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_FW_OFLD_TX_DATA_WR_ULPMODE(ulp_mode >> 4) |
505 V_FW_OFLD_TX_DATA_WR_ULPSUBMODE(ulp_mode & 3);
507 wr_ulp_mode = V_FW_OFLD_TX_DATA_WR_ULPMODE(toep->ulp_mode);
509 txwr->lsodisable_to_proxy =
510 htobe32(wr_ulp_mode |
511 V_FW_OFLD_TX_DATA_WR_URGENT(0) | /* XXX */
512 V_FW_OFLD_TX_DATA_WR_SHOVE(shove));
513 txwr->plen = htobe32(plen);
516 struct tcpcb *tp = intotcpcb(toep->inp);
518 if (plen < 2 * tp->t_maxseg || is_10G_port(toep->port))
519 txwr->lsodisable_to_proxy |=
520 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
522 txwr->lsodisable_to_proxy |=
523 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
524 (tp->t_flags & TF_NODELAY ? 0 :
525 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
530 * Generate a DSGL from a starting mbuf. The total number of segments and the
531 * maximum segments in any one mbuf are provided.
534 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
537 struct ulptx_sgl *usgl = dst;
540 struct sglist_seg segs[n];
542 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
544 sglist_init(&sg, n, segs);
545 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
546 V_ULPTX_NSGE(nsegs));
549 for (m = start; m != stop; m = m->m_next) {
550 rc = sglist_append(&sg, mtod(m, void *), m->m_len);
551 if (__predict_false(rc != 0))
552 panic("%s: sglist_append %d", __func__, rc);
554 for (j = 0; j < sg.sg_nseg; i++, j++) {
556 usgl->len0 = htobe32(segs[j].ss_len);
557 usgl->addr0 = htobe64(segs[j].ss_paddr);
559 usgl->sge[i / 2].len[i & 1] =
560 htobe32(segs[j].ss_len);
561 usgl->sge[i / 2].addr[i & 1] =
562 htobe64(segs[j].ss_paddr);
571 usgl->sge[i / 2].len[1] = htobe32(0);
572 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
573 __func__, nsegs, start, stop));
577 * Max number of SGL entries an offload tx work request can have. This is 41
578 * (1 + 40) for a full 512B work request.
579 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
581 #define OFLD_SGL_LEN (41)
584 * Send data and/or a FIN to the peer.
586 * The socket's so_snd buffer consists of a stream of data starting with sb_mb
587 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that
590 * drop indicates the number of bytes that should be dropped from the head of
591 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating
592 * contention on the send buffer lock (before this change it used to do
593 * sowwakeup and then t4_push_frames right after that when recovering from tx
594 * stalls). When drop is set this function MUST drop the bytes and wake up any
598 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
600 struct mbuf *sndptr, *m, *sb_sndptr;
601 struct fw_ofld_tx_data_wr *txwr;
603 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
604 struct inpcb *inp = toep->inp;
605 struct tcpcb *tp = intotcpcb(inp);
606 struct socket *so = inp->inp_socket;
607 struct sockbuf *sb = &so->so_snd;
608 int tx_credits, shove, compl, space, sowwakeup;
609 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
611 INP_WLOCK_ASSERT(inp);
612 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
613 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
615 KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
616 toep->ulp_mode == ULP_MODE_TCPDDP ||
617 toep->ulp_mode == ULP_MODE_RDMA,
618 ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
621 * This function doesn't resume by itself. Someone else must clear the
622 * flag and call this function.
