
532 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
533 __func__, nsegs, start, stop));
537 * Max number of SGL entries an offload tx work request can have. This is 41
538 * (1 + 40) for a full 512B work request.
539 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
541 #define OFLD_SGL_LEN (41)
544 * Send data and/or a FIN to the peer.
546 * The socket's so_snd buffer consists of a stream of data starting with sb_mb
547 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that
550 * drop indicates the number of bytes that should be dropped from the head of
551 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating
552 * contention on the send buffer lock (before this change it used to do
553 * sowwakeup and then t4_push_frames right after that when recovering from tx
554 * stalls). When drop is set this function MUST drop the bytes and wake up any
558 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
560 struct mbuf *sndptr, *m, *sb_sndptr;
561 struct fw_ofld_tx_data_wr *txwr;
563 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
564 struct inpcb *inp = toep->inp;
565 struct tcpcb *tp = intotcpcb(inp);
566 struct socket *so = inp->inp_socket;
567 struct sockbuf *sb = &so->so_snd;
568 int tx_credits, shove, compl, space, sowwakeup;
569 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
571 INP_WLOCK_ASSERT(inp);
572 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
573 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
575 KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
576 toep->ulp_mode == ULP_MODE_TCPDDP ||
577 toep->ulp_mode == ULP_MODE_RDMA,
578 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
580 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
584 * This function doesn't resume by itself. Someone else must clear the
585 * flag and call this function.
587 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
589 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
594 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
595 max_imm = max_imm_payload(tx_credits);
596 max_nsegs = max_dsgl_nsegs(tx_credits);
601 sbdrop_locked(sb, drop);
604 sb_sndptr = sb->sb_sndptr;
605 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
608 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
609 for (m = sndptr; m != NULL; m = m->m_next) {
610 int n = sglist_count(mtod(m, void *), m->m_len);
615 /* This mbuf sent us _over_ the nsegs limit, back out */
616 if (plen > max_imm && nsegs > max_nsegs) {
620 /* Too few credits */
621 toep->flags |= TPF_TX_SUSPENDED;
623 sowwakeup_locked(so);
626 SOCKBUF_UNLOCK_ASSERT(sb);
632 if (max_nsegs_1mbuf < n)
634 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */
636 /* This mbuf put us right at the max_nsegs limit */
637 if (plen > max_imm && nsegs == max_nsegs) {
645 if (space <= sb->sb_hiwat * 3 / 8 &&
646 toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
651 if (sb->sb_flags & SB_AUTOSIZE &&
652 V_tcp_do_autosndbuf &&
653 sb->sb_hiwat < V_tcp_autosndbuf_max &&
654 space < sb->sb_hiwat / 8) {
655 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
656 V_tcp_autosndbuf_max);
658 if (!sbreserve_locked(sb, newsize, so, NULL))
659 sb->sb_flags &= ~SB_AUTOSIZE;
661 sowwakeup = 1; /* room available */
664 sowwakeup_locked(so);
667 SOCKBUF_UNLOCK_ASSERT(sb);
669 /* nothing to send */
672 ("%s: nothing to send, but m != NULL", __func__));
676 if (__predict_false(toep->flags & TPF_FIN_SENT))
677 panic("%s: excess tx.", __func__);
679 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
680 if (plen <= max_imm) {
682 /* Immediate data tx */
684 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
687 /* XXX: how will we recover from this? */
688 toep->flags |= TPF_TX_SUSPENDED;
692 credits = howmany(wr->wr_len, 16);
693 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
695 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
702 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
703 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
704 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
706 /* XXX: how will we recover from this? */
707 toep->flags |= TPF_TX_SUSPENDED;
711 credits = howmany(wr_len, 16);
712 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
714 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
717 uint64_t *pad = (uint64_t *)
718 ((uintptr_t)txwr + wr_len);
723 KASSERT(toep->tx_credits >= credits,
724 ("%s: not enough credits", __func__));
726 toep->tx_credits -= credits;
727 toep->tx_nocompl += credits;
728 toep->plen_nocompl += plen;
729 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
