2 * Copyright (c) 2012 Chelsio Communications, Inc.
4 * Written by: Navdeep Parhar <np@FreeBSD.org>
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
34 #include <sys/param.h>
35 #include <sys/types.h>
36 #include <sys/kernel.h>
38 #include <sys/module.h>
39 #include <sys/protosw.h>
40 #include <sys/domain.h>
41 #include <sys/socket.h>
42 #include <sys/socketvar.h>
43 #include <sys/sglist.h>
44 #include <netinet/in.h>
45 #include <netinet/in_pcb.h>
46 #include <netinet/ip.h>
47 #include <netinet/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 * For ULP connections HW may add headers, e.g., for digests, that aren't part
76 * of the messages sent by the host but that are part of the TCP payload and
77 * therefore consume TCP sequence space. Tx connection parameters that
78 * operate in TCP sequence space are affected by the HW additions and need to
79 * compensate for them to accurately track TCP sequence numbers. This array
80 * contains the compensating extra lengths for ULP packets. It is indexed by
81 * a packet's ULP submode.
83 const unsigned int t4_ulp_extra_len[] = {0, 4, 4, 8};
86 * Return the length of any HW additions that will be made to a Tx packet.
87 * Such additions can happen for some types of ULP packets.
89 static inline unsigned int
90 ulp_extra_len(struct mbuf *m, int *ulp_mode)
94 if ((mtag = m_tag_find(m, CXGBE_ISCSI_MBUF_TAG, NULL)) == NULL)
96 *ulp_mode = *((int *)(mtag + 1));
98 return (t4_ulp_extra_len[*ulp_mode & 3]);
102 send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp)
105 struct fw_flowc_wr *flowc;
106 unsigned int nparams = ftxp ? 8 : 6, flowclen;
107 struct port_info *pi = toep->port;
108 struct adapter *sc = pi->adapter;
109 unsigned int pfvf = G_FW_VIID_PFN(pi->viid) << S_FW_VIID_PFN;
110 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
112 KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
113 ("%s: flowc for tid %u sent already", __func__, toep->tid));
115 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
117 wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
120 panic("%s: allocation failure.", __func__);
123 memset(flowc, 0, wr->wr_len);
125 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
126 V_FW_FLOWC_WR_NPARAMS(nparams));
127 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
128 V_FW_WR_FLOWID(toep->tid));
130 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
131 flowc->mnemval[0].val = htobe32(pfvf);
132 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
133 flowc->mnemval[1].val = htobe32(pi->tx_chan);
134 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
135 flowc->mnemval[2].val = htobe32(pi->tx_chan);
136 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
137 flowc->mnemval[3].val = htobe32(toep->ofld_rxq->iq.abs_id);
139 uint32_t sndbuf = min(ftxp->snd_space, sc->tt.sndbuf);
141 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDNXT;
142 flowc->mnemval[4].val = htobe32(ftxp->snd_nxt);
143 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_RCVNXT;
144 flowc->mnemval[5].val = htobe32(ftxp->rcv_nxt);
145 flowc->mnemval[6].mnemonic = FW_FLOWC_MNEM_SNDBUF;
146 flowc->mnemval[6].val = htobe32(sndbuf);
147 flowc->mnemval[7].mnemonic = FW_FLOWC_MNEM_MSS;
148 flowc->mnemval[7].val = htobe32(ftxp->mss);
151 "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x",
152 __func__, toep->tid, ftxp->mss, sndbuf, ftxp->snd_nxt,
155 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF;
156 flowc->mnemval[4].val = htobe32(512);
157 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS;
158 flowc->mnemval[5].val = htobe32(512);
160 CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid);
163 txsd->tx_credits = howmany(flowclen, 16);
165 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
166 ("%s: not enough credits (%d)", __func__, toep->tx_credits));
167 toep->tx_credits -= txsd->tx_credits;
168 if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
172 toep->flags |= TPF_FLOWC_WR_SENT;
177 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
180 struct cpl_abort_req *req;
182 struct inpcb *inp = toep->inp;
183 struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */
185 INP_WLOCK_ASSERT(inp);
187 CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
189 inp->inp_flags & INP_DROPPED ? "inp dropped" :
190 tcpstates[tp->t_state],
191 toep->flags, inp->inp_flags,
192 toep->flags & TPF_ABORT_SHUTDOWN ?
193 " (abort already in progress)" : "");
195 if (toep->flags & TPF_ABORT_SHUTDOWN)
196 return; /* abort already in progress */
198 toep->flags |= TPF_ABORT_SHUTDOWN;
200 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
201 ("%s: flowc_wr not sent for tid %d.", __func__, tid));
203 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
206 panic("%s: allocation failure.", __func__);
210 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
211 if (inp->inp_flags & INP_DROPPED)
212 req->rsvd0 = htobe32(snd_nxt);
214 req->rsvd0 = htobe32(tp->snd_nxt);
215 req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
216 req->cmd = CPL_ABORT_SEND_RST;
219 * XXX: What's the correct way to tell that the inp hasn't been detached
220 * from its socket? Should I even be flushing the snd buffer here?