624 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
626 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
631 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
632 max_imm = max_imm_payload(tx_credits);
633 max_nsegs = max_dsgl_nsegs(tx_credits);
638 sbdrop_locked(sb, drop);
641 sb_sndptr = sb->sb_sndptr;
642 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
645 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
646 for (m = sndptr; m != NULL; m = m->m_next) {
647 int n = sglist_count(mtod(m, void *), m->m_len);
652 /* This mbuf sent us _over_ the nsegs limit, back out */
653 if (plen > max_imm && nsegs > max_nsegs) {
657 /* Too few credits */
658 toep->flags |= TPF_TX_SUSPENDED;
660 sowwakeup_locked(so);
663 SOCKBUF_UNLOCK_ASSERT(sb);
669 if (max_nsegs_1mbuf < n)
671 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */
673 /* This mbuf put us right at the max_nsegs limit */
674 if (plen > max_imm && nsegs == max_nsegs) {
680 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
683 if (space <= sb->sb_hiwat * 3 / 8 &&
684 toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
689 if (sb->sb_flags & SB_AUTOSIZE &&
690 V_tcp_do_autosndbuf &&
691 sb->sb_hiwat < V_tcp_autosndbuf_max &&
692 space < sb->sb_hiwat / 8) {
693 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
694 V_tcp_autosndbuf_max);
696 if (!sbreserve_locked(sb, newsize, so, NULL))
697 sb->sb_flags &= ~SB_AUTOSIZE;
699 sowwakeup = 1; /* room available */
702 sowwakeup_locked(so);
705 SOCKBUF_UNLOCK_ASSERT(sb);
707 /* nothing to send */
710 ("%s: nothing to send, but m != NULL", __func__));
714 if (__predict_false(toep->flags & TPF_FIN_SENT))
715 panic("%s: excess tx.", __func__);
717 if (plen <= max_imm) {
719 /* Immediate data tx */
721 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
724 /* XXX: how will we recover from this? */
725 toep->flags |= TPF_TX_SUSPENDED;
729 credits = howmany(wr->wr_len, 16);
730 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
732 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
739 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
740 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
741 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
743 /* XXX: how will we recover from this? */
744 toep->flags |= TPF_TX_SUSPENDED;
748 credits = howmany(wr_len, 16);
749 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
751 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
754 uint64_t *pad = (uint64_t *)
755 ((uintptr_t)txwr + wr_len);
760 KASSERT(toep->tx_credits >= credits,
761 ("%s: not enough credits", __func__));
763 toep->tx_credits -= credits;
764 toep->tx_nocompl += credits;
765 toep->plen_nocompl += plen;
766 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
767 toep->tx_nocompl >= toep->tx_total / 4)
770 if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
771 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
772 toep->tx_nocompl = 0;
773 toep->plen_nocompl = 0;
780 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
781 sb->sb_sndptr = sb_sndptr;
784 toep->flags |= TPF_TX_DATA_SENT;
785 if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
786 toep->flags |= TPF_TX_SUSPENDED;
788 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
790 txsd->tx_credits = credits;
792 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
794 txsd = &toep->txsd[0];
798 t4_l2t_send(sc, wr, toep->l2te);
801 /* Send a FIN if requested, but only if there's no more data to send */
802 if (m == NULL && toep->flags & TPF_SEND_FIN)
803 close_conn(sc, toep);
806 /* Send ULP data over TOE using TX_DATA_WR. We send whole mbuf at once */
808 t4_ulp_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
810 struct mbuf *sndptr, *m = NULL;
811 struct fw_ofld_tx_data_wr *txwr;
813 unsigned int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
814 struct inpcb *inp = toep->inp;
818 int tx_credits, ulp_len = 0, ulp_mode = 0, qlen = 0;
820 struct ofld_tx_sdesc *txsd;
822 INP_WLOCK_ASSERT(inp);
823 if (toep->flags & TPF_ABORT_SHUTDOWN)
827 so = inp->inp_socket;
829 txsd = &toep->txsd[toep->txsd_pidx];
831 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
832 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
835 * This function doesn't resume by itself. Someone else must clear the
836 * flag and call this function.
