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 credits = toep->rx_credits;
397 (credits + 16384 >= tp->rcv_wnd || credits >= 15 * 1024)) {
399 credits = send_rx_credits(sc, toep, credits);
401 toep->rx_credits -= credits;
403 tp->rcv_wnd += credits;
404 tp->rcv_adv += credits;
409 * Close a connection by sending a CPL_CLOSE_CON_REQ message.
412 close_conn(struct adapter *sc, struct toepcb *toep)
415 struct cpl_close_con_req *req;
416 unsigned int tid = toep->tid;
418 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
419 toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
421 if (toep->flags & TPF_FIN_SENT)
424 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
425 ("%s: flowc_wr not sent for tid %u.", __func__, tid));
427 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
430 panic("%s: allocation failure.", __func__);
434 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
435 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
436 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
437 V_FW_WR_FLOWID(tid));
438 req->wr.wr_lo = cpu_to_be64(0);
439 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
442 toep->flags |= TPF_FIN_SENT;
443 toep->flags &= ~TPF_SEND_FIN;
444 t4_l2t_send(sc, wr, toep->l2te);
449 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
450 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
452 /* Maximum amount of immediate data we could stuff in a WR */
454 max_imm_payload(int tx_credits)
456 const int n = 2; /* Use only up to 2 desc for imm. data WR */
458 KASSERT(tx_credits >= 0 &&
459 tx_credits <= MAX_OFLD_TX_CREDITS,
460 ("%s: %d credits", __func__, tx_credits));
462 if (tx_credits < MIN_OFLD_TX_CREDITS)
465 if (tx_credits >= (n * EQ_ESIZE) / 16)
466 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
468 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
471 /* Maximum number of SGL entries we could stuff in a WR */
473 max_dsgl_nsegs(int tx_credits)
475 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
476 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
478 KASSERT(tx_credits >= 0 &&
479 tx_credits <= MAX_OFLD_TX_CREDITS,
480 ("%s: %d credits", __func__, tx_credits));
482 if (tx_credits < MIN_OFLD_TX_CREDITS)
485 nseg += 2 * (sge_pair_credits * 16 / 24);
486 if ((sge_pair_credits * 16) % 24 == 16)
493 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
494 unsigned int plen, uint8_t credits, int shove, int ulp_mode, int txalign)
496 struct fw_ofld_tx_data_wr *txwr = dst;
497 unsigned int wr_ulp_mode;
499 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
500 V_FW_WR_IMMDLEN(immdlen));
501 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
502 V_FW_WR_LEN16(credits));
504 /* for iscsi, the mode & submode setting is per-packet */
505 if (toep->ulp_mode == ULP_MODE_ISCSI)
506 wr_ulp_mode = V_FW_OFLD_TX_DATA_WR_ULPMODE(ulp_mode >> 4) |
507 V_FW_OFLD_TX_DATA_WR_ULPSUBMODE(ulp_mode & 3);
509 wr_ulp_mode = V_FW_OFLD_TX_DATA_WR_ULPMODE(toep->ulp_mode);
511 txwr->lsodisable_to_proxy =
512 htobe32(wr_ulp_mode |
513 V_FW_OFLD_TX_DATA_WR_URGENT(0) | /* XXX */
514 V_FW_OFLD_TX_DATA_WR_SHOVE(shove));
515 txwr->plen = htobe32(plen);
518 struct tcpcb *tp = intotcpcb(toep->inp);
520 if (plen < 2 * tp->t_maxseg || is_10G_port(toep->port))
521 txwr->lsodisable_to_proxy |=
522 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
524 txwr->lsodisable_to_proxy |=
525 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
526 (tp->t_flags & TF_NODELAY ? 0 :
527 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
532 * Generate a DSGL from a starting mbuf. The total number of segments and the
533 * maximum segments in any one mbuf are provided.