222 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
223 struct socket *so = inp->inp_socket;
225 if (so != NULL) /* because I'm not sure. See comment above */
226 sbflush(&so->so_snd);
229 t4_l2t_send(sc, wr, toep->l2te);
233 * Called when a connection is established to translate the TCP options
234 * reported by HW to FreeBSD's native format.
237 assign_rxopt(struct tcpcb *tp, unsigned int opt)
239 struct toepcb *toep = tp->t_toe;
240 struct inpcb *inp = tp->t_inpcb;
241 struct adapter *sc = td_adapter(toep->td);
244 INP_LOCK_ASSERT(inp);
246 if (inp->inp_inc.inc_flags & INC_ISIPV6)
247 n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
249 n = sizeof(struct ip) + sizeof(struct tcphdr);
250 tp->t_maxseg = tp->t_maxopd = sc->params.mtus[G_TCPOPT_MSS(opt)] - n;
252 CTR4(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u)", __func__, toep->tid,
253 G_TCPOPT_MSS(opt), sc->params.mtus[G_TCPOPT_MSS(opt)]);
255 if (G_TCPOPT_TSTAMP(opt)) {
256 tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */
257 tp->ts_recent = 0; /* hmmm */
258 tp->ts_recent_age = tcp_ts_getticks();
259 tp->t_maxseg -= TCPOLEN_TSTAMP_APPA;
262 if (G_TCPOPT_SACK(opt))
263 tp->t_flags |= TF_SACK_PERMIT; /* should already be set */
265 tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */
267 if (G_TCPOPT_WSCALE_OK(opt))
268 tp->t_flags |= TF_RCVD_SCALE;
270 /* Doing window scaling? */
271 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
272 (TF_RCVD_SCALE | TF_REQ_SCALE)) {
273 tp->rcv_scale = tp->request_r_scale;
274 tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
279 * Completes some final bits of initialization for just established connections
280 * and changes their state to TCPS_ESTABLISHED.
282 * The ISNs are from after the exchange of SYNs. i.e., the true ISN + 1.
285 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
288 struct inpcb *inp = toep->inp;
289 struct socket *so = inp->inp_socket;
290 struct tcpcb *tp = intotcpcb(inp);
292 uint32_t iss = be32toh(snd_isn) - 1; /* true ISS */
293 uint32_t irs = be32toh(rcv_isn) - 1; /* true IRS */
294 uint16_t tcpopt = be16toh(opt);
295 struct flowc_tx_params ftxp;
297 INP_WLOCK_ASSERT(inp);
298 KASSERT(tp->t_state == TCPS_SYN_SENT ||
299 tp->t_state == TCPS_SYN_RECEIVED,
300 ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
302 CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p",
303 __func__, toep->tid, toep, inp);
305 tp->t_state = TCPS_ESTABLISHED;
306 tp->t_starttime = ticks;
307 TCPSTAT_INC(tcps_connects);
311 tp->rcv_wnd = toep->rx_credits << 10;
312 tp->rcv_adv += tp->rcv_wnd;
313 tp->last_ack_sent = tp->rcv_nxt;
316 * If we were unable to send all rx credits via opt0, save the remainder
317 * in rx_credits so that they can be handed over with the next credit
320 SOCKBUF_LOCK(&so->so_rcv);
321 bufsize = select_rcv_wnd(so);
322 SOCKBUF_UNLOCK(&so->so_rcv);
323 toep->rx_credits = bufsize - tp->rcv_wnd;
327 tp->snd_una = iss + 1;
328 tp->snd_nxt = iss + 1;
329 tp->snd_max = iss + 1;
331 assign_rxopt(tp, tcpopt);
333 SOCKBUF_LOCK(&so->so_snd);
334 if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf)
335 bufsize = V_tcp_autosndbuf_max;
337 bufsize = sbspace(&so->so_snd);
338 SOCKBUF_UNLOCK(&so->so_snd);
340 ftxp.snd_nxt = tp->snd_nxt;
341 ftxp.rcv_nxt = tp->rcv_nxt;
342 ftxp.snd_space = bufsize;
343 ftxp.mss = tp->t_maxseg;
344 send_flowc_wr(toep, &ftxp);
350 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
353 struct cpl_rx_data_ack *req;
354 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
356 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
358 wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
363 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
364 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
371 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
373 struct adapter *sc = tod->tod_softc;
374 struct inpcb *inp = tp->t_inpcb;
375 struct socket *so = inp->inp_socket;
376 struct sockbuf *sb = &so->so_rcv;
377 struct toepcb *toep = tp->t_toe;
380 INP_WLOCK_ASSERT(inp);
383 KASSERT(toep->sb_cc >= sb->sb_cc,
384 ("%s: sb %p has more data (%d) than last time (%d).",
385 __func__, sb, sb->sb_cc, toep->sb_cc));
386 if (toep->ulp_mode == ULP_MODE_ISCSI) {
387 toep->rx_credits += toep->sb_cc;
390 toep->rx_credits += toep->sb_cc - sb->sb_cc;
391 toep->sb_cc = sb->sb_cc;
393 if (toep->rx_credits > 0 &&
394 (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
395 (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
396 toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
398 credits = send_rx_credits(sc, toep, toep->rx_credits);
399 toep->rx_credits -= credits;
400 tp->rcv_wnd += credits;
401 tp->rcv_adv += credits;
407 * Close a connection by sending a CPL_CLOSE_CON_REQ message.
