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)
74 static inline struct mbuf *
75 mbufq_dequeue(struct mbufq *q)
83 q->head = m->m_nextpkt;
90 mbufq_enqueue(struct mbufq *q, struct mbuf *m)
95 q->tail->m_nextpkt = m;
101 static inline struct mbuf *
102 mbufq_first(const struct mbufq *q)
109 send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp)
112 struct fw_flowc_wr *flowc;
113 unsigned int nparams = ftxp ? 8 : 6, flowclen;
114 struct vi_info *vi = toep->vi;
115 struct port_info *pi = vi->pi;
116 struct adapter *sc = pi->adapter;
117 unsigned int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
118 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
120 KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
121 ("%s: flowc for tid %u sent already", __func__, toep->tid));
123 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
125 wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
128 panic("%s: allocation failure.", __func__);
131 memset(flowc, 0, wr->wr_len);
133 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
134 V_FW_FLOWC_WR_NPARAMS(nparams));
135 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
136 V_FW_WR_FLOWID(toep->tid));
138 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
139 flowc->mnemval[0].val = htobe32(pfvf);
140 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
141 flowc->mnemval[1].val = htobe32(pi->tx_chan);
142 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
143 flowc->mnemval[2].val = htobe32(pi->tx_chan);
144 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
145 flowc->mnemval[3].val = htobe32(toep->ofld_rxq->iq.abs_id);
147 uint32_t sndbuf = min(ftxp->snd_space, sc->tt.sndbuf);
149 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDNXT;
150 flowc->mnemval[4].val = htobe32(ftxp->snd_nxt);
151 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_RCVNXT;
152 flowc->mnemval[5].val = htobe32(ftxp->rcv_nxt);
153 flowc->mnemval[6].mnemonic = FW_FLOWC_MNEM_SNDBUF;
154 flowc->mnemval[6].val = htobe32(sndbuf);
155 flowc->mnemval[7].mnemonic = FW_FLOWC_MNEM_MSS;
156 flowc->mnemval[7].val = htobe32(ftxp->mss);
159 "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x",
160 __func__, toep->tid, ftxp->mss, sndbuf, ftxp->snd_nxt,
163 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF;
164 flowc->mnemval[4].val = htobe32(512);
165 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS;
166 flowc->mnemval[5].val = htobe32(512);
168 CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid);
171 txsd->tx_credits = howmany(flowclen, 16);
173 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
174 ("%s: not enough credits (%d)", __func__, toep->tx_credits));
175 toep->tx_credits -= txsd->tx_credits;
176 if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
180 toep->flags |= TPF_FLOWC_WR_SENT;
185 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
188 struct cpl_abort_req *req;
190 struct inpcb *inp = toep->inp;
191 struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */
193 INP_WLOCK_ASSERT(inp);
195 CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
197 inp->inp_flags & INP_DROPPED ? "inp dropped" :
198 tcpstates[tp->t_state],
199 toep->flags, inp->inp_flags,
200 toep->flags & TPF_ABORT_SHUTDOWN ?
201 " (abort already in progress)" : "");
203 if (toep->flags & TPF_ABORT_SHUTDOWN)
204 return; /* abort already in progress */
206 toep->flags |= TPF_ABORT_SHUTDOWN;
208 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
209 ("%s: flowc_wr not sent for tid %d.", __func__, tid));
211 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
214 panic("%s: allocation failure.", __func__);
218 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
219 if (inp->inp_flags & INP_DROPPED)
220 req->rsvd0 = htobe32(snd_nxt);
222 req->rsvd0 = htobe32(tp->snd_nxt);
223 req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
224 req->cmd = CPL_ABORT_SEND_RST;
227 * XXX: What's the correct way to tell that the inp hasn't been detached
228 * from its socket? Should I even be flushing the snd buffer here?
230 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
231 struct socket *so = inp->inp_socket;
233 if (so != NULL) /* because I'm not sure. See comment above */
234 sbflush(&so->so_snd);
237 t4_l2t_send(sc, wr, toep->l2te);
241 * Called when a connection is established to translate the TCP options
242 * reported by HW to FreeBSD's native format.