536 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
539 struct ulptx_sgl *usgl = dst;
542 struct sglist_seg segs[n];
544 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
546 sglist_init(&sg, n, segs);
547 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
548 V_ULPTX_NSGE(nsegs));
551 for (m = start; m != stop; m = m->m_next) {
552 rc = sglist_append(&sg, mtod(m, void *), m->m_len);
553 if (__predict_false(rc != 0))
554 panic("%s: sglist_append %d", __func__, rc);
556 for (j = 0; j < sg.sg_nseg; i++, j++) {
558 usgl->len0 = htobe32(segs[j].ss_len);
559 usgl->addr0 = htobe64(segs[j].ss_paddr);
561 usgl->sge[i / 2].len[i & 1] =
562 htobe32(segs[j].ss_len);
563 usgl->sge[i / 2].addr[i & 1] =
564 htobe64(segs[j].ss_paddr);
573 usgl->sge[i / 2].len[1] = htobe32(0);
574 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
575 __func__, nsegs, start, stop));
579 * Max number of SGL entries an offload tx work request can have. This is 41
580 * (1 + 40) for a full 512B work request.
581 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
583 #define OFLD_SGL_LEN (41)
586 * Send data and/or a FIN to the peer.
588 * The socket's so_snd buffer consists of a stream of data starting with sb_mb
589 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that
592 * drop indicates the number of bytes that should be dropped from the head of
593 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating
594 * contention on the send buffer lock (before this change it used to do
595 * sowwakeup and then t4_push_frames right after that when recovering from tx
596 * stalls). When drop is set this function MUST drop the bytes and wake up any
600 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
602 struct mbuf *sndptr, *m, *sb_sndptr;
603 struct fw_ofld_tx_data_wr *txwr;
605 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
606 struct inpcb *inp = toep->inp;
607 struct tcpcb *tp = intotcpcb(inp);
608 struct socket *so = inp->inp_socket;
609 struct sockbuf *sb = &so->so_snd;
610 int tx_credits, shove, compl, space, sowwakeup;
611 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
613 INP_WLOCK_ASSERT(inp);
614 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
615 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
617 KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
618 toep->ulp_mode == ULP_MODE_TCPDDP ||
619 toep->ulp_mode == ULP_MODE_RDMA,
620 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
623 * This function doesn't resume by itself. Someone else must clear the
624 * flag and call this function.
626 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
628 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
633 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
634 max_imm = max_imm_payload(tx_credits);
635 max_nsegs = max_dsgl_nsegs(tx_credits);
640 sbdrop_locked(sb, drop);
643 sb_sndptr = sb->sb_sndptr;
644 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
647 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
648 for (m = sndptr; m != NULL; m = m->m_next) {
649 int n = sglist_count(mtod(m, void *), m->m_len);
654 /* This mbuf sent us _over_ the nsegs limit, back out */
655 if (plen > max_imm && nsegs > max_nsegs) {
659 /* Too few credits */
660 toep->flags |= TPF_TX_SUSPENDED;
662 sowwakeup_locked(so);
665 SOCKBUF_UNLOCK_ASSERT(sb);
671 if (max_nsegs_1mbuf < n)
673 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */
675 /* This mbuf put us right at the max_nsegs limit */
676 if (plen > max_imm && nsegs == max_nsegs) {
682 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
685 if (space <= sb->sb_hiwat * 3 / 8 &&
686 toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
691 if (sb->sb_flags & SB_AUTOSIZE &&
692 V_tcp_do_autosndbuf &&
693 sb->sb_hiwat < V_tcp_autosndbuf_max &&
694 space < sb->sb_hiwat / 8) {
695 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
696 V_tcp_autosndbuf_max);
698 if (!sbreserve_locked(sb, newsize, so, NULL))
699 sb->sb_flags &= ~SB_AUTOSIZE;
701 sowwakeup = 1; /* room available */
704 sowwakeup_locked(so);