410 close_conn(struct adapter *sc, struct toepcb *toep)
413 struct cpl_close_con_req *req;
414 unsigned int tid = toep->tid;
416 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
417 toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
419 if (toep->flags & TPF_FIN_SENT)
422 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
423 ("%s: flowc_wr not sent for tid %u.", __func__, tid));
425 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
428 panic("%s: allocation failure.", __func__);
432 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
433 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
434 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
435 V_FW_WR_FLOWID(tid));
436 req->wr.wr_lo = cpu_to_be64(0);
437 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
440 toep->flags |= TPF_FIN_SENT;
441 toep->flags &= ~TPF_SEND_FIN;
442 t4_l2t_send(sc, wr, toep->l2te);
447 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
448 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
450 /* Maximum amount of immediate data we could stuff in a WR */
452 max_imm_payload(int tx_credits)
454 const int n = 2; /* Use only up to 2 desc for imm. data WR */
456 KASSERT(tx_credits >= 0 &&
457 tx_credits <= MAX_OFLD_TX_CREDITS,
458 ("%s: %d credits", __func__, tx_credits));
460 if (tx_credits < MIN_OFLD_TX_CREDITS)
463 if (tx_credits >= (n * EQ_ESIZE) / 16)
464 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
466 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
469 /* Maximum number of SGL entries we could stuff in a WR */
471 max_dsgl_nsegs(int tx_credits)
473 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
474 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
476 KASSERT(tx_credits >= 0 &&
477 tx_credits <= MAX_OFLD_TX_CREDITS,
478 ("%s: %d credits", __func__, tx_credits));
480 if (tx_credits < MIN_OFLD_TX_CREDITS)
483 nseg += 2 * (sge_pair_credits * 16 / 24);
484 if ((sge_pair_credits * 16) % 24 == 16)
491 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
492 unsigned int plen, uint8_t credits, int shove, int ulp_mode, int txalign)
494 struct fw_ofld_tx_data_wr *txwr = dst;
495 unsigned int wr_ulp_mode;
497 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
498 V_FW_WR_IMMDLEN(immdlen));
499 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
500 V_FW_WR_LEN16(credits));
502 /* for iscsi, the mode & submode setting is per-packet */
503 if (toep->ulp_mode == ULP_MODE_ISCSI)
504 wr_ulp_mode = V_TX_ULP_MODE(ulp_mode >> 4) |
505 V_TX_ULP_SUBMODE(ulp_mode & 3);
507 wr_ulp_mode = V_TX_ULP_MODE(toep->ulp_mode);
509 txwr->lsodisable_to_flags = htobe32(wr_ulp_mode | V_TX_URG(0) | /*XXX*/
511 txwr->plen = htobe32(plen);
514 struct tcpcb *tp = intotcpcb(toep->inp);
516 if (plen < 2 * tp->t_maxseg || is_10G_port(toep->port))
517 txwr->lsodisable_to_flags |=
518 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
520 txwr->lsodisable_to_flags |=
521 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
522 (tp->t_flags & TF_NODELAY ? 0 :
523 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
528 * Generate a DSGL from a starting mbuf. The total number of segments and the
529 * maximum segments in any one mbuf are provided.
532 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
535 struct ulptx_sgl *usgl = dst;
538 struct sglist_seg segs[n];
540 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
542 sglist_init(&sg, n, segs);
543 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
544 V_ULPTX_NSGE(nsegs));
547 for (m = start; m != stop; m = m->m_next) {
548 rc = sglist_append(&sg, mtod(m, void *), m->m_len);
549 if (__predict_false(rc != 0))
550 panic("%s: sglist_append %d", __func__, rc);
552 for (j = 0; j < sg.sg_nseg; i++, j++) {
554 usgl->len0 = htobe32(segs[j].ss_len);
555 usgl->addr0 = htobe64(segs[j].ss_paddr);
557 usgl->sge[i / 2].len[i & 1] =
558 htobe32(segs[j].ss_len);
559 usgl->sge[i / 2].addr[i & 1] =
560 htobe64(segs[j].ss_paddr);
569 usgl->sge[i / 2].len[1] = htobe32(0);
570 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
571 __func__, nsegs, start, stop));
575 * Max number of SGL entries an offload tx work request can have. This is 41
576 * (1 + 40) for a full 512B work request.
577 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
579 #define OFLD_SGL_LEN (41)
582 * Send data and/or a FIN to the peer.