245 assign_rxopt(struct tcpcb *tp, unsigned int opt)
247 struct toepcb *toep = tp->t_toe;
248 struct inpcb *inp = tp->t_inpcb;
249 struct adapter *sc = td_adapter(toep->td);
252 INP_LOCK_ASSERT(inp);
254 if (inp->inp_inc.inc_flags & INC_ISIPV6)
255 n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
257 n = sizeof(struct ip) + sizeof(struct tcphdr);
258 tp->t_maxseg = tp->t_maxopd = sc->params.mtus[G_TCPOPT_MSS(opt)] - n;
260 CTR4(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u)", __func__, toep->tid,
261 G_TCPOPT_MSS(opt), sc->params.mtus[G_TCPOPT_MSS(opt)]);
263 if (G_TCPOPT_TSTAMP(opt)) {
264 tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */
265 tp->ts_recent = 0; /* hmmm */
266 tp->ts_recent_age = tcp_ts_getticks();
267 tp->t_maxseg -= TCPOLEN_TSTAMP_APPA;
270 if (G_TCPOPT_SACK(opt))
271 tp->t_flags |= TF_SACK_PERMIT; /* should already be set */
273 tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */
275 if (G_TCPOPT_WSCALE_OK(opt))
276 tp->t_flags |= TF_RCVD_SCALE;
278 /* Doing window scaling? */
279 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
280 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
281 tp->rcv_scale = tp->request_r_scale;
282 tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
287 * Completes some final bits of initialization for just established connections
288 * and changes their state to TCPS_ESTABLISHED.
290 * The ISNs are from after the exchange of SYNs. i.e., the true ISN + 1.
293 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
296 struct inpcb *inp = toep->inp;
297 struct socket *so = inp->inp_socket;
298 struct tcpcb *tp = intotcpcb(inp);
300 uint32_t iss = be32toh(snd_isn) - 1; /* true ISS */
301 uint32_t irs = be32toh(rcv_isn) - 1; /* true IRS */
302 uint16_t tcpopt = be16toh(opt);
303 struct flowc_tx_params ftxp;
305 CURVNET_SET(so->so_vnet);
306 INP_WLOCK_ASSERT(inp);
307 KASSERT(tp->t_state == TCPS_SYN_SENT ||
308 tp->t_state == TCPS_SYN_RECEIVED,
309 ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
311 CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p",
312 __func__, toep->tid, toep, inp);
314 tp->t_state = TCPS_ESTABLISHED;
315 tp->t_starttime = ticks;
316 TCPSTAT_INC(tcps_connects);
320 tp->rcv_wnd = toep->rx_credits << 10;
321 tp->rcv_adv += tp->rcv_wnd;
322 tp->last_ack_sent = tp->rcv_nxt;
325 * If we were unable to send all rx credits via opt0, save the remainder
326 * in rx_credits so that they can be handed over with the next credit
329 SOCKBUF_LOCK(&so->so_rcv);
330 bufsize = select_rcv_wnd(so);
331 SOCKBUF_UNLOCK(&so->so_rcv);
332 toep->rx_credits = bufsize - tp->rcv_wnd;
336 tp->snd_una = iss + 1;
337 tp->snd_nxt = iss + 1;
338 tp->snd_max = iss + 1;
340 assign_rxopt(tp, tcpopt);
342 SOCKBUF_LOCK(&so->so_snd);
343 if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf)
344 bufsize = V_tcp_autosndbuf_max;
346 bufsize = sbspace(&so->so_snd);
347 SOCKBUF_UNLOCK(&so->so_snd);
349 ftxp.snd_nxt = tp->snd_nxt;
350 ftxp.rcv_nxt = tp->rcv_nxt;
351 ftxp.snd_space = bufsize;
352 ftxp.mss = tp->t_maxseg;
353 send_flowc_wr(toep, &ftxp);
360 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
363 struct cpl_rx_data_ack *req;
364 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
366 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
368 wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
373 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
374 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
381 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
383 struct adapter *sc = tod->tod_softc;
384 struct inpcb *inp = tp->t_inpcb;
385 struct socket *so = inp->inp_socket;
386 struct sockbuf *sb = &so->so_rcv;
387 struct toepcb *toep = tp->t_toe;
390 INP_WLOCK_ASSERT(inp);
393 KASSERT(toep->sb_cc >= sb->sb_cc,
394 ("%s: sb %p has more data (%d) than last time (%d).",
395 __func__, sb, sb->sb_cc, toep->sb_cc));
396 toep->rx_credits += toep->sb_cc - sb->sb_cc;
397 toep->sb_cc = sb->sb_cc;
399 if (toep->rx_credits > 0 &&
400 (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
401 (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
402 toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
404 credits = send_rx_credits(sc, toep, toep->rx_credits);
405 toep->rx_credits -= credits;
406 tp->rcv_wnd += credits;
407 tp->rcv_adv += credits;
413 * Close a connection by sending a CPL_CLOSE_CON_REQ message.