730 toep->tx_nocompl >= toep->tx_total / 4)
733 if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
734 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
735 toep->tx_nocompl = 0;
736 toep->plen_nocompl = 0;
743 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
744 sb->sb_sndptr = sb_sndptr;
747 toep->flags |= TPF_TX_DATA_SENT;
748 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
749 toep->flags |= TPF_TX_SUSPENDED;
751 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
753 txsd->tx_credits = credits;
755 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
757 txsd = &toep->txsd[0];
761 t4_l2t_send(sc, wr, toep->l2te);
764 /* Send a FIN if requested, but only if there's no more data to send */
765 if (m == NULL && toep->flags & TPF_SEND_FIN)
766 close_conn(sc, toep);
770 rqdrop_locked(struct mbufq *q, int plen)
775 m = mbufq_dequeue(q);
777 /* Too many credits. */
781 /* Partial credits. */
782 MPASS(plen >= m->m_pkthdr.len);
784 plen -= m->m_pkthdr.len;
790 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
792 struct mbuf *sndptr, *m;
793 struct fw_ofld_tx_data_wr *txwr;
795 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
796 u_int adjusted_plen, ulp_submode;
797 struct inpcb *inp = toep->inp;
798 struct tcpcb *tp = intotcpcb(inp);
799 int tx_credits, shove;
800 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
801 struct mbufq *pduq = &toep->ulp_pduq;
802 static const u_int ulp_extra_len[] = {0, 4, 4, 8};
804 INP_WLOCK_ASSERT(inp);
805 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
806 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
807 KASSERT(toep->ulp_mode == ULP_MODE_ISCSI,
808 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
810 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
814 * This function doesn't resume by itself. Someone else must clear the
815 * flag and call this function.
817 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
819 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
824 rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
826 while ((sndptr = mbufq_first(pduq)) != NULL) {
827 M_ASSERTPKTHDR(sndptr);
829 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
830 max_imm = max_imm_payload(tx_credits);
831 max_nsegs = max_dsgl_nsegs(tx_credits);
835 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
836 for (m = sndptr; m != NULL; m = m->m_next) {
837 int n = sglist_count(mtod(m, void *), m->m_len);
843 * This mbuf would send us _over_ the nsegs limit.
844 * Suspend tx because the PDU can't be sent out.
846 if (plen > max_imm && nsegs > max_nsegs) {
847 toep->flags |= TPF_TX_SUSPENDED;
851 if (max_nsegs_1mbuf < n)
855 if (__predict_false(toep->flags & TPF_FIN_SENT))
856 panic("%s: excess tx.", __func__);
859 * We have a PDU to send. All of it goes out in one WR so 'm'
860 * is NULL. A PDU's length is always a multiple of 4.
863 MPASS((plen & 3) == 0);
864 MPASS(sndptr->m_pkthdr.len == plen);
866 shove = !(tp->t_flags & TF_MORETOCOME);
867 ulp_submode = mbuf_ulp_submode(sndptr);
868 MPASS(ulp_submode < nitems(ulp_extra_len));
871 * plen doesn't include header and data digests, which are
872 * generated and inserted in the right places by the TOE, but
873 * they do occupy TCP sequence space and need to be accounted
876 adjusted_plen = plen + ulp_extra_len[ulp_submode];
877 if (plen <= max_imm) {
879 /* Immediate data tx */
881 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
884 /* XXX: how will we recover from this? */
885 toep->flags |= TPF_TX_SUSPENDED;
889 credits = howmany(wr->wr_len, 16);
890 write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
891 shove, ulp_submode, sc->tt.tx_align);
892 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
898 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
899 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
900 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
902 /* XXX: how will we recover from this? */
903 toep->flags |= TPF_TX_SUSPENDED;
907 credits = howmany(wr_len, 16);
908 write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
909 shove, ulp_submode, sc->tt.tx_align);
910 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
913 uint64_t *pad = (uint64_t *)
914 ((uintptr_t)txwr + wr_len);
919 KASSERT(toep->tx_credits >= credits,
920 ("%s: not enough credits", __func__));
922 m = mbufq_dequeue(pduq);
924 mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
926 toep->tx_credits -= credits;
927 toep->tx_nocompl += credits;
928 toep->plen_nocompl += plen;
929 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