707 SOCKBUF_UNLOCK_ASSERT(sb);
709 /* nothing to send */
712 ("%s: nothing to send, but m != NULL", __func__));
716 if (__predict_false(toep->flags & TPF_FIN_SENT))
717 panic("%s: excess tx.", __func__);
719 if (plen <= max_imm) {
721 /* Immediate data tx */
723 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
726 /* XXX: how will we recover from this? */
727 toep->flags |= TPF_TX_SUSPENDED;
731 credits = howmany(wr->wr_len, 16);
732 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
734 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
741 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
742 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
743 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
745 /* XXX: how will we recover from this? */
746 toep->flags |= TPF_TX_SUSPENDED;
750 credits = howmany(wr_len, 16);
751 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
753 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
756 uint64_t *pad = (uint64_t *)
757 ((uintptr_t)txwr + wr_len);
762 KASSERT(toep->tx_credits >= credits,
763 ("%s: not enough credits", __func__));
765 toep->tx_credits -= credits;
766 toep->tx_nocompl += credits;
767 toep->plen_nocompl += plen;
768 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
769 toep->tx_nocompl >= toep->tx_total / 4)
773 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
774 toep->tx_nocompl = 0;
775 toep->plen_nocompl = 0;
782 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
783 sb->sb_sndptr = sb_sndptr;
786 toep->flags |= TPF_TX_DATA_SENT;
787 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
788 toep->flags |= TPF_TX_SUSPENDED;
790 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
792 txsd->tx_credits = credits;
794 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
796 txsd = &toep->txsd[0];
800 t4_l2t_send(sc, wr, toep->l2te);
803 /* Send a FIN if requested, but only if there's no more data to send */
804 if (m == NULL && toep->flags & TPF_SEND_FIN)
805 close_conn(sc, toep);
808 /* Send ULP data over TOE using TX_DATA_WR. We send whole mbuf at once */
810 t4_ulp_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
812 struct mbuf *sndptr, *m = NULL;
813 struct fw_ofld_tx_data_wr *txwr;
815 unsigned int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
816 struct inpcb *inp = toep->inp;
820 int tx_credits, ulp_len = 0, ulp_mode = 0, qlen = 0;
822 struct ofld_tx_sdesc *txsd;
824 INP_WLOCK_ASSERT(inp);
825 if (toep->flags & TPF_ABORT_SHUTDOWN)
829 so = inp->inp_socket;
831 txsd = &toep->txsd[toep->txsd_pidx];
833 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
834 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
837 * This function doesn't resume by itself. Someone else must clear the
838 * flag and call this function.
840 if (__predict_false(toep->flags & TPF_TX_SUSPENDED))
843 sndptr = t4_queue_iscsi_callback(so, toep, 1, &qlen);
848 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
849 max_imm = max_imm_payload(tx_credits);
850 max_nsegs = max_dsgl_nsegs(tx_credits);
853 t4_cpl_iscsi_callback(toep->td, toep, &drop,
860 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
861 for (m = sndptr; m != NULL; m = m->m_next) {
862 int n = sglist_count(mtod(m, void *), m->m_len);
867 /* This mbuf sent us _over_ the nsegs limit, return */
868 if (plen > max_imm && nsegs > max_nsegs) {
869 toep->flags |= TPF_TX_SUSPENDED;
873 if (max_nsegs_1mbuf < n)
876 /* This mbuf put us right at the max_nsegs limit */
877 if (plen > max_imm && nsegs == max_nsegs) {
878 toep->flags |= TPF_TX_SUSPENDED;
883 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
884 /* nothing to send */
887 ("%s: nothing to send, but m != NULL", __func__));
891 if (__predict_false(toep->flags & TPF_FIN_SENT))
892 panic("%s: excess tx.", __func__);
894 ulp_len = plen + ulp_extra_len(sndptr, &ulp_mode);
895 if (plen <= max_imm) {
897 /* Immediate data tx */
898 wr = alloc_wrqe(roundup(sizeof(*txwr) + plen, 16),
901 /* XXX: how will we recover from this? */
902 toep->flags |= TPF_TX_SUSPENDED;
906 credits = howmany(wr->wr_len, 16);