838 if (__predict_false(toep->flags & TPF_TX_SUSPENDED))
841 sndptr = t4_queue_iscsi_callback(so, toep, 1, &qlen);
846 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
847 max_imm = max_imm_payload(tx_credits);
848 max_nsegs = max_dsgl_nsegs(tx_credits);
851 t4_cpl_iscsi_callback(toep->td, toep, &drop,
858 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
859 for (m = sndptr; m != NULL; m = m->m_next) {
860 int n = sglist_count(mtod(m, void *), m->m_len);
865 /* This mbuf sent us _over_ the nsegs limit, return */
866 if (plen > max_imm && nsegs > max_nsegs) {
867 toep->flags |= TPF_TX_SUSPENDED;
871 if (max_nsegs_1mbuf < n)
874 /* This mbuf put us right at the max_nsegs limit */
875 if (plen > max_imm && nsegs == max_nsegs) {
876 toep->flags |= TPF_TX_SUSPENDED;
881 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
882 /* nothing to send */
885 ("%s: nothing to send, but m != NULL", __func__));
889 if (__predict_false(toep->flags & TPF_FIN_SENT))
890 panic("%s: excess tx.", __func__);
892 ulp_len = plen + ulp_extra_len(sndptr, &ulp_mode);
893 if (plen <= max_imm) {
895 /* Immediate data tx */
896 wr = alloc_wrqe(roundup(sizeof(*txwr) + plen, 16),
899 /* XXX: how will we recover from this? */
900 toep->flags |= TPF_TX_SUSPENDED;
904 credits = howmany(wr->wr_len, 16);
905 write_tx_wr(txwr, toep, plen, ulp_len, credits, shove,
907 m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
912 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
913 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
914 wr = alloc_wrqe(roundup(wr_len, 16), toep->ofld_txq);
916 /* XXX: how will we recover from this? */
917 toep->flags |= TPF_TX_SUSPENDED;
921 credits = howmany(wr_len, 16);
922 write_tx_wr(txwr, toep, 0, ulp_len, credits, shove,
924 write_tx_sgl(txwr + 1, sndptr, m, nsegs,
927 uint64_t *pad = (uint64_t *)
928 ((uintptr_t)txwr + wr_len);
933 KASSERT(toep->tx_credits >= credits,
934 ("%s: not enough credits", __func__));
936 toep->tx_credits -= credits;
937 toep->tx_nocompl += credits;
938 toep->plen_nocompl += plen;
939 if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
940 toep->tx_nocompl >= toep->tx_total / 4)
944 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
945 toep->tx_nocompl = 0;
946 toep->plen_nocompl = 0;
948 tp->snd_nxt += ulp_len;
949 tp->snd_max += ulp_len;
952 sndptr = m = t4_queue_iscsi_callback(so, toep, 2, &qlen);
954 toep->flags |= TPF_TX_DATA_SENT;
955 if (toep->tx_credits < MIN_OFLD_TX_CREDITS) {
956 toep->flags |= TPF_TX_SUSPENDED;
959 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
961 txsd->tx_credits = credits;
963 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
965 txsd = &toep->txsd[0];
969 t4_l2t_send(sc, wr, toep->l2te);
972 /* Send a FIN if requested, but only if there's no more data to send */
973 if (m == NULL && toep->flags & TPF_SEND_FIN)
974 close_conn(sc, toep);
978 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
980 struct adapter *sc = tod->tod_softc;
982 struct inpcb *inp = tp->t_inpcb;
984 struct toepcb *toep = tp->t_toe;
986 INP_WLOCK_ASSERT(inp);
987 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
988 ("%s: inp %p dropped.", __func__, inp));
989 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
991 t4_push_frames(sc, toep, 0);
997 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
999 struct adapter *sc = tod->tod_softc;
1001 struct inpcb *inp = tp->t_inpcb;
1003 struct toepcb *toep = tp->t_toe;
1005 INP_WLOCK_ASSERT(inp);
1006 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1007 ("%s: inp %p dropped.", __func__, inp));
1008 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1010 toep->flags |= TPF_SEND_FIN;
1011 if (tp->t_state >= TCPS_ESTABLISHED) {
1012 if (toep->ulp_mode == ULP_MODE_ISCSI)
1013 t4_ulp_push_frames(sc, toep, 0);
1015 t4_push_frames(sc, toep, 0);
1022 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1024 struct adapter *sc = tod->tod_softc;
1025 #if defined(INVARIANTS)
1026 struct inpcb *inp = tp->t_inpcb;
1028 struct toepcb *toep = tp->t_toe;
1030 INP_WLOCK_ASSERT(inp);
1031 KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1032 ("%s: inp %p dropped.", __func__, inp));
1033 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1036 KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1037 ("%s: flowc for tid %u [%s] not sent already",
1038 __func__, toep->tid, tcpstates[tp->t_state]));
1040 send_reset(sc, toep, 0);
1045 * Peer has sent us a FIN.