584 * The socket's so_snd buffer consists of a stream of data starting with sb_mb
585 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that
588 * drop indicates the number of bytes that should be dropped from the head of
589 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating
590 * contention on the send buffer lock (before this change it used to do
591 * sowwakeup and then t4_push_frames right after that when recovering from tx
592 * stalls). When drop is set this function MUST drop the bytes and wake up any
596 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
598 struct mbuf *sndptr, *m, *sb_sndptr;
599 struct fw_ofld_tx_data_wr *txwr;
601 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
602 struct inpcb *inp = toep->inp;
603 struct tcpcb *tp = intotcpcb(inp);
604 struct socket *so = inp->inp_socket;
605 struct sockbuf *sb = &so->so_snd;
606 int tx_credits, shove, compl, space, sowwakeup;
607 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
609 INP_WLOCK_ASSERT(inp);
610 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
611 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
613 KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
614 toep->ulp_mode == ULP_MODE_TCPDDP ||
615 toep->ulp_mode == ULP_MODE_RDMA,
616 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
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 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
681 if (space <= sb->sb_hiwat * 3 / 8 &&
682 toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
687 if (sb->sb_flags & SB_AUTOSIZE &&
688 V_tcp_do_autosndbuf &&
689 sb->sb_hiwat < V_tcp_autosndbuf_max &&
690 space < sb->sb_hiwat / 8) {
691 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
692 V_tcp_autosndbuf_max);
694 if (!sbreserve_locked(sb, newsize, so, NULL))
695 sb->sb_flags &= ~SB_AUTOSIZE;
697 sowwakeup = 1; /* room available */
700 sowwakeup_locked(so);
703 SOCKBUF_UNLOCK_ASSERT(sb);
705 /* nothing to send */
708 ("%s: nothing to send, but m != NULL", __func__));
712 if (__predict_false(toep->flags & TPF_FIN_SENT))
713 panic("%s: excess tx.", __func__);
715 if (plen <= max_imm) {
717 /* Immediate data tx */
719 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
722 /* XXX: how will we recover from this? */
723 toep->flags |= TPF_TX_SUSPENDED;
727 credits = howmany(wr->wr_len, 16);
728 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
730 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
737 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
738 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
739 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
741 /* XXX: how will we recover from this? */
742 toep->flags |= TPF_TX_SUSPENDED;
746 credits = howmany(wr_len, 16);
747 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
749 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
752 uint64_t *pad = (uint64_t *)
753 ((uintptr_t)txwr + wr_len);
758 KASSERT(toep->tx_credits >= credits,
759 ("%s: not enough credits", __func__));
761 toep->tx_credits -= credits;
762 toep->tx_nocompl += credits;
763 toep->plen_nocompl += plen;
764 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
765 toep->tx_nocompl >= toep->tx_total / 4)
768 if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
769 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
770 toep->tx_nocompl = 0;
771 toep->plen_nocompl = 0;
778 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
779 sb->sb_sndptr = sb_sndptr;
782 toep->flags |= TPF_TX_DATA_SENT;
783 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
784 toep->flags |= TPF_TX_SUSPENDED;
786 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
788 txsd->tx_credits = credits;
790 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
792 txsd = &toep->txsd[0];
796 t4_l2t_send(sc, wr, toep->l2te);
799 /* Send a FIN if requested, but only if there's no more data to send */
800 if (m == NULL && toep->flags & TPF_SEND_FIN)
801 close_conn(sc, toep);
804 /* Send ULP data over TOE using TX_DATA_WR. We send whole mbuf at once */
806 t4_ulp_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
808 struct mbuf *sndptr, *m = NULL;
809 struct fw_ofld_tx_data_wr *txwr;
811 unsigned int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
812 struct inpcb *inp = toep->inp;
816 int tx_credits, ulp_len = 0, ulp_mode = 0, qlen = 0;
818 struct ofld_tx_sdesc *txsd;
820 INP_WLOCK_ASSERT(inp);
821 if (toep->flags & TPF_ABORT_SHUTDOWN)
825 so = inp->inp_socket;
827 txsd = &toep->txsd[toep->txsd_pidx];
829 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
830 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
833 * This function doesn't resume by itself. Someone else must clear the
834 * flag and call this function.