416 close_conn(struct adapter *sc, struct toepcb *toep)
419 struct cpl_close_con_req *req;
420 unsigned int tid = toep->tid;
422 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
423 toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
425 if (toep->flags & TPF_FIN_SENT)
428 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
429 ("%s: flowc_wr not sent for tid %u.", __func__, tid));
431 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
434 panic("%s: allocation failure.", __func__);
438 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
439 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
440 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
441 V_FW_WR_FLOWID(tid));
442 req->wr.wr_lo = cpu_to_be64(0);
443 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
446 toep->flags |= TPF_FIN_SENT;
447 toep->flags &= ~TPF_SEND_FIN;
448 t4_l2t_send(sc, wr, toep->l2te);
453 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
454 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
456 /* Maximum amount of immediate data we could stuff in a WR */
458 max_imm_payload(int tx_credits)
460 const int n = 2; /* Use only up to 2 desc for imm. data WR */
462 KASSERT(tx_credits >= 0 &&
463 tx_credits <= MAX_OFLD_TX_CREDITS,
464 ("%s: %d credits", __func__, tx_credits));
466 if (tx_credits < MIN_OFLD_TX_CREDITS)
469 if (tx_credits >= (n * EQ_ESIZE) / 16)
470 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
472 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
475 /* Maximum number of SGL entries we could stuff in a WR */
477 max_dsgl_nsegs(int tx_credits)
479 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
480 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
482 KASSERT(tx_credits >= 0 &&
483 tx_credits <= MAX_OFLD_TX_CREDITS,
484 ("%s: %d credits", __func__, tx_credits));
486 if (tx_credits < MIN_OFLD_TX_CREDITS)
489 nseg += 2 * (sge_pair_credits * 16 / 24);
490 if ((sge_pair_credits * 16) % 24 == 16)
497 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
498 unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign)
500 struct fw_ofld_tx_data_wr *txwr = dst;
502 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
503 V_FW_WR_IMMDLEN(immdlen));
504 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
505 V_FW_WR_LEN16(credits));
506 txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) |
507 V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
508 txwr->plen = htobe32(plen);
511 struct tcpcb *tp = intotcpcb(toep->inp);
513 if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi))
514 txwr->lsodisable_to_flags |=
515 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
517 txwr->lsodisable_to_flags |=
518 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
519 (tp->t_flags & TF_NODELAY ? 0 :
520 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
525 * Generate a DSGL from a starting mbuf. The total number of segments and the
526 * maximum segments in any one mbuf are provided.
529 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
532 struct ulptx_sgl *usgl = dst;
535 struct sglist_seg segs[n];
537 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
539 sglist_init(&sg, n, segs);
540 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
541 V_ULPTX_NSGE(nsegs));
544 for (m = start; m != stop; m = m->m_next) {
545 rc = sglist_append(&sg, mtod(m, void *), m->m_len);
546 if (__predict_false(rc != 0))
547 panic("%s: sglist_append %d", __func__, rc);
549 for (j = 0; j < sg.sg_nseg; i++, j++) {
551 usgl->len0 = htobe32(segs[j].ss_len);
552 usgl->addr0 = htobe64(segs[j].ss_paddr);
554 usgl->sge[i / 2].len[i & 1] =
555 htobe32(segs[j].ss_len);
556 usgl->sge[i / 2].addr[i & 1] =
557 htobe64(segs[j].ss_paddr);
566 usgl->sge[i / 2].len[1] = htobe32(0);
567 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
568 __func__, nsegs, start, stop));
572 * Max number of SGL entries an offload tx work request can have. This is 41
573 * (1 + 40) for a full 512B work request.
574 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
576 #define OFLD_SGL_LEN (41)
579 * Send data and/or a FIN to the peer.
581 * The socket's so_snd buffer consists of a stream of data starting with sb_mb
582 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that
585 * drop indicates the number of bytes that should be dropped from the head of
586 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating
587 * contention on the send buffer lock (before this change it used to do
588 * sowwakeup and then t4_push_frames right after that when recovering from tx
589 * stalls). When drop is set this function MUST drop the bytes and wake up any
593 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
595 struct mbuf *sndptr, *m, *sb_sndptr;
596 struct fw_ofld_tx_data_wr *txwr;
598 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
599 struct inpcb *inp = toep->inp;
600 struct tcpcb *tp = intotcpcb(inp);
601 struct socket *so = inp->inp_socket;
602 struct sockbuf *sb = &so->so_snd;
603 int tx_credits, shove, compl, sowwakeup;
604 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
606 INP_WLOCK_ASSERT(inp);
607 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
608 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
610 KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
611 toep->ulp_mode == ULP_MODE_TCPDDP ||
612 toep->ulp_mode == ULP_MODE_RDMA,
613 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
615 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
619 * This function doesn't resume by itself. Someone else must clear the
620 * flag and call this function.