1048 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1050 struct adapter *sc = iq->adapter;
1051 const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1052 unsigned int tid = GET_TID(cpl);
1053 struct toepcb *toep = lookup_tid(sc, tid);
1054 struct inpcb *inp = toep->inp;
1055 struct tcpcb *tp = NULL;
1059 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1062 KASSERT(opcode == CPL_PEER_CLOSE,
1063 ("%s: unexpected opcode 0x%x", __func__, opcode));
1064 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1066 if (__predict_false(toep->flags & TPF_SYNQE)) {
1068 struct synq_entry *synqe = (void *)toep;
1070 INP_WLOCK(synqe->lctx->inp);
1071 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1072 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1073 ("%s: listen socket closed but tid %u not aborted.",
1077 * do_pass_accept_req is still running and will
1078 * eventually take care of this tid.
1081 INP_WUNLOCK(synqe->lctx->inp);
1083 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1088 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1090 INP_INFO_WLOCK(&V_tcbinfo);
1092 tp = intotcpcb(inp);
1094 CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1095 tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1097 if (toep->flags & TPF_ABORT_SHUTDOWN)
1100 tp->rcv_nxt++; /* FIN */
1102 so = inp->inp_socket;
1105 if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
1106 m = get_ddp_mbuf(be32toh(cpl->rcv_nxt) - tp->rcv_nxt);
1107 tp->rcv_nxt = be32toh(cpl->rcv_nxt);
1108 toep->ddp_flags &= ~(DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE);
1110 KASSERT(toep->sb_cc >= sb->sb_cc,
1111 ("%s: sb %p has more data (%d) than last time (%d).",
1112 __func__, sb, sb->sb_cc, toep->sb_cc));
1113 toep->rx_credits += toep->sb_cc - sb->sb_cc;
1114 #ifdef USE_DDP_RX_FLOW_CONTROL
1115 toep->rx_credits -= m->m_len; /* adjust for F_RX_FC_DDP */
1117 sbappendstream_locked(sb, m);
1118 toep->sb_cc = sb->sb_cc;
1120 socantrcvmore_locked(so); /* unlocks the sockbuf */
1122 if (toep->ulp_mode != ULP_MODE_RDMA) {
1123 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1124 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1125 be32toh(cpl->rcv_nxt)));
1128 switch (tp->t_state) {
1129 case TCPS_SYN_RECEIVED:
1130 tp->t_starttime = ticks;
1133 case TCPS_ESTABLISHED:
1134 tp->t_state = TCPS_CLOSE_WAIT;
1137 case TCPS_FIN_WAIT_1:
1138 tp->t_state = TCPS_CLOSING;
1141 case TCPS_FIN_WAIT_2:
1143 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1144 INP_INFO_WUNLOCK(&V_tcbinfo);
1147 final_cpl_received(toep);
1151 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1152 __func__, tid, tp->t_state);
1156 INP_INFO_WUNLOCK(&V_tcbinfo);
1161 * Peer has ACK'd our FIN.
1164 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1167 struct adapter *sc = iq->adapter;
1168 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1169 unsigned int tid = GET_TID(cpl);
1170 struct toepcb *toep = lookup_tid(sc, tid);
1171 struct inpcb *inp = toep->inp;
1172 struct tcpcb *tp = NULL;
1173 struct socket *so = NULL;
1175 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1178 KASSERT(opcode == CPL_CLOSE_CON_RPL,
1179 ("%s: unexpected opcode 0x%x", __func__, opcode));
1180 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1181 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1183 INP_INFO_WLOCK(&V_tcbinfo);
1185 tp = intotcpcb(inp);
1187 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1188 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1190 if (toep->flags & TPF_ABORT_SHUTDOWN)
1193 so = inp->inp_socket;
1194 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */
1196 switch (tp->t_state) {
1197 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */
1200 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */
1201 INP_INFO_WUNLOCK(&V_tcbinfo);
1204 final_cpl_received(toep); /* no more CPLs expected */
1212 case TCPS_FIN_WAIT_1:
1213 if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1214 soisdisconnected(so);
1215 tp->t_state = TCPS_FIN_WAIT_2;
1220 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1221 __func__, tid, tcpstates[tp->t_state]);
1225 INP_INFO_WUNLOCK(&V_tcbinfo);
1230 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1234 struct cpl_abort_rpl *cpl;
1236 wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1239 panic("%s: allocation failure.", __func__);
1243 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1244 cpl->cmd = rst_status;
1250 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1252 switch (abort_reason) {
1253 case CPL_ERR_BAD_SYN:
1254 case CPL_ERR_CONN_RESET:
1255 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1256 case CPL_ERR_XMIT_TIMEDOUT:
1257 case CPL_ERR_PERSIST_TIMEDOUT:
1258 case CPL_ERR_FINWAIT2_TIMEDOUT:
1259 case CPL_ERR_KEEPALIVE_TIMEDOUT:
1267 cpl_not_handled(struct sge_iq *, const struct rss_header *, struct mbuf *);
1269 * tom_cpl_iscsi_callback -
1270 * iscsi and tom would share the following cpl messages, so when any of these
1271 * message is received, after tom is done with processing it, the messages
1272 * needs to be forwarded to iscsi for further processing:
1276 void (*tom_cpl_iscsi_callback)(struct tom_data *, struct socket *, void *,
1279 struct mbuf *(*tom_queue_iscsi_callback)(struct socket *, unsigned int, int *);
1281 * Check if the handler function is set for a given CPL
1282 * return 0 if the function is NULL or cpl_not_handled, 1 otherwise.