836 if (__predict_false(toep->flags & TPF_TX_SUSPENDED))
839 sndptr = t4_queue_iscsi_callback(so, toep, 1, &qlen);
844 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
845 max_imm = max_imm_payload(tx_credits);
846 max_nsegs = max_dsgl_nsegs(tx_credits);
849 t4_cpl_iscsi_callback(toep->td, toep, &drop,
856 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
857 for (m = sndptr; m != NULL; m = m->m_next) {
858 int n = sglist_count(mtod(m, void *), m->m_len);
863 /* This mbuf sent us _over_ the nsegs limit, return */
864 if (plen > max_imm && nsegs > max_nsegs) {
865 toep->flags |= TPF_TX_SUSPENDED;
869 if (max_nsegs_1mbuf < n)
872 /* This mbuf put us right at the max_nsegs limit */
873 if (plen > max_imm && nsegs == max_nsegs) {
874 toep->flags |= TPF_TX_SUSPENDED;
879 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
880 /* nothing to send */
883 ("%s: nothing to send, but m != NULL", __func__));
887 if (__predict_false(toep->flags & TPF_FIN_SENT))
888 panic("%s: excess tx.", __func__);
890 ulp_len = plen + ulp_extra_len(sndptr, &ulp_mode);
891 if (plen <= max_imm) {
893 /* Immediate data tx */
894 wr = alloc_wrqe(roundup(sizeof(*txwr) + plen, 16),
897 /* XXX: how will we recover from this? */
898 toep->flags |= TPF_TX_SUSPENDED;
902 credits = howmany(wr->wr_len, 16);
903 write_tx_wr(txwr, toep, plen, ulp_len, credits, shove,
905 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(roundup(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, ulp_len, credits, shove,
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 toep->tx_credits -= credits;
935 toep->tx_nocompl += credits;
936 toep->plen_nocompl += plen;
937 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
938 toep->tx_nocompl >= toep->tx_total / 4)
942 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
943 toep->tx_nocompl = 0;
944 toep->plen_nocompl = 0;
946 tp->snd_nxt += ulp_len;
947 tp->snd_max += ulp_len;
950 sndptr = m = t4_queue_iscsi_callback(so, toep, 2, &qlen);
952 toep->flags |= TPF_TX_DATA_SENT;
953 if (toep->tx_credits < MIN_OFLD_TX_CREDITS) {
954 toep->flags |= TPF_TX_SUSPENDED;
957 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
959 txsd->tx_credits = credits;
961 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
963 txsd = &toep->txsd[0];
967 t4_l2t_send(sc, wr, toep->l2te);
970 /* Send a FIN if requested, but only if there's no more data to send */
971 if (m == NULL && toep->flags & TPF_SEND_FIN)
972 close_conn(sc, toep);
976 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
978 struct adapter *sc = tod->tod_softc;
980 struct inpcb *inp = tp->t_inpcb;
982 struct toepcb *toep = tp->t_toe;
984 INP_WLOCK_ASSERT(inp);
985 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
986 ("%s: inp %p dropped.", __func__, inp));
987 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
989 t4_push_frames(sc, toep, 0);
995 t4_send_fin(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 toep->flags |= TPF_SEND_FIN;
1009 if (tp->t_state >= TCPS_ESTABLISHED) {
1010 if (toep->ulp_mode == ULP_MODE_ISCSI)
1011 t4_ulp_push_frames(sc, toep, 0);
1013 t4_push_frames(sc, toep, 0);
1020 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1022 struct adapter *sc = tod->tod_softc;
1023 #if defined(INVARIANTS)
1024 struct inpcb *inp = tp->t_inpcb;
1026 struct toepcb *toep = tp->t_toe;
1028 INP_WLOCK_ASSERT(inp);
1029 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1030 ("%s: inp %p dropped.", __func__, inp));
1031 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1034 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1035 ("%s: flowc for tid %u [%s] not sent already",
1036 __func__, toep->tid, tcpstates[tp->t_state]));
1038 send_reset(sc, toep, 0);
1043 * Peer has sent us a FIN.
1046 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1048 struct adapter *sc = iq->adapter;
1049 const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1050 unsigned int tid = GET_TID(cpl);
1051 struct toepcb *toep = lookup_tid(sc, tid);
1052 struct inpcb *inp = toep->inp;
1053 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_WLOCK(&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;
1103 if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
1104 m = get_ddp_mbuf(be32toh(cpl->rcv_nxt) - tp->rcv_nxt);
1105 tp->rcv_nxt = be32toh(cpl->rcv_nxt);
1106 toep->ddp_flags &= ~(DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE);
1108 KASSERT(toep->sb_cc >= sb->sb_cc,
1109 ("%s: sb %p has more data (%d) than last time (%d).",
1110 __func__, sb, sb->sb_cc, toep->sb_cc));
1111 toep->rx_credits += toep->sb_cc - sb->sb_cc;
1112 #ifdef USE_DDP_RX_FLOW_CONTROL
1113 toep->rx_credits -= m->m_len; /* adjust for F_RX_FC_DDP */
1115 sbappendstream_locked(sb, m);
1116 toep->sb_cc = sb->sb_cc;
1118 socantrcvmore_locked(so); /* unlocks the sockbuf */
1120 if (toep->ulp_mode != ULP_MODE_RDMA) {
1121 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1122 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1123 be32toh(cpl->rcv_nxt)));
1126 switch (tp->t_state) {
1127 case TCPS_SYN_RECEIVED:
1128 tp->t_starttime = ticks;
1131 case TCPS_ESTABLISHED:
1132 tp->t_state = TCPS_CLOSE_WAIT;
1135 case TCPS_FIN_WAIT_1:
1136 tp->t_state = TCPS_CLOSING;
1139 case TCPS_FIN_WAIT_2:
1141 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1142 INP_INFO_WUNLOCK(&V_tcbinfo);
1145 final_cpl_received(toep);
1149 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1150 __func__, tid, tp->t_state);
1154 INP_INFO_WUNLOCK(&V_tcbinfo);
1159 * Peer has ACK'd our FIN.