907 write_tx_wr(txwr, toep, plen, ulp_len, credits, shove,
909 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
914 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
915 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
916 wr = alloc_wrqe(roundup(wr_len, 16), toep->ofld_txq);
918 /* XXX: how will we recover from this? */
919 toep->flags |= TPF_TX_SUSPENDED;
923 credits = howmany(wr_len, 16);
924 write_tx_wr(txwr, toep, 0, ulp_len, credits, shove,
926 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
929 uint64_t *pad = (uint64_t *)
930 ((uintptr_t)txwr + wr_len);
935 KASSERT(toep->tx_credits >= credits,
936 ("%s: not enough credits", __func__));
938 toep->tx_credits -= credits;
939 toep->tx_nocompl += credits;
940 toep->plen_nocompl += plen;
941 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
942 toep->tx_nocompl >= toep->tx_total / 4)
946 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
947 toep->tx_nocompl = 0;
948 toep->plen_nocompl = 0;
950 tp->snd_nxt += ulp_len;
951 tp->snd_max += ulp_len;
954 sndptr = m = t4_queue_iscsi_callback(so, toep, 2, &qlen);
956 toep->flags |= TPF_TX_DATA_SENT;
957 if (toep->tx_credits < MIN_OFLD_TX_CREDITS) {
958 toep->flags |= TPF_TX_SUSPENDED;
961 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
963 txsd->tx_credits = credits;
965 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
967 txsd = &toep->txsd[0];
971 t4_l2t_send(sc, wr, toep->l2te);
974 /* Send a FIN if requested, but only if there's no more data to send */
975 if (m == NULL && toep->flags & TPF_SEND_FIN)
976 close_conn(sc, toep);
980 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
982 struct adapter *sc = tod->tod_softc;
984 struct inpcb *inp = tp->t_inpcb;
986 struct toepcb *toep = tp->t_toe;
988 INP_WLOCK_ASSERT(inp);
989 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
990 ("%s: inp %p dropped.", __func__, inp));
991 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
993 t4_push_frames(sc, toep, 0);
999 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
1001 struct adapter *sc = tod->tod_softc;
1003 struct inpcb *inp = tp->t_inpcb;
1005 struct toepcb *toep = tp->t_toe;
1007 INP_WLOCK_ASSERT(inp);
1008 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1009 ("%s: inp %p dropped.", __func__, inp));
1010 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1012 toep->flags |= TPF_SEND_FIN;
1013 if (tp->t_state >= TCPS_ESTABLISHED) {
1014 if (toep->ulp_mode == ULP_MODE_ISCSI)
1015 t4_ulp_push_frames(sc, toep, 0);
1017 t4_push_frames(sc, toep, 0);
1024 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1026 struct adapter *sc = tod->tod_softc;
1027 #if defined(INVARIANTS)
1028 struct inpcb *inp = tp->t_inpcb;
1030 struct toepcb *toep = tp->t_toe;
1032 INP_WLOCK_ASSERT(inp);
1033 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1034 ("%s: inp %p dropped.", __func__, inp));
1035 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1038 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1039 ("%s: flowc for tid %u [%s] not sent already",
1040 __func__, toep->tid, tcpstates[tp->t_state]));
1042 send_reset(sc, toep, 0);
1047 * Peer has sent us a FIN.
1050 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1052 struct adapter *sc = iq->adapter;
1053 const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1054 unsigned int tid = GET_TID(cpl);
1055 struct toepcb *toep = lookup_tid(sc, tid);
1056 struct inpcb *inp = toep->inp;
1057 struct tcpcb *tp = NULL;
1061 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1064 KASSERT(opcode == CPL_PEER_CLOSE,
1065 ("%s: unexpected opcode 0x%x", __func__, opcode));
1066 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1068 if (__predict_false(toep->flags & TPF_SYNQE)) {
1070 struct synq_entry *synqe = (void *)toep;
1072 INP_WLOCK(synqe->lctx->inp);
1073 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1074 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1075 ("%s: listen socket closed but tid %u not aborted.",
1079 * do_pass_accept_req is still running and will
1080 * eventually take care of this tid.