1285 t4tom_cpl_handler_registered(struct adapter *sc, unsigned int opcode)
1288 MPASS(opcode < nitems(sc->cpl_handler));
1290 return (sc->cpl_handler[opcode] &&
1291 sc->cpl_handler[opcode] != cpl_not_handled);
1295 * set the tom_cpl_iscsi_callback function, this function should be used
1296 * whenever both toe and iscsi need to process the same cpl msg.
1299 t4tom_register_cpl_iscsi_callback(void (*fp)(struct tom_data *, struct socket *,
1300 void *, unsigned int))
1303 tom_cpl_iscsi_callback = fp;
1307 t4tom_register_queue_iscsi_callback(struct mbuf *(*fp)(struct socket *,
1308 unsigned int, int *qlen))
1311 tom_queue_iscsi_callback = fp;
1315 t4_cpl_iscsi_callback(struct tom_data *td, struct toepcb *toep, void *m,
1316 unsigned int opcode)
1320 if (opcode == CPL_FW4_ACK)
1321 so = toep->inp->inp_socket;
1323 INP_WLOCK(toep->inp);
1324 so = toep->inp->inp_socket;
1325 INP_WUNLOCK(toep->inp);
1328 if (tom_cpl_iscsi_callback && so) {
1329 if (toep->ulp_mode == ULP_MODE_ISCSI) {
1330 tom_cpl_iscsi_callback(td, so, m, opcode);
1339 t4_queue_iscsi_callback(struct socket *so, struct toepcb *toep,
1340 unsigned int cmd, int *qlen)
1343 if (tom_queue_iscsi_callback && so) {
1344 if (toep->ulp_mode == ULP_MODE_ISCSI)
1345 return (tom_queue_iscsi_callback(so, cmd, qlen));
1352 * TCP RST from the peer, timeout, or some other such critical error.
1355 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1357 struct adapter *sc = iq->adapter;
1358 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1359 unsigned int tid = GET_TID(cpl);
1360 struct toepcb *toep = lookup_tid(sc, tid);
1361 struct sge_wrq *ofld_txq = toep->ofld_txq;
1365 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1368 KASSERT(opcode == CPL_ABORT_REQ_RSS,
1369 ("%s: unexpected opcode 0x%x", __func__, opcode));
1370 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1372 if (toep->flags & TPF_SYNQE)
1373 return (do_abort_req_synqe(iq, rss, m));
1375 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1377 if (negative_advice(cpl->status)) {
1378 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1379 __func__, cpl->status, tid, toep->flags);
1380 return (0); /* Ignore negative advice */
1384 INP_INFO_WLOCK(&V_tcbinfo); /* for tcp_close */
1387 tp = intotcpcb(inp);
1390 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1391 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1392 inp->inp_flags, cpl->status);
1395 * If we'd initiated an abort earlier the reply to it is responsible for
1396 * cleaning up resources. Otherwise we tear everything down right here
1397 * right now. We owe the T4 a CPL_ABORT_RPL no matter what.