1162 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1165 struct adapter *sc = iq->adapter;
1166 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1167 unsigned int tid = GET_TID(cpl);
1168 struct toepcb *toep = lookup_tid(sc, tid);
1169 struct inpcb *inp = toep->inp;
1170 struct tcpcb *tp = NULL;
1171 struct socket *so = NULL;
1173 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1176 KASSERT(opcode == CPL_CLOSE_CON_RPL,
1177 ("%s: unexpected opcode 0x%x", __func__, opcode));
1178 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1179 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1181 INP_INFO_WLOCK(&V_tcbinfo);
1183 tp = intotcpcb(inp);
1185 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1186 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1188 if (toep->flags & TPF_ABORT_SHUTDOWN)
1191 so = inp->inp_socket;
1192 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */
1194 switch (tp->t_state) {
1195 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */
1198 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1199 INP_INFO_WUNLOCK(&V_tcbinfo);
1202 final_cpl_received(toep); /* no more CPLs expected */
1210 case TCPS_FIN_WAIT_1:
1211 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1212 soisdisconnected(so);
1213 tp->t_state = TCPS_FIN_WAIT_2;
1218 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1219 __func__, tid, tcpstates[tp->t_state]);
1223 INP_INFO_WUNLOCK(&V_tcbinfo);
1228 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1232 struct cpl_abort_rpl *cpl;
1234 wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1237 panic("%s: allocation failure.", __func__);
1241 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1242 cpl->cmd = rst_status;
1248 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1250 switch (abort_reason) {
1251 case CPL_ERR_BAD_SYN:
1252 case CPL_ERR_CONN_RESET:
1253 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1254 case CPL_ERR_XMIT_TIMEDOUT:
1255 case CPL_ERR_PERSIST_TIMEDOUT:
1256 case CPL_ERR_FINWAIT2_TIMEDOUT:
1257 case CPL_ERR_KEEPALIVE_TIMEDOUT:
1265 cpl_not_handled(struct sge_iq *, const struct rss_header *, struct mbuf *);
1267 * tom_cpl_iscsi_callback -
1268 * iscsi and tom would share the following cpl messages, so when any of these
1269 * message is received, after tom is done with processing it, the messages
1270 * needs to be forwarded to iscsi for further processing:
1274 void (*tom_cpl_iscsi_callback)(struct tom_data *, struct socket *, void *,
1277 struct mbuf *(*tom_queue_iscsi_callback)(struct socket *, unsigned int, int *);
1279 * Check if the handler function is set for a given CPL
1280 * return 0 if the function is NULL or cpl_not_handled, 1 otherwise.
1283 t4tom_cpl_handler_registered(struct adapter *sc, unsigned int opcode)
1286 MPASS(opcode < nitems(sc->cpl_handler));
1288 return (sc->cpl_handler[opcode] &&
1289 sc->cpl_handler[opcode] != cpl_not_handled);
1293 * set the tom_cpl_iscsi_callback function, this function should be used
1294 * whenever both toe and iscsi need to process the same cpl msg.
1297 t4tom_register_cpl_iscsi_callback(void (*fp)(struct tom_data *, struct socket *,
1298 void *, unsigned int))
1301 tom_cpl_iscsi_callback = fp;
1305 t4tom_register_queue_iscsi_callback(struct mbuf *(*fp)(struct socket *,
1306 unsigned int, int *qlen))
1309 tom_queue_iscsi_callback = fp;
1313 t4_cpl_iscsi_callback(struct tom_data *td, struct toepcb *toep, void *m,
1314 unsigned int opcode)
1318 if (opcode == CPL_FW4_ACK)
1319 so = toep->inp->inp_socket;
1321 INP_WLOCK(toep->inp);
1322 so = toep->inp->inp_socket;
1323 INP_WUNLOCK(toep->inp);
1326 if (tom_cpl_iscsi_callback && so) {
1327 if (toep->ulp_mode == ULP_MODE_ISCSI) {
1328 tom_cpl_iscsi_callback(td, so, m, opcode);
1337 t4_queue_iscsi_callback(struct socket *so, struct toepcb *toep,
1338 unsigned int cmd, int *qlen)
1341 if (tom_queue_iscsi_callback && so) {
1342 if (toep->ulp_mode == ULP_MODE_ISCSI)
1343 return (tom_queue_iscsi_callback(so, cmd, qlen));
1350 * TCP RST from the peer, timeout, or some other such critical error.