622 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
624 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
629 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
630 max_imm = max_imm_payload(tx_credits);
631 max_nsegs = max_dsgl_nsegs(tx_credits);
636 sbdrop_locked(sb, drop);
639 sb_sndptr = sb->sb_sndptr;
640 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
643 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
644 for (m = sndptr; m != NULL; m = m->m_next) {
645 int n = sglist_count(mtod(m, void *), m->m_len);
650 /* This mbuf sent us _over_ the nsegs limit, back out */
651 if (plen > max_imm && nsegs > max_nsegs) {
655 /* Too few credits */
656 toep->flags |= TPF_TX_SUSPENDED;
658 sowwakeup_locked(so);
661 SOCKBUF_UNLOCK_ASSERT(sb);
667 if (max_nsegs_1mbuf < n)
669 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */
671 /* This mbuf put us right at the max_nsegs limit */
672 if (plen > max_imm && nsegs == max_nsegs) {
678 if (sb->sb_cc > sb->sb_hiwat * 5 / 8 &&
679 toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
684 if (sb->sb_flags & SB_AUTOSIZE &&
685 V_tcp_do_autosndbuf &&
686 sb->sb_hiwat < V_tcp_autosndbuf_max &&
687 sb->sb_cc >= sb->sb_hiwat * 7 / 8) {
688 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
689 V_tcp_autosndbuf_max);
691 if (!sbreserve_locked(sb, newsize, so, NULL))
692 sb->sb_flags &= ~SB_AUTOSIZE;
694 sowwakeup = 1; /* room available */
697 sowwakeup_locked(so);
700 SOCKBUF_UNLOCK_ASSERT(sb);
702 /* nothing to send */
705 ("%s: nothing to send, but m != NULL", __func__));
709 if (__predict_false(toep->flags & TPF_FIN_SENT))
710 panic("%s: excess tx.", __func__);
712 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
713 if (plen <= max_imm) {
715 /* Immediate data tx */
717 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
720 /* XXX: how will we recover from this? */
721 toep->flags |= TPF_TX_SUSPENDED;
725 credits = howmany(wr->wr_len, 16);
726 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
728 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
735 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
736 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
737 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
739 /* XXX: how will we recover from this? */
740 toep->flags |= TPF_TX_SUSPENDED;
744 credits = howmany(wr_len, 16);
745 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
747 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
750 uint64_t *pad = (uint64_t *)
751 ((uintptr_t)txwr + wr_len);
756 KASSERT(toep->tx_credits >= credits,
757 ("%s: not enough credits", __func__));
759 toep->tx_credits -= credits;
760 toep->tx_nocompl += credits;
761 toep->plen_nocompl += plen;
762 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
763 toep->tx_nocompl >= toep->tx_total / 4)
766 if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
767 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
768 toep->tx_nocompl = 0;
769 toep->plen_nocompl = 0;
776 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
777 sb->sb_sndptr = sb_sndptr;
780 toep->flags |= TPF_TX_DATA_SENT;
781 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
782 toep->flags |= TPF_TX_SUSPENDED;
784 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
786 txsd->tx_credits = credits;
788 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
790 txsd = &toep->txsd[0];
794 t4_l2t_send(sc, wr, toep->l2te);
797 /* Send a FIN if requested, but only if there's no more data to send */
798 if (m == NULL && toep->flags & TPF_SEND_FIN)
799 close_conn(sc, toep);
803 rqdrop_locked(struct mbufq *q, int plen)
808 m = mbufq_dequeue(q);
810 /* Too many credits. */
814 /* Partial credits. */
815 MPASS(plen >= m->m_pkthdr.len);
817 plen -= m->m_pkthdr.len;
823 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
825 struct mbuf *sndptr, *m;
826 struct fw_ofld_tx_data_wr *txwr;
828 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
829 u_int adjusted_plen, ulp_submode;
830 struct inpcb *inp = toep->inp;
831 struct tcpcb *tp = intotcpcb(inp);
832 int tx_credits, shove;
833 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
834 struct mbufq *pduq = &toep->ulp_pduq;
835 static const u_int ulp_extra_len[] = {0, 4, 4, 8};
837 INP_WLOCK_ASSERT(inp);
838 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
839 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
840 KASSERT(toep->ulp_mode == ULP_MODE_ISCSI,
841 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
843 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
847 * This function doesn't resume by itself. Someone else must clear the
848 * flag and call this function.
850 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
852 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
857 rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
859 while ((sndptr = mbufq_first(pduq)) != NULL) {
860 M_ASSERTPKTHDR(sndptr);
862 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
863 max_imm = max_imm_payload(tx_credits);
864 max_nsegs = max_dsgl_nsegs(tx_credits);
868 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
869 for (m = sndptr; m != NULL; m = m->m_next) {
870 int n = sglist_count(mtod(m, void *), m->m_len);
876 * This mbuf would send us _over_ the nsegs limit.
877 * Suspend tx because the PDU can't be sent out.
879 if (plen > max_imm && nsegs > max_nsegs) {
880 toep->flags |= TPF_TX_SUSPENDED;
884 if (max_nsegs_1mbuf < n)
888 if (__predict_false(toep->flags & TPF_FIN_SENT))
889 panic("%s: excess tx.", __func__);
892 * We have a PDU to send. All of it goes out in one WR so 'm'
893 * is NULL. A PDU's length is always a multiple of 4.