930 toep->tx_nocompl >= toep->tx_total / 4) {
931 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
932 toep->tx_nocompl = 0;
933 toep->plen_nocompl = 0;
936 tp->snd_nxt += adjusted_plen;
937 tp->snd_max += adjusted_plen;
939 toep->flags |= TPF_TX_DATA_SENT;
940 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
941 toep->flags |= TPF_TX_SUSPENDED;
943 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
945 txsd->tx_credits = credits;
947 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
949 txsd = &toep->txsd[0];
953 t4_l2t_send(sc, wr, toep->l2te);
956 /* Send a FIN if requested, but only if there are no more PDUs to send */
957 if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
958 close_conn(sc, toep);
962 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
964 struct adapter *sc = tod->tod_softc;
966 struct inpcb *inp = tp->t_inpcb;
968 struct toepcb *toep = tp->t_toe;
970 INP_WLOCK_ASSERT(inp);
971 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
972 ("%s: inp %p dropped.", __func__, inp));
973 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
975 if (toep->ulp_mode == ULP_MODE_ISCSI)
976 t4_push_pdus(sc, toep, 0);
978 t4_push_frames(sc, toep, 0);
984 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
986 struct adapter *sc = tod->tod_softc;
988 struct inpcb *inp = tp->t_inpcb;
990 struct toepcb *toep = tp->t_toe;
992 INP_WLOCK_ASSERT(inp);
993 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
994 ("%s: inp %p dropped.", __func__, inp));
995 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
997 toep->flags |= TPF_SEND_FIN;
998 if (tp->t_state >= TCPS_ESTABLISHED) {
999 if (toep->ulp_mode == ULP_MODE_ISCSI)
1000 t4_push_pdus(sc, toep, 0);
1002 t4_push_frames(sc, toep, 0);
1009 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1011 struct adapter *sc = tod->tod_softc;
1012 #if defined(INVARIANTS)
1013 struct inpcb *inp = tp->t_inpcb;
1015 struct toepcb *toep = tp->t_toe;
1017 INP_WLOCK_ASSERT(inp);
1018 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1019 ("%s: inp %p dropped.", __func__, inp));
1020 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1023 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1024 ("%s: flowc for tid %u [%s] not sent already",
1025 __func__, toep->tid, tcpstates[tp->t_state]));
1027 send_reset(sc, toep, 0);
1032 * Peer has sent us a FIN.
1035 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1037 struct adapter *sc = iq->adapter;
1038 const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1039 unsigned int tid = GET_TID(cpl);
1040 struct toepcb *toep = lookup_tid(sc, tid);
1041 struct inpcb *inp = toep->inp;
1042 struct tcpcb *tp = NULL;
1046 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1049 KASSERT(opcode == CPL_PEER_CLOSE,
1050 ("%s: unexpected opcode 0x%x", __func__, opcode));
1051 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1053 if (__predict_false(toep->flags & TPF_SYNQE)) {
1055 struct synq_entry *synqe = (void *)toep;
1057 INP_WLOCK(synqe->lctx->inp);
1058 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1059 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1060 ("%s: listen socket closed but tid %u not aborted.",
1064 * do_pass_accept_req is still running and will
1065 * eventually take care of this tid.
1068 INP_WUNLOCK(synqe->lctx->inp);
1070 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1075 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1077 INP_INFO_RLOCK(&V_tcbinfo);
1079 tp = intotcpcb(inp);
1081 CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1082 tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1084 if (toep->flags & TPF_ABORT_SHUTDOWN)
1087 tp->rcv_nxt++; /* FIN */
1089 so = inp->inp_socket;
1092 if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
1093 handle_ddp_close(toep, tp, sb, cpl->rcv_nxt);
1095 socantrcvmore_locked(so); /* unlocks the sockbuf */
1097 if (toep->ulp_mode != ULP_MODE_RDMA) {
1098 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1099 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1100 be32toh(cpl->rcv_nxt)));
1103 switch (tp->t_state) {
1104 case TCPS_SYN_RECEIVED:
1105 tp->t_starttime = ticks;
1108 case TCPS_ESTABLISHED:
1109 tp->t_state = TCPS_CLOSE_WAIT;
1112 case TCPS_FIN_WAIT_1:
1113 tp->t_state = TCPS_CLOSING;
1116 case TCPS_FIN_WAIT_2:
1118 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1119 INP_INFO_RUNLOCK(&V_tcbinfo);
1122 final_cpl_received(toep);
1126 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1127 __func__, tid, tp->t_state);
1131 INP_INFO_RUNLOCK(&V_tcbinfo);
1136 * Peer has ACK'd our FIN.