1083 INP_WUNLOCK(synqe->lctx->inp);
1085 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1090 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1092 INP_INFO_WLOCK(&V_tcbinfo);
1094 tp = intotcpcb(inp);
1096 CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1097 tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1099 if (toep->flags & TPF_ABORT_SHUTDOWN)
1102 tp->rcv_nxt++; /* FIN */
1104 so = inp->inp_socket;
1107 if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
1108 m = get_ddp_mbuf(be32toh(cpl->rcv_nxt) - tp->rcv_nxt);
1109 tp->rcv_nxt = be32toh(cpl->rcv_nxt);
1110 toep->ddp_flags &= ~(DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE);
1112 KASSERT(toep->sb_cc >= sb->sb_cc,
1113 ("%s: sb %p has more data (%d) than last time (%d).",
1114 __func__, sb, sb->sb_cc, toep->sb_cc));
1115 toep->rx_credits += toep->sb_cc - sb->sb_cc;
1116 #ifdef USE_DDP_RX_FLOW_CONTROL
1117 toep->rx_credits -= m->m_len; /* adjust for F_RX_FC_DDP */
1119 sbappendstream_locked(sb, m);
1120 toep->sb_cc = sb->sb_cc;
1122 socantrcvmore_locked(so); /* unlocks the sockbuf */
1124 if (toep->ulp_mode != ULP_MODE_RDMA) {
1125 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1126 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1127 be32toh(cpl->rcv_nxt)));
1130 switch (tp->t_state) {
1131 case TCPS_SYN_RECEIVED:
1132 tp->t_starttime = ticks;
1135 case TCPS_ESTABLISHED:
1136 tp->t_state = TCPS_CLOSE_WAIT;
1139 case TCPS_FIN_WAIT_1:
1140 tp->t_state = TCPS_CLOSING;
1143 case TCPS_FIN_WAIT_2:
1145 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1146 INP_INFO_WUNLOCK(&V_tcbinfo);
1149 final_cpl_received(toep);
1153 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1154 __func__, tid, tp->t_state);
1158 INP_INFO_WUNLOCK(&V_tcbinfo);
1163 * Peer has ACK'd our FIN.
1166 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1169 struct adapter *sc = iq->adapter;
1170 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1171 unsigned int tid = GET_TID(cpl);
1172 struct toepcb *toep = lookup_tid(sc, tid);
1173 struct inpcb *inp = toep->inp;
1174 struct tcpcb *tp = NULL;
1175 struct socket *so = NULL;
1177 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1180 KASSERT(opcode == CPL_CLOSE_CON_RPL,
1181 ("%s: unexpected opcode 0x%x", __func__, opcode));
1182 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1183 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1185 INP_INFO_WLOCK(&V_tcbinfo);
1187 tp = intotcpcb(inp);
1189 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1190 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1192 if (toep->flags & TPF_ABORT_SHUTDOWN)
1195 so = inp->inp_socket;
1196 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */
1198 switch (tp->t_state) {
1199 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */
1202 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1203 INP_INFO_WUNLOCK(&V_tcbinfo);
1206 final_cpl_received(toep); /* no more CPLs expected */
1214 case TCPS_FIN_WAIT_1:
1215 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1216 soisdisconnected(so);
1217 tp->t_state = TCPS_FIN_WAIT_2;
1222 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1223 __func__, tid, tcpstates[tp->t_state]);
1227 INP_INFO_WUNLOCK(&V_tcbinfo);
1232 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1236 struct cpl_abort_rpl *cpl;
1238 wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1241 panic("%s: allocation failure.", __func__);
1245 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1246 cpl->cmd = rst_status;
1252 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1254 switch (abort_reason) {
1255 case CPL_ERR_BAD_SYN:
1256 case CPL_ERR_CONN_RESET:
1257 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1258 case CPL_ERR_XMIT_TIMEDOUT:
1259 case CPL_ERR_PERSIST_TIMEDOUT:
1260 case CPL_ERR_FINWAIT2_TIMEDOUT:
1261 case CPL_ERR_KEEPALIVE_TIMEDOUT:
1269 cpl_not_handled(struct sge_iq *, const struct rss_header *, struct mbuf *);
1271 * tom_cpl_iscsi_callback -
1272 * iscsi and tom would share the following cpl messages, so when any of these
1273 * message is received, after tom is done with processing it, the messages
1274 * needs to be forwarded to iscsi for further processing:
1278 void (*tom_cpl_iscsi_callback)(struct tom_data *, struct socket *, void *,
1281 struct mbuf *(*tom_queue_iscsi_callback)(struct socket *, unsigned int, int *);
1283 * Check if the handler function is set for a given CPL
1284 * return 0 if the function is NULL or cpl_not_handled, 1 otherwise.