896 MPASS((plen & 3) == 0);
897 MPASS(sndptr->m_pkthdr.len == plen);
899 shove = !(tp->t_flags & TF_MORETOCOME);
900 ulp_submode = mbuf_ulp_submode(sndptr);
901 MPASS(ulp_submode < nitems(ulp_extra_len));
904 * plen doesn't include header and data digests, which are
905 * generated and inserted in the right places by the TOE, but
906 * they do occupy TCP sequence space and need to be accounted
909 adjusted_plen = plen + ulp_extra_len[ulp_submode];
910 if (plen <= max_imm) {
912 /* Immediate data tx */
914 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
917 /* XXX: how will we recover from this? */
918 toep->flags |= TPF_TX_SUSPENDED;
922 credits = howmany(wr->wr_len, 16);
923 write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
924 shove, ulp_submode, sc->tt.tx_align);
925 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
931 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
932 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
933 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
935 /* XXX: how will we recover from this? */
936 toep->flags |= TPF_TX_SUSPENDED;
940 credits = howmany(wr_len, 16);
941 write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
942 shove, ulp_submode, sc->tt.tx_align);
943 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
946 uint64_t *pad = (uint64_t *)
947 ((uintptr_t)txwr + wr_len);
952 KASSERT(toep->tx_credits >= credits,
953 ("%s: not enough credits", __func__));
955 m = mbufq_dequeue(pduq);
957 mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
959 toep->tx_credits -= credits;
960 toep->tx_nocompl += credits;
961 toep->plen_nocompl += plen;
962 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
963 toep->tx_nocompl >= toep->tx_total / 4) {
964 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
965 toep->tx_nocompl = 0;
966 toep->plen_nocompl = 0;
969 tp->snd_nxt += adjusted_plen;
970 tp->snd_max += adjusted_plen;
972 toep->flags |= TPF_TX_DATA_SENT;
973 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
974 toep->flags |= TPF_TX_SUSPENDED;
976 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
978 txsd->tx_credits = credits;
980 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
982 txsd = &toep->txsd[0];
986 t4_l2t_send(sc, wr, toep->l2te);
989 /* Send a FIN if requested, but only if there are no more PDUs to send */
990 if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
991 close_conn(sc, toep);
995 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
997 struct adapter *sc = tod->tod_softc;
999 struct inpcb *inp = tp->t_inpcb;
1001 struct toepcb *toep = tp->t_toe;
1003 INP_WLOCK_ASSERT(inp);
1004 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1005 ("%s: inp %p dropped.", __func__, inp));
1006 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1008 if (toep->ulp_mode == ULP_MODE_ISCSI)
1009 t4_push_pdus(sc, toep, 0);
1011 t4_push_frames(sc, toep, 0);
1017 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
1019 struct adapter *sc = tod->tod_softc;
1021 struct inpcb *inp = tp->t_inpcb;
1023 struct toepcb *toep = tp->t_toe;
1025 INP_WLOCK_ASSERT(inp);
1026 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1027 ("%s: inp %p dropped.", __func__, inp));
1028 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1030 toep->flags |= TPF_SEND_FIN;
1031 if (tp->t_state >= TCPS_ESTABLISHED) {
1032 if (toep->ulp_mode == ULP_MODE_ISCSI)
1033 t4_push_pdus(sc, toep, 0);
1035 t4_push_frames(sc, toep, 0);
1042 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1044 struct adapter *sc = tod->tod_softc;
1045 #if defined(INVARIANTS)
1046 struct inpcb *inp = tp->t_inpcb;
1048 struct toepcb *toep = tp->t_toe;
1050 INP_WLOCK_ASSERT(inp);
1051 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1052 ("%s: inp %p dropped.", __func__, inp));
1053 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1056 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1057 ("%s: flowc for tid %u [%s] not sent already",
1058 __func__, toep->tid, tcpstates[tp->t_state]));
1060 send_reset(sc, toep, 0);
1065 * Peer has sent us a FIN.
1068 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1070 struct adapter *sc = iq->adapter;
1071 const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1072 unsigned int tid = GET_TID(cpl);
1073 struct toepcb *toep = lookup_tid(sc, tid);
1074 struct inpcb *inp = toep->inp;
1075 struct tcpcb *tp = NULL;
1079 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1082 KASSERT(opcode == CPL_PEER_CLOSE,
1083 ("%s: unexpected opcode 0x%x", __func__, opcode));
1084 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1086 if (__predict_false(toep->flags & TPF_SYNQE)) {
1088 struct synq_entry *synqe = (void *)toep;
1090 INP_WLOCK(synqe->lctx->inp);
1091 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1092 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1093 ("%s: listen socket closed but tid %u not aborted.",
1097 * do_pass_accept_req is still running and will
1098 * eventually take care of this tid.