1139 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1142 struct adapter *sc = iq->adapter;
1143 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1144 unsigned int tid = GET_TID(cpl);
1145 struct toepcb *toep = lookup_tid(sc, tid);
1146 struct inpcb *inp = toep->inp;
1147 struct tcpcb *tp = NULL;
1148 struct socket *so = NULL;
1150 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1153 KASSERT(opcode == CPL_CLOSE_CON_RPL,
1154 ("%s: unexpected opcode 0x%x", __func__, opcode));
1155 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1156 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1158 INP_INFO_RLOCK(&V_tcbinfo);
1160 tp = intotcpcb(inp);
1162 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1163 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1165 if (toep->flags & TPF_ABORT_SHUTDOWN)
1168 so = inp->inp_socket;
1169 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */
1171 switch (tp->t_state) {
1172 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */
1175 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1176 INP_INFO_RUNLOCK(&V_tcbinfo);
1179 final_cpl_received(toep); /* no more CPLs expected */
1187 case TCPS_FIN_WAIT_1:
1188 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1189 soisdisconnected(so);
1190 tp->t_state = TCPS_FIN_WAIT_2;
1195 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1196 __func__, tid, tcpstates[tp->t_state]);
1200 INP_INFO_RUNLOCK(&V_tcbinfo);
1205 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1209 struct cpl_abort_rpl *cpl;
1211 wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1214 panic("%s: allocation failure.", __func__);
1218 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1219 cpl->cmd = rst_status;
1225 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1227 switch (abort_reason) {
1228 case CPL_ERR_BAD_SYN:
1229 case CPL_ERR_CONN_RESET:
1230 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1231 case CPL_ERR_XMIT_TIMEDOUT:
1232 case CPL_ERR_PERSIST_TIMEDOUT:
1233 case CPL_ERR_FINWAIT2_TIMEDOUT:
1234 case CPL_ERR_KEEPALIVE_TIMEDOUT:
1242 * TCP RST from the peer, timeout, or some other such critical error.
1245 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1247 struct adapter *sc = iq->adapter;
1248 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1249 unsigned int tid = GET_TID(cpl);
1250 struct toepcb *toep = lookup_tid(sc, tid);
1251 struct sge_wrq *ofld_txq = toep->ofld_txq;
1255 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1258 KASSERT(opcode == CPL_ABORT_REQ_RSS,
1259 ("%s: unexpected opcode 0x%x", __func__, opcode));
1260 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1262 if (toep->flags & TPF_SYNQE)
1263 return (do_abort_req_synqe(iq, rss, m));
1265 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1267 if (negative_advice(cpl->status)) {
1268 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1269 __func__, cpl->status, tid, toep->flags);
1270 return (0); /* Ignore negative advice */
1274 INP_INFO_RLOCK(&V_tcbinfo); /* for tcp_close */
1277 tp = intotcpcb(inp);
1280 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1281 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1282 inp->inp_flags, cpl->status);
1285 * If we'd initiated an abort earlier the reply to it is responsible for
1286 * cleaning up resources. Otherwise we tear everything down right here
1287 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
1289 if (toep->flags & TPF_ABORT_SHUTDOWN) {
1293 toep->flags |= TPF_ABORT_SHUTDOWN;
1295 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1296 struct socket *so = inp->inp_socket;
1299 so_error_set(so, abort_status_to_errno(tp,
1303 INP_WLOCK(inp); /* re-acquire */
1306 final_cpl_received(toep);
1308 INP_INFO_RUNLOCK(&V_tcbinfo);
1309 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1314 * Reply to the CPL_ABORT_REQ (send_reset)
1317 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1319 struct adapter *sc = iq->adapter;
1320 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1321 unsigned int tid = GET_TID(cpl);
1322 struct toepcb *toep = lookup_tid(sc, tid);
1323 struct inpcb *inp = toep->inp;
1325 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1328 KASSERT(opcode == CPL_ABORT_RPL_RSS,
1329 ("%s: unexpected opcode 0x%x", __func__, opcode));
1330 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1332 if (toep->flags & TPF_SYNQE)
1333 return (do_abort_rpl_synqe(iq, rss, m));
1335 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1337 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1338 __func__, tid, toep, inp, cpl->status);
1340 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1341 ("%s: wasn't expecting abort reply", __func__));
1344 final_cpl_received(toep);
1350 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1352 struct adapter *sc = iq->adapter;
1353 const struct cpl_rx_data *cpl = mtod(m, const void *);
1354 unsigned int tid = GET_TID(cpl);
1355 struct toepcb *toep = lookup_tid(sc, tid);
1356 struct inpcb *inp = toep->inp;
1361 uint32_t ddp_placed = 0;
1363 if (__predict_false(toep->flags & TPF_SYNQE)) {
1365 struct synq_entry *synqe = (void *)toep;
1367 INP_WLOCK(synqe->lctx->inp);
1368 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1369 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1370 ("%s: listen socket closed but tid %u not aborted.",
1374 * do_pass_accept_req is still running and will
1375 * eventually take care of this tid.