1287 t4tom_cpl_handler_registered(struct adapter *sc, unsigned int opcode)
1290 MPASS(opcode < nitems(sc->cpl_handler));
1292 return (sc->cpl_handler[opcode] &&
1293 sc->cpl_handler[opcode] != cpl_not_handled);
1297 * set the tom_cpl_iscsi_callback function, this function should be used
1298 * whenever both toe and iscsi need to process the same cpl msg.
1301 t4tom_register_cpl_iscsi_callback(void (*fp)(struct tom_data *, struct socket *,
1302 void *, unsigned int))
1305 tom_cpl_iscsi_callback = fp;
1309 t4tom_register_queue_iscsi_callback(struct mbuf *(*fp)(struct socket *,
1310 unsigned int, int *qlen))
1313 tom_queue_iscsi_callback = fp;
1317 t4_cpl_iscsi_callback(struct tom_data *td, struct toepcb *toep, void *m,
1318 unsigned int opcode)
1322 if (opcode == CPL_FW4_ACK)
1323 so = toep->inp->inp_socket;
1325 INP_WLOCK(toep->inp);
1326 so = toep->inp->inp_socket;
1327 INP_WUNLOCK(toep->inp);
1330 if (tom_cpl_iscsi_callback && so) {
1331 if (toep->ulp_mode == ULP_MODE_ISCSI) {
1332 tom_cpl_iscsi_callback(td, so, m, opcode);
1341 t4_queue_iscsi_callback(struct socket *so, struct toepcb *toep,
1342 unsigned int cmd, int *qlen)
1345 if (tom_queue_iscsi_callback && so) {
1346 if (toep->ulp_mode == ULP_MODE_ISCSI)
1347 return (tom_queue_iscsi_callback(so, cmd, qlen));
1354 * TCP RST from the peer, timeout, or some other such critical error.
1357 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1359 struct adapter *sc = iq->adapter;
1360 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1361 unsigned int tid = GET_TID(cpl);
1362 struct toepcb *toep = lookup_tid(sc, tid);
1363 struct sge_wrq *ofld_txq = toep->ofld_txq;
1367 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1370 KASSERT(opcode == CPL_ABORT_REQ_RSS,
1371 ("%s: unexpected opcode 0x%x", __func__, opcode));
1372 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1374 if (toep->flags & TPF_SYNQE)
1375 return (do_abort_req_synqe(iq, rss, m));
1377 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1379 if (negative_advice(cpl->status)) {
1380 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1381 __func__, cpl->status, tid, toep->flags);
1382 return (0); /* Ignore negative advice */
1386 INP_INFO_WLOCK(&V_tcbinfo); /* for tcp_close */
1389 tp = intotcpcb(inp);
1392 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1393 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1394 inp->inp_flags, cpl->status);
1397 * If we'd initiated an abort earlier the reply to it is responsible for
1398 * cleaning up resources. Otherwise we tear everything down right here
1399 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
1401 if (toep->flags & TPF_ABORT_SHUTDOWN) {
1405 toep->flags |= TPF_ABORT_SHUTDOWN;
1407 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1408 struct socket *so = inp->inp_socket;
1411 so_error_set(so, abort_status_to_errno(tp,
1415 INP_WLOCK(inp); /* re-acquire */
1418 final_cpl_received(toep);
1420 INP_INFO_WUNLOCK(&V_tcbinfo);
1421 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1426 * Reply to the CPL_ABORT_REQ (send_reset)
1429 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1431 struct adapter *sc = iq->adapter;
1432 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1433 unsigned int tid = GET_TID(cpl);
1434 struct toepcb *toep = lookup_tid(sc, tid);
1435 struct inpcb *inp = toep->inp;
1437 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1440 KASSERT(opcode == CPL_ABORT_RPL_RSS,
1441 ("%s: unexpected opcode 0x%x", __func__, opcode));
1442 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1444 if (toep->flags & TPF_SYNQE)
1445 return (do_abort_rpl_synqe(iq, rss, m));
1447 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1449 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1450 __func__, tid, toep, inp, cpl->status);
1452 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1453 ("%s: wasn't expecting abort reply", __func__));
1456 final_cpl_received(toep);
1462 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1464 struct adapter *sc = iq->adapter;
1465 const struct cpl_rx_data *cpl = mtod(m, const void *);
1466 unsigned int tid = GET_TID(cpl);
1467 struct toepcb *toep = lookup_tid(sc, tid);
1468 struct inpcb *inp = toep->inp;
1473 uint32_t ddp_placed = 0;
1475 if (__predict_false(toep->flags & TPF_SYNQE)) {
1477 struct synq_entry *synqe = (void *)toep;
1479 INP_WLOCK(synqe->lctx->inp);
1480 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1481 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1482 ("%s: listen socket closed but tid %u not aborted.",
1486 * do_pass_accept_req is still running and will
1487 * eventually take care of this tid.