1101 INP_WUNLOCK(synqe->lctx->inp);
1103 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1108 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1110 INP_INFO_RLOCK(&V_tcbinfo);
1112 tp = intotcpcb(inp);
1114 CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1115 tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1117 if (toep->flags & TPF_ABORT_SHUTDOWN)
1120 tp->rcv_nxt++; /* FIN */
1122 so = inp->inp_socket;
1125 if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
1126 handle_ddp_close(toep, tp, sb, cpl->rcv_nxt);
1128 socantrcvmore_locked(so); /* unlocks the sockbuf */
1130 if (toep->ulp_mode != ULP_MODE_RDMA) {
1131 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1132 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1133 be32toh(cpl->rcv_nxt)));
1136 switch (tp->t_state) {
1137 case TCPS_SYN_RECEIVED:
1138 tp->t_starttime = ticks;
1141 case TCPS_ESTABLISHED:
1142 tp->t_state = TCPS_CLOSE_WAIT;
1145 case TCPS_FIN_WAIT_1:
1146 tp->t_state = TCPS_CLOSING;
1149 case TCPS_FIN_WAIT_2:
1151 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1152 INP_INFO_RUNLOCK(&V_tcbinfo);
1155 final_cpl_received(toep);
1159 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1160 __func__, tid, tp->t_state);
1164 INP_INFO_RUNLOCK(&V_tcbinfo);
1169 * Peer has ACK'd our FIN.
1172 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1175 struct adapter *sc = iq->adapter;
1176 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1177 unsigned int tid = GET_TID(cpl);
1178 struct toepcb *toep = lookup_tid(sc, tid);
1179 struct inpcb *inp = toep->inp;
1180 struct tcpcb *tp = NULL;
1181 struct socket *so = NULL;
1183 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1186 KASSERT(opcode == CPL_CLOSE_CON_RPL,
1187 ("%s: unexpected opcode 0x%x", __func__, opcode));
1188 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1189 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1191 INP_INFO_RLOCK(&V_tcbinfo);
1193 tp = intotcpcb(inp);
1195 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1196 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1198 if (toep->flags & TPF_ABORT_SHUTDOWN)
1201 so = inp->inp_socket;
1202 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */
1204 switch (tp->t_state) {
1205 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */
1208 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1209 INP_INFO_RUNLOCK(&V_tcbinfo);
1212 final_cpl_received(toep); /* no more CPLs expected */
1220 case TCPS_FIN_WAIT_1:
1221 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1222 soisdisconnected(so);
1223 tp->t_state = TCPS_FIN_WAIT_2;
1228 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1229 __func__, tid, tcpstates[tp->t_state]);
1233 INP_INFO_RUNLOCK(&V_tcbinfo);
1238 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1242 struct cpl_abort_rpl *cpl;
1244 wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1247 panic("%s: allocation failure.", __func__);
1251 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1252 cpl->cmd = rst_status;
1258 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1260 switch (abort_reason) {
1261 case CPL_ERR_BAD_SYN:
1262 case CPL_ERR_CONN_RESET:
1263 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1264 case CPL_ERR_XMIT_TIMEDOUT:
1265 case CPL_ERR_PERSIST_TIMEDOUT:
1266 case CPL_ERR_FINWAIT2_TIMEDOUT:
1267 case CPL_ERR_KEEPALIVE_TIMEDOUT:
1275 * TCP RST from the peer, timeout, or some other such critical error.
1278 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1280 struct adapter *sc = iq->adapter;
1281 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1282 unsigned int tid = GET_TID(cpl);
1283 struct toepcb *toep = lookup_tid(sc, tid);
1284 struct sge_wrq *ofld_txq = toep->ofld_txq;
1288 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1291 KASSERT(opcode == CPL_ABORT_REQ_RSS,
1292 ("%s: unexpected opcode 0x%x", __func__, opcode));
1293 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1295 if (toep->flags & TPF_SYNQE)
1296 return (do_abort_req_synqe(iq, rss, m));
1298 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1300 if (negative_advice(cpl->status)) {
1301 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1302 __func__, cpl->status, tid, toep->flags);
1303 return (0); /* Ignore negative advice */
1307 INP_INFO_RLOCK(&V_tcbinfo); /* for tcp_close */
1310 tp = intotcpcb(inp);
1313 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1314 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1315 inp->inp_flags, cpl->status);
1318 * If we'd initiated an abort earlier the reply to it is responsible for
1319 * cleaning up resources. Otherwise we tear everything down right here
1320 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
1322 if (toep->flags & TPF_ABORT_SHUTDOWN) {
1326 toep->flags |= TPF_ABORT_SHUTDOWN;
1328 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1329 struct socket *so = inp->inp_socket;
1332 so_error_set(so, abort_status_to_errno(tp,
1336 INP_WLOCK(inp); /* re-acquire */
1339 final_cpl_received(toep);
1341 INP_INFO_RUNLOCK(&V_tcbinfo);
1342 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1347 * Reply to the CPL_ABORT_REQ (send_reset)
1350 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1352 struct adapter *sc = iq->adapter;
1353 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1354 unsigned int tid = GET_TID(cpl);
1355 struct toepcb *toep = lookup_tid(sc, tid);
1356 struct inpcb *inp = toep->inp;
1358 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1361 KASSERT(opcode == CPL_ABORT_RPL_RSS,
1362 ("%s: unexpected opcode 0x%x", __func__, opcode));
1363 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1365 if (toep->flags & TPF_SYNQE)
1366 return (do_abort_rpl_synqe(iq, rss, m));
1368 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1370 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1371 __func__, tid, toep, inp, cpl->status);
1373 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1374 ("%s: wasn't expecting abort reply", __func__));
1377 final_cpl_received(toep);
1383 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1385 struct adapter *sc = iq->adapter;
1386 const struct cpl_rx_data *cpl = mtod(m, const void *);
1387 unsigned int tid = GET_TID(cpl);
1388 struct toepcb *toep = lookup_tid(sc, tid);
1389 struct inpcb *inp = toep->inp;
1394 uint32_t ddp_placed = 0;
1396 if (__predict_false(toep->flags & TPF_SYNQE)) {
1398 struct synq_entry *synqe = (void *)toep;
1400 INP_WLOCK(synqe->lctx->inp);
1401 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1402 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1403 ("%s: listen socket closed but tid %u not aborted.",
1407 * do_pass_accept_req is still running and will
1408 * eventually take care of this tid.