1353 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1355 struct adapter *sc = iq->adapter;
1356 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1357 unsigned int tid = GET_TID(cpl);
1358 struct toepcb *toep = lookup_tid(sc, tid);
1359 struct sge_wrq *ofld_txq = toep->ofld_txq;
1363 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1366 KASSERT(opcode == CPL_ABORT_REQ_RSS,
1367 ("%s: unexpected opcode 0x%x", __func__, opcode));
1368 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1370 if (toep->flags & TPF_SYNQE)
1371 return (do_abort_req_synqe(iq, rss, m));
1373 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1375 if (negative_advice(cpl->status)) {
1376 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1377 __func__, cpl->status, tid, toep->flags);
1378 return (0); /* Ignore negative advice */
1382 INP_INFO_WLOCK(&V_tcbinfo); /* for tcp_close */
1385 tp = intotcpcb(inp);
1388 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1389 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1390 inp->inp_flags, cpl->status);
1393 * If we'd initiated an abort earlier the reply to it is responsible for
1394 * cleaning up resources. Otherwise we tear everything down right here
1395 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
1397 if (toep->flags & TPF_ABORT_SHUTDOWN) {
1401 toep->flags |= TPF_ABORT_SHUTDOWN;
1403 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1404 struct socket *so = inp->inp_socket;
1407 so_error_set(so, abort_status_to_errno(tp,
1411 INP_WLOCK(inp); /* re-acquire */
1414 final_cpl_received(toep);
1416 INP_INFO_WUNLOCK(&V_tcbinfo);
1417 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1422 * Reply to the CPL_ABORT_REQ (send_reset)
1425 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1427 struct adapter *sc = iq->adapter;
1428 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1429 unsigned int tid = GET_TID(cpl);
1430 struct toepcb *toep = lookup_tid(sc, tid);
1431 struct inpcb *inp = toep->inp;
1433 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1436 KASSERT(opcode == CPL_ABORT_RPL_RSS,
1437 ("%s: unexpected opcode 0x%x", __func__, opcode));
1438 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1440 if (toep->flags & TPF_SYNQE)
1441 return (do_abort_rpl_synqe(iq, rss, m));
1443 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1445 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1446 __func__, tid, toep, inp, cpl->status);
1448 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1449 ("%s: wasn't expecting abort reply", __func__));
1452 final_cpl_received(toep);
1458 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1460 struct adapter *sc = iq->adapter;
1461 const struct cpl_rx_data *cpl = mtod(m, const void *);
1462 unsigned int tid = GET_TID(cpl);
1463 struct toepcb *toep = lookup_tid(sc, tid);
1464 struct inpcb *inp = toep->inp;
1469 uint32_t ddp_placed = 0;
1471 if (__predict_false(toep->flags & TPF_SYNQE)) {
1473 struct synq_entry *synqe = (void *)toep;
1475 INP_WLOCK(synqe->lctx->inp);
1476 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1477 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1478 ("%s: listen socket closed but tid %u not aborted.",
1482 * do_pass_accept_req is still running and will
1483 * eventually take care of this tid.
1486 INP_WUNLOCK(synqe->lctx->inp);
1488 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1494 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1496 /* strip off CPL header */
1497 m_adj(m, sizeof(*cpl));
1498 len = m->m_pkthdr.len;
1501 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1502 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1503 __func__, tid, len, inp->inp_flags);
1509 tp = intotcpcb(inp);
1511 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1512 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1515 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
1517 tp->t_rcvtime = ticks;
1519 so = inp_inpcbtosocket(inp);
1523 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1524 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1525 __func__, tid, len);
1530 INP_INFO_WLOCK(&V_tcbinfo);
1532 tp = tcp_drop(tp, ECONNRESET);
1535 INP_INFO_WUNLOCK(&V_tcbinfo);
1540 /* receive buffer autosize */
1541 if (sb->sb_flags & SB_AUTOSIZE &&
1542 V_tcp_do_autorcvbuf &&
1543 sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1544 len > (sbspace(sb) / 8 * 7)) {
1545 unsigned int hiwat = sb->sb_hiwat;
1546 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1547 V_tcp_autorcvbuf_max);
1549 if (!sbreserve_locked(sb, newsize, so, NULL))
1550 sb->sb_flags &= ~SB_AUTOSIZE;
1552 toep->rx_credits += newsize - hiwat;
1555 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1556 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1559 if (toep->ddp_flags & DDP_SC_REQ)
1560 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1562 KASSERT(cpl->ddp_off == 1,
1563 ("%s: DDP switched on by itself.",
1566 /* Fell out of DDP mode */
1567 toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1571 insert_ddp_data(toep, ddp_placed);
1575 if ((toep->ddp_flags & DDP_OK) == 0 &&
1576 time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1577 toep->ddp_score = DDP_LOW_SCORE;
1578 toep->ddp_flags |= DDP_OK;
1579 CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1580 __func__, tid, time_uptime);
1583 if (toep->ddp_flags & DDP_ON) {
1586 * CPL_RX_DATA with DDP on can only be an indicate. Ask
1587 * soreceive to post a buffer or disable DDP. The
1588 * payload that arrived in this indicate is appended to
1589 * the socket buffer as usual.
1594 "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1595 __func__, tid, toep->flags, be32toh(cpl->seq), len);
1597 sb->sb_flags |= SB_DDP_INDICATE;
1598 } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1599 tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1602 * DDP allowed but isn't on (and a request to switch it
1603 * on isn't pending either), and conditions are ripe for
1604 * it to work. Switch it on.