1490 INP_WUNLOCK(synqe->lctx->inp);
1492 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1498 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1500 /* strip off CPL header */
1501 m_adj(m, sizeof(*cpl));
1502 len = m->m_pkthdr.len;
1505 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1506 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1507 __func__, tid, len, inp->inp_flags);
1513 tp = intotcpcb(inp);
1515 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1516 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1519 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
1521 tp->t_rcvtime = ticks;
1523 so = inp_inpcbtosocket(inp);
1527 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1528 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1529 __func__, tid, len);
1534 INP_INFO_WLOCK(&V_tcbinfo);
1536 tp = tcp_drop(tp, ECONNRESET);
1539 INP_INFO_WUNLOCK(&V_tcbinfo);
1544 /* receive buffer autosize */
1545 if (sb->sb_flags & SB_AUTOSIZE &&
1546 V_tcp_do_autorcvbuf &&
1547 sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1548 len > (sbspace(sb) / 8 * 7)) {
1549 unsigned int hiwat = sb->sb_hiwat;
1550 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1551 V_tcp_autorcvbuf_max);
1553 if (!sbreserve_locked(sb, newsize, so, NULL))
1554 sb->sb_flags &= ~SB_AUTOSIZE;
1556 toep->rx_credits += newsize - hiwat;
1559 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1560 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1563 if (toep->ddp_flags & DDP_SC_REQ)
1564 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1566 KASSERT(cpl->ddp_off == 1,
1567 ("%s: DDP switched on by itself.",
1570 /* Fell out of DDP mode */
1571 toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1575 insert_ddp_data(toep, ddp_placed);
1579 if ((toep->ddp_flags & DDP_OK) == 0 &&
1580 time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1581 toep->ddp_score = DDP_LOW_SCORE;
1582 toep->ddp_flags |= DDP_OK;
1583 CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1584 __func__, tid, time_uptime);
1587 if (toep->ddp_flags & DDP_ON) {
1590 * CPL_RX_DATA with DDP on can only be an indicate. Ask
1591 * soreceive to post a buffer or disable DDP. The
1592 * payload that arrived in this indicate is appended to
1593 * the socket buffer as usual.
1598 "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1599 __func__, tid, toep->flags, be32toh(cpl->seq), len);
1601 sb->sb_flags |= SB_DDP_INDICATE;
1602 } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1603 tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1606 * DDP allowed but isn't on (and a request to switch it
1607 * on isn't pending either), and conditions are ripe for
1608 * it to work. Switch it on.
1611 enable_ddp(sc, toep);
1615 KASSERT(toep->sb_cc >= sb->sb_cc,
1616 ("%s: sb %p has more data (%d) than last time (%d).",
1617 __func__, sb, sb->sb_cc, toep->sb_cc));
1618 toep->rx_credits += toep->sb_cc - sb->sb_cc;
1619 sbappendstream_locked(sb, m);
1620 toep->sb_cc = sb->sb_cc;
1621 sorwakeup_locked(so);
1622 SOCKBUF_UNLOCK_ASSERT(sb);
1628 #define S_CPL_FW4_ACK_OPCODE 24
1629 #define M_CPL_FW4_ACK_OPCODE 0xff
1630 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1631 #define G_CPL_FW4_ACK_OPCODE(x) \
1632 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1634 #define S_CPL_FW4_ACK_FLOWID 0
1635 #define M_CPL_FW4_ACK_FLOWID 0xffffff
1636 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1637 #define G_CPL_FW4_ACK_FLOWID(x) \
1638 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1640 #define S_CPL_FW4_ACK_CR 24
1641 #define M_CPL_FW4_ACK_CR 0xff
1642 #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR)
1643 #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1645 #define S_CPL_FW4_ACK_SEQVAL 0
1646 #define M_CPL_FW4_ACK_SEQVAL 0x1
1647 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1648 #define G_CPL_FW4_ACK_SEQVAL(x) \
1649 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1650 #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U)
1653 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1655 struct adapter *sc = iq->adapter;
1656 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1657 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1658 struct toepcb *toep = lookup_tid(sc, tid);
1662 uint8_t credits = cpl->credits;
1663 struct ofld_tx_sdesc *txsd;
1666 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1670 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1671 * now this comes back carrying the credits for the flowc.