1411 INP_WUNLOCK(synqe->lctx->inp);
1413 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1419 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1421 /* strip off CPL header */
1422 m_adj(m, sizeof(*cpl));
1423 len = m->m_pkthdr.len;
1426 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1427 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1428 __func__, tid, len, inp->inp_flags);
1434 tp = intotcpcb(inp);
1436 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1437 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1440 if (tp->rcv_wnd < len) {
1441 KASSERT(toep->ulp_mode == ULP_MODE_RDMA,
1442 ("%s: negative window size", __func__));
1446 tp->t_rcvtime = ticks;
1448 so = inp_inpcbtosocket(inp);
1452 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1453 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1454 __func__, tid, len);
1459 INP_INFO_RLOCK(&V_tcbinfo);
1461 tp = tcp_drop(tp, ECONNRESET);
1464 INP_INFO_RUNLOCK(&V_tcbinfo);
1469 /* receive buffer autosize */
1470 CURVNET_SET(so->so_vnet);
1471 if (sb->sb_flags & SB_AUTOSIZE &&
1472 V_tcp_do_autorcvbuf &&
1473 sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1474 len > (sbspace(sb) / 8 * 7)) {
1475 unsigned int hiwat = sb->sb_hiwat;
1476 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1477 V_tcp_autorcvbuf_max);
1479 if (!sbreserve_locked(sb, newsize, so, NULL))
1480 sb->sb_flags &= ~SB_AUTOSIZE;
1482 toep->rx_credits += newsize - hiwat;
1485 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1486 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1489 if (toep->ddp_flags & DDP_SC_REQ)
1490 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1492 KASSERT(cpl->ddp_off == 1,
1493 ("%s: DDP switched on by itself.",
1496 /* Fell out of DDP mode */
1497 toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1501 insert_ddp_data(toep, ddp_placed);
1505 if ((toep->ddp_flags & DDP_OK) == 0 &&
1506 time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1507 toep->ddp_score = DDP_LOW_SCORE;
1508 toep->ddp_flags |= DDP_OK;
1509 CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1510 __func__, tid, time_uptime);
1513 if (toep->ddp_flags & DDP_ON) {
1516 * CPL_RX_DATA with DDP on can only be an indicate. Ask
1517 * soreceive to post a buffer or disable DDP. The
1518 * payload that arrived in this indicate is appended to
1519 * the socket buffer as usual.
1524 "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1525 __func__, tid, toep->flags, be32toh(cpl->seq), len);
1527 sb->sb_flags |= SB_DDP_INDICATE;
1528 } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1529 tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1532 * DDP allowed but isn't on (and a request to switch it
1533 * on isn't pending either), and conditions are ripe for
1534 * it to work. Switch it on.