1378 INP_WUNLOCK(synqe->lctx->inp);
1380 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1386 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1388 /* strip off CPL header */
1389 m_adj(m, sizeof(*cpl));
1390 len = m->m_pkthdr.len;
1393 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1394 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1395 __func__, tid, len, inp->inp_flags);
1401 tp = intotcpcb(inp);
1403 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1404 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1407 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
1409 tp->t_rcvtime = ticks;
1411 so = inp_inpcbtosocket(inp);
1415 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1416 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1417 __func__, tid, len);
1422 INP_INFO_RLOCK(&V_tcbinfo);
1424 tp = tcp_drop(tp, ECONNRESET);
1427 INP_INFO_RUNLOCK(&V_tcbinfo);
1432 /* receive buffer autosize */
1433 if (sb->sb_flags & SB_AUTOSIZE &&
1434 V_tcp_do_autorcvbuf &&
1435 sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1436 len > (sbspace(sb) / 8 * 7)) {
1437 unsigned int hiwat = sb->sb_hiwat;
1438 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1439 V_tcp_autorcvbuf_max);
1441 if (!sbreserve_locked(sb, newsize, so, NULL))
1442 sb->sb_flags &= ~SB_AUTOSIZE;
1444 toep->rx_credits += newsize - hiwat;
1447 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1448 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1451 if (toep->ddp_flags & DDP_SC_REQ)
1452 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1454 KASSERT(cpl->ddp_off == 1,
1455 ("%s: DDP switched on by itself.",
1458 /* Fell out of DDP mode */
1459 toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1463 insert_ddp_data(toep, ddp_placed);
1467 if ((toep->ddp_flags & DDP_OK) == 0 &&
1468 time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1469 toep->ddp_score = DDP_LOW_SCORE;
1470 toep->ddp_flags |= DDP_OK;
1471 CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1472 __func__, tid, time_uptime);
1475 if (toep->ddp_flags & DDP_ON) {
1478 * CPL_RX_DATA with DDP on can only be an indicate. Ask
1479 * soreceive to post a buffer or disable DDP. The
1480 * payload that arrived in this indicate is appended to
1481 * the socket buffer as usual.
1486 "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1487 __func__, tid, toep->flags, be32toh(cpl->seq), len);
1489 sb->sb_flags |= SB_DDP_INDICATE;
1490 } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1491 tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1494 * DDP allowed but isn't on (and a request to switch it
1495 * on isn't pending either), and conditions are ripe for
1496 * it to work. Switch it on.
1499 enable_ddp(sc, toep);
1503 KASSERT(toep->sb_cc >= sbused(sb),
1504 ("%s: sb %p has more data (%d) than last time (%d).",
1505 __func__, sb, sbused(sb), toep->sb_cc));
1506 toep->rx_credits += toep->sb_cc - sbused(sb);
1507 sbappendstream_locked(sb, m, 0);
1508 toep->sb_cc = sbused(sb);
1509 if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1512 credits = send_rx_credits(sc, toep, toep->rx_credits);
1513 toep->rx_credits -= credits;
1514 tp->rcv_wnd += credits;
1515 tp->rcv_adv += credits;
1517 sorwakeup_locked(so);
1518 SOCKBUF_UNLOCK_ASSERT(sb);
1524 #define S_CPL_FW4_ACK_OPCODE 24
1525 #define M_CPL_FW4_ACK_OPCODE 0xff
1526 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1527 #define G_CPL_FW4_ACK_OPCODE(x) \
1528 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1530 #define S_CPL_FW4_ACK_FLOWID 0
1531 #define M_CPL_FW4_ACK_FLOWID 0xffffff
1532 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1533 #define G_CPL_FW4_ACK_FLOWID(x) \
1534 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1536 #define S_CPL_FW4_ACK_CR 24
1537 #define M_CPL_FW4_ACK_CR 0xff
1538 #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR)
1539 #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1541 #define S_CPL_FW4_ACK_SEQVAL 0
1542 #define M_CPL_FW4_ACK_SEQVAL 0x1
1543 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1544 #define G_CPL_FW4_ACK_SEQVAL(x) \
1545 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1546 #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U)
1549 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1551 struct adapter *sc = iq->adapter;
1552 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1553 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1554 struct toepcb *toep = lookup_tid(sc, tid);
1558 uint8_t credits = cpl->credits;
1559 struct ofld_tx_sdesc *txsd;
1562 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1566 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1567 * now this comes back carrying the credits for the flowc.