1607 enable_ddp(sc, toep);
1611 KASSERT(toep->sb_cc >= sb->sb_cc,
1612 ("%s: sb %p has more data (%d) than last time (%d).",
1613 __func__, sb, sb->sb_cc, toep->sb_cc));
1614 toep->rx_credits += toep->sb_cc - sb->sb_cc;
1615 sbappendstream_locked(sb, m);
1616 toep->sb_cc = sb->sb_cc;
1617 if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1620 credits = send_rx_credits(sc, toep, toep->rx_credits);
1621 toep->rx_credits -= credits;
1622 tp->rcv_wnd += credits;
1623 tp->rcv_adv += credits;
1625 sorwakeup_locked(so);
1626 SOCKBUF_UNLOCK_ASSERT(sb);
1632 #define S_CPL_FW4_ACK_OPCODE 24
1633 #define M_CPL_FW4_ACK_OPCODE 0xff
1634 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1635 #define G_CPL_FW4_ACK_OPCODE(x) \
1636 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1638 #define S_CPL_FW4_ACK_FLOWID 0
1639 #define M_CPL_FW4_ACK_FLOWID 0xffffff
1640 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1641 #define G_CPL_FW4_ACK_FLOWID(x) \
1642 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1644 #define S_CPL_FW4_ACK_CR 24
1645 #define M_CPL_FW4_ACK_CR 0xff
1646 #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR)
1647 #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1649 #define S_CPL_FW4_ACK_SEQVAL 0
1650 #define M_CPL_FW4_ACK_SEQVAL 0x1
1651 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1652 #define G_CPL_FW4_ACK_SEQVAL(x) \
1653 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1654 #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U)
1657 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1659 struct adapter *sc = iq->adapter;
1660 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1661 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1662 struct toepcb *toep = lookup_tid(sc, tid);
1666 uint8_t credits = cpl->credits;
1667 struct ofld_tx_sdesc *txsd;
1670 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1674 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1675 * now this comes back carrying the credits for the flowc.
1677 if (__predict_false(toep->flags & TPF_SYNQE)) {
1678 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1679 ("%s: credits for a synq entry %p", __func__, toep));
1685 KASSERT(opcode == CPL_FW4_ACK,
1686 ("%s: unexpected opcode 0x%x", __func__, opcode));
1687 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1688 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1692 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1697 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1698 ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1700 tp = intotcpcb(inp);
1702 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1703 tcp_seq snd_una = be32toh(cpl->snd_una);
1706 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1708 "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1709 __func__, snd_una, toep->tid, tp->snd_una);
1713 if (tp->snd_una != snd_una) {
1714 tp->snd_una = snd_una;
1715 tp->ts_recent_age = tcp_ts_getticks();
1719 so = inp->inp_socket;
1720 txsd = &toep->txsd[toep->txsd_cidx];
1723 KASSERT(credits >= txsd->tx_credits,
1724 ("%s: too many (or partial) credits", __func__));
1725 credits -= txsd->tx_credits;
1726 toep->tx_credits += txsd->tx_credits;
1730 KASSERT(toep->txsd_avail <= toep->txsd_total,
1731 ("%s: txsd avail > total", __func__));
1732 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1733 txsd = &toep->txsd[0];
1734 toep->txsd_cidx = 0;
1738 if (toep->tx_credits == toep->tx_total) {
1739 toep->tx_nocompl = 0;
1740 toep->plen_nocompl = 0;
1743 if (toep->flags & TPF_TX_SUSPENDED &&
1744 toep->tx_credits >= toep->tx_total / 4) {
1745 toep->flags &= ~TPF_TX_SUSPENDED;
1746 if (toep->ulp_mode == ULP_MODE_ISCSI)
1747 t4_ulp_push_frames(sc, toep, plen);
1749 t4_push_frames(sc, toep, plen);
1750 } else if (plen > 0) {
1751 struct sockbuf *sb = &so->so_snd;
1753 if (toep->ulp_mode == ULP_MODE_ISCSI)
1754 t4_cpl_iscsi_callback(toep->td, toep, &plen,
1758 sbdrop_locked(sb, plen);
1759 sowwakeup_locked(so);
1760 SOCKBUF_UNLOCK_ASSERT(sb);
1770 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1772 struct adapter *sc = iq->adapter;
1773 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1774 unsigned int tid = GET_TID(cpl);
1776 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1779 KASSERT(opcode == CPL_SET_TCB_RPL,
1780 ("%s: unexpected opcode 0x%x", __func__, opcode));
1781 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1783 if (is_ftid(sc, tid))
1784 return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1786 struct toepcb *toep = lookup_tid(sc, tid);
1788 t4_cpl_iscsi_callback(toep->td, toep, m, CPL_SET_TCB_RPL);
1792 CXGBE_UNIMPLEMENTED(__func__);
1796 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1797 uint16_t word, uint64_t mask, uint64_t val)
1800 struct cpl_set_tcb_field *req;
1802 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1805 panic("%s: allocation failure.", __func__);
1809 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1810 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1811 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1812 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1813 req->mask = htobe64(mask);
1814 req->val = htobe64(val);
1820 t4_init_cpl_io_handlers(struct adapter *sc)
1823 t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1824 t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1825 t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1826 t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1827 t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1828 t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1829 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1833 t4_uninit_cpl_io_handlers(struct adapter *sc)
1836 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);