1537 enable_ddp(sc, toep);
1541 KASSERT(toep->sb_cc >= sb->sb_cc,
1542 ("%s: sb %p has more data (%d) than last time (%d).",
1543 __func__, sb, sb->sb_cc, toep->sb_cc));
1544 toep->rx_credits += toep->sb_cc - sb->sb_cc;
1545 sbappendstream_locked(sb, m);
1546 toep->sb_cc = sb->sb_cc;
1547 if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1550 credits = send_rx_credits(sc, toep, toep->rx_credits);
1551 toep->rx_credits -= credits;
1552 tp->rcv_wnd += credits;
1553 tp->rcv_adv += credits;
1555 sorwakeup_locked(so);
1556 SOCKBUF_UNLOCK_ASSERT(sb);
1563 #define S_CPL_FW4_ACK_OPCODE 24
1564 #define M_CPL_FW4_ACK_OPCODE 0xff
1565 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1566 #define G_CPL_FW4_ACK_OPCODE(x) \
1567 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1569 #define S_CPL_FW4_ACK_FLOWID 0
1570 #define M_CPL_FW4_ACK_FLOWID 0xffffff
1571 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1572 #define G_CPL_FW4_ACK_FLOWID(x) \
1573 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1575 #define S_CPL_FW4_ACK_CR 24
1576 #define M_CPL_FW4_ACK_CR 0xff
1577 #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR)
1578 #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1580 #define S_CPL_FW4_ACK_SEQVAL 0
1581 #define M_CPL_FW4_ACK_SEQVAL 0x1
1582 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1583 #define G_CPL_FW4_ACK_SEQVAL(x) \
1584 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1585 #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U)
1588 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1590 struct adapter *sc = iq->adapter;
1591 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1592 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1593 struct toepcb *toep = lookup_tid(sc, tid);
1597 uint8_t credits = cpl->credits;
1598 struct ofld_tx_sdesc *txsd;
1601 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1605 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1606 * now this comes back carrying the credits for the flowc.
1608 if (__predict_false(toep->flags & TPF_SYNQE)) {
1609 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1610 ("%s: credits for a synq entry %p", __func__, toep));
1616 KASSERT(opcode == CPL_FW4_ACK,
1617 ("%s: unexpected opcode 0x%x", __func__, opcode));
1618 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1619 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1623 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1628 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1629 ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1631 tp = intotcpcb(inp);
1633 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1634 tcp_seq snd_una = be32toh(cpl->snd_una);
1637 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1639 "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1640 __func__, snd_una, toep->tid, tp->snd_una);
1644 if (tp->snd_una != snd_una) {
1645 tp->snd_una = snd_una;
1646 tp->ts_recent_age = tcp_ts_getticks();
1650 so = inp->inp_socket;
1651 txsd = &toep->txsd[toep->txsd_cidx];
1654 KASSERT(credits >= txsd->tx_credits,
1655 ("%s: too many (or partial) credits", __func__));
1656 credits -= txsd->tx_credits;
1657 toep->tx_credits += txsd->tx_credits;
1661 KASSERT(toep->txsd_avail <= toep->txsd_total,
1662 ("%s: txsd avail > total", __func__));
1663 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1664 txsd = &toep->txsd[0];
1665 toep->txsd_cidx = 0;
1669 if (toep->tx_credits == toep->tx_total) {
1670 toep->tx_nocompl = 0;
1671 toep->plen_nocompl = 0;
1674 if (toep->flags & TPF_TX_SUSPENDED &&
1675 toep->tx_credits >= toep->tx_total / 4) {
1676 toep->flags &= ~TPF_TX_SUSPENDED;
1677 if (toep->ulp_mode == ULP_MODE_ISCSI)
1678 t4_push_pdus(sc, toep, plen);
1680 t4_push_frames(sc, toep, plen);
1681 } else if (plen > 0) {
1682 struct sockbuf *sb = &so->so_snd;
1687 if (toep->ulp_mode == ULP_MODE_ISCSI) {
1689 if (__predict_false(sbu > 0)) {
1691 * The data trasmitted before the tid's ULP mode
1692 * changed to ISCSI is still in so_snd.
1693 * Incoming credits should account for so_snd
1696 sbdrop_locked(sb, min(sbu, plen));
1697 plen -= min(sbu, plen);
1699 sowwakeup_locked(so); /* unlocks so_snd */
1700 rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1702 sbdrop_locked(sb, plen);
1703 sowwakeup_locked(so); /* unlocks so_snd */
1705 SOCKBUF_UNLOCK_ASSERT(sb);
1714 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1716 struct adapter *sc = iq->adapter;
1717 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1718 unsigned int tid = GET_TID(cpl);
1720 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1723 KASSERT(opcode == CPL_SET_TCB_RPL,
1724 ("%s: unexpected opcode 0x%x", __func__, opcode));
1725 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1727 if (is_ftid(sc, tid))
1728 return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1731 * TOM and/or other ULPs don't request replies for CPL_SET_TCB or
1732 * CPL_SET_TCB_FIELD requests. This can easily change and when it does
1733 * the dispatch code will go here.
1736 panic("%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p", __func__,
1739 log(LOG_ERR, "%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p\n",
1747 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1748 uint16_t word, uint64_t mask, uint64_t val)
1751 struct cpl_set_tcb_field *req;
1753 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1756 panic("%s: allocation failure.", __func__);
1760 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1761 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1762 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1763 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1764 req->mask = htobe64(mask);
1765 req->val = htobe64(val);
1771 t4_init_cpl_io_handlers(struct adapter *sc)
1774 t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1775 t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1776 t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1777 t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1778 t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1779 t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1780 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1784 t4_uninit_cpl_io_handlers(struct adapter *sc)
1787 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);