1569 if (__predict_false(toep->flags & TPF_SYNQE)) {
1570 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1571 ("%s: credits for a synq entry %p", __func__, toep));
1577 KASSERT(opcode == CPL_FW4_ACK,
1578 ("%s: unexpected opcode 0x%x", __func__, opcode));
1579 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1580 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1584 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1589 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1590 ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1592 tp = intotcpcb(inp);
1594 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1595 tcp_seq snd_una = be32toh(cpl->snd_una);
1598 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1600 "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1601 __func__, snd_una, toep->tid, tp->snd_una);
1605 if (tp->snd_una != snd_una) {
1606 tp->snd_una = snd_una;
1607 tp->ts_recent_age = tcp_ts_getticks();
1611 so = inp->inp_socket;
1612 txsd = &toep->txsd[toep->txsd_cidx];
1615 KASSERT(credits >= txsd->tx_credits,
1616 ("%s: too many (or partial) credits", __func__));
1617 credits -= txsd->tx_credits;
1618 toep->tx_credits += txsd->tx_credits;
1622 KASSERT(toep->txsd_avail <= toep->txsd_total,
1623 ("%s: txsd avail > total", __func__));
1624 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1625 txsd = &toep->txsd[0];
1626 toep->txsd_cidx = 0;
1630 if (toep->tx_credits == toep->tx_total) {
1631 toep->tx_nocompl = 0;
1632 toep->plen_nocompl = 0;
1635 if (toep->flags & TPF_TX_SUSPENDED &&
1636 toep->tx_credits >= toep->tx_total / 4) {
1637 toep->flags &= ~TPF_TX_SUSPENDED;
1638 if (toep->ulp_mode == ULP_MODE_ISCSI)
1639 t4_push_pdus(sc, toep, plen);
1641 t4_push_frames(sc, toep, plen);
1642 } else if (plen > 0) {
1643 struct sockbuf *sb = &so->so_snd;
1648 if (toep->ulp_mode == ULP_MODE_ISCSI) {
1650 if (__predict_false(sbu > 0)) {
1652 * The data trasmitted before the tid's ULP mode
1653 * changed to ISCSI is still in so_snd.
1654 * Incoming credits should account for so_snd
1657 sbdrop_locked(sb, min(sbu, plen));
1658 plen -= min(sbu, plen);
1660 sowwakeup_locked(so); /* unlocks so_snd */
1661 rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1663 sbdrop_locked(sb, plen);
1664 sowwakeup_locked(so); /* unlocks so_snd */
1666 SOCKBUF_UNLOCK_ASSERT(sb);
1675 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1677 struct adapter *sc = iq->adapter;
1678 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1679 unsigned int tid = GET_TID(cpl);
1681 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1684 KASSERT(opcode == CPL_SET_TCB_RPL,
1685 ("%s: unexpected opcode 0x%x", __func__, opcode));
1686 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1688 if (is_ftid(sc, tid))
1689 return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1692 * TOM and/or other ULPs don't request replies for CPL_SET_TCB or
1693 * CPL_SET_TCB_FIELD requests. This can easily change and when it does
1694 * the dispatch code will go here.
1697 panic("%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p", __func__,
1700 log(LOG_ERR, "%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p\n",
1708 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1709 uint16_t word, uint64_t mask, uint64_t val)
1712 struct cpl_set_tcb_field *req;
1714 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1717 panic("%s: allocation failure.", __func__);
1721 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1722 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1723 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1724 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1725 req->mask = htobe64(mask);
1726 req->val = htobe64(val);
1732 t4_init_cpl_io_handlers(struct adapter *sc)
1735 t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1736 t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1737 t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1738 t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1739 t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1740 t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1741 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1745 t4_uninit_cpl_io_handlers(struct adapter *sc)
1748 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);