1673 if (__predict_false(toep->flags & TPF_SYNQE)) {
1674 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1675 ("%s: credits for a synq entry %p", __func__, toep));
1681 KASSERT(opcode == CPL_FW4_ACK,
1682 ("%s: unexpected opcode 0x%x", __func__, opcode));
1683 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1684 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1688 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1693 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1694 ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1696 tp = intotcpcb(inp);
1698 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1699 tcp_seq snd_una = be32toh(cpl->snd_una);
1702 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1704 "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1705 __func__, snd_una, toep->tid, tp->snd_una);
1709 if (tp->snd_una != snd_una) {
1710 tp->snd_una = snd_una;
1711 tp->ts_recent_age = tcp_ts_getticks();
1715 so = inp->inp_socket;
1716 txsd = &toep->txsd[toep->txsd_cidx];
1719 KASSERT(credits >= txsd->tx_credits,
1720 ("%s: too many (or partial) credits", __func__));
1721 credits -= txsd->tx_credits;
1722 toep->tx_credits += txsd->tx_credits;
1726 KASSERT(toep->txsd_avail <= toep->txsd_total,
1727 ("%s: txsd avail > total", __func__));
1728 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1729 txsd = &toep->txsd[0];
1730 toep->txsd_cidx = 0;
1734 if (toep->tx_credits == toep->tx_total) {
1735 toep->tx_nocompl = 0;
1736 toep->plen_nocompl = 0;
1739 if (toep->flags & TPF_TX_SUSPENDED &&
1740 toep->tx_credits >= toep->tx_total / 4) {
1741 toep->flags &= ~TPF_TX_SUSPENDED;
1742 if (toep->ulp_mode == ULP_MODE_ISCSI)
1743 t4_ulp_push_frames(sc, toep, plen);
1745 t4_push_frames(sc, toep, plen);
1746 } else if (plen > 0) {
1747 struct sockbuf *sb = &so->so_snd;
1749 if (toep->ulp_mode == ULP_MODE_ISCSI)
1750 t4_cpl_iscsi_callback(toep->td, toep, &plen,
1754 sbdrop_locked(sb, plen);
1755 sowwakeup_locked(so);
1756 SOCKBUF_UNLOCK_ASSERT(sb);
1766 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1768 struct adapter *sc = iq->adapter;
1769 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1770 unsigned int tid = GET_TID(cpl);
1772 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1775 KASSERT(opcode == CPL_SET_TCB_RPL,
1776 ("%s: unexpected opcode 0x%x", __func__, opcode));
1777 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1779 if (is_ftid(sc, tid))
1780 return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1782 struct toepcb *toep = lookup_tid(sc, tid);
1784 t4_cpl_iscsi_callback(toep->td, toep, m, CPL_SET_TCB_RPL);
1788 CXGBE_UNIMPLEMENTED(__func__);
1792 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1793 uint16_t word, uint64_t mask, uint64_t val)
1796 struct cpl_set_tcb_field *req;
1798 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1801 panic("%s: allocation failure.", __func__);
1805 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1806 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1807 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1808 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1809 req->mask = htobe64(mask);
1810 req->val = htobe64(val);
1816 t4_init_cpl_io_handlers(struct adapter *sc)
1819 t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1820 t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1821 t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1822 t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1823 t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1824 t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1825 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1829 t4_uninit_cpl_io_handlers(struct adapter *sc)
1832 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);