1399 if (toep->flags & TPF_ABORT_SHUTDOWN) {
1403 toep->flags |= TPF_ABORT_SHUTDOWN;
1405 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1406 struct socket *so = inp->inp_socket;
1409 so_error_set(so, abort_status_to_errno(tp,
1413 INP_WLOCK(inp); /* re-acquire */
1416 final_cpl_received(toep);
1418 INP_INFO_WUNLOCK(&V_tcbinfo);
1419 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1424 * Reply to the CPL_ABORT_REQ (send_reset)
1427 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1429 struct adapter *sc = iq->adapter;
1430 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1431 unsigned int tid = GET_TID(cpl);
1432 struct toepcb *toep = lookup_tid(sc, tid);
1433 struct inpcb *inp = toep->inp;
1435 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1438 KASSERT(opcode == CPL_ABORT_RPL_RSS,
1439 ("%s: unexpected opcode 0x%x", __func__, opcode));
1440 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1442 if (toep->flags & TPF_SYNQE)
1443 return (do_abort_rpl_synqe(iq, rss, m));
1445 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1447 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1448 __func__, tid, toep, inp, cpl->status);
1450 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1451 ("%s: wasn't expecting abort reply", __func__));
1454 final_cpl_received(toep);
1460 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1462 struct adapter *sc = iq->adapter;
1463 const struct cpl_rx_data *cpl = mtod(m, const void *);
1464 unsigned int tid = GET_TID(cpl);
1465 struct toepcb *toep = lookup_tid(sc, tid);
1466 struct inpcb *inp = toep->inp;
1471 uint32_t ddp_placed = 0;
1473 if (__predict_false(toep->flags & TPF_SYNQE)) {
1475 struct synq_entry *synqe = (void *)toep;
1477 INP_WLOCK(synqe->lctx->inp);
1478 if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1479 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1480 ("%s: listen socket closed but tid %u not aborted.",
1484 * do_pass_accept_req is still running and will
1485 * eventually take care of this tid.
1488 INP_WUNLOCK(synqe->lctx->inp);
1490 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1496 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1498 /* strip off CPL header */
1499 m_adj(m, sizeof(*cpl));
1500 len = m->m_pkthdr.len;
1503 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1504 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1505 __func__, tid, len, inp->inp_flags);
1511 tp = intotcpcb(inp);
1513 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1514 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1517 KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
1519 tp->t_rcvtime = ticks;
1521 so = inp_inpcbtosocket(inp);
1525 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1526 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1527 __func__, tid, len);
1532 INP_INFO_WLOCK(&V_tcbinfo);
1534 tp = tcp_drop(tp, ECONNRESET);
1537 INP_INFO_WUNLOCK(&V_tcbinfo);
1542 /* receive buffer autosize */
1543 if (sb->sb_flags & SB_AUTOSIZE &&
1544 V_tcp_do_autorcvbuf &&
1545 sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1546 len > (sbspace(sb) / 8 * 7)) {
1547 unsigned int hiwat = sb->sb_hiwat;
1548 unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1549 V_tcp_autorcvbuf_max);
1551 if (!sbreserve_locked(sb, newsize, so, NULL))
1552 sb->sb_flags &= ~SB_AUTOSIZE;
1554 toep->rx_credits += newsize - hiwat;
1557 if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1558 int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1561 if (toep->ddp_flags & DDP_SC_REQ)
1562 toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1564 KASSERT(cpl->ddp_off == 1,
1565 ("%s: DDP switched on by itself.",
1568 /* Fell out of DDP mode */
1569 toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1573 insert_ddp_data(toep, ddp_placed);
1577 if ((toep->ddp_flags & DDP_OK) == 0 &&
1578 time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1579 toep->ddp_score = DDP_LOW_SCORE;
1580 toep->ddp_flags |= DDP_OK;
1581 CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1582 __func__, tid, time_uptime);
1585 if (toep->ddp_flags & DDP_ON) {
1588 * CPL_RX_DATA with DDP on can only be an indicate. Ask
1589 * soreceive to post a buffer or disable DDP. The
1590 * payload that arrived in this indicate is appended to
1591 * the socket buffer as usual.
1596 "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1597 __func__, tid, toep->flags, be32toh(cpl->seq), len);
1599 sb->sb_flags |= SB_DDP_INDICATE;
1600 } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1601 tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1604 * DDP allowed but isn't on (and a request to switch it
1605 * on isn't pending either), and conditions are ripe for
1606 * it to work. Switch it on.
1609 enable_ddp(sc, toep);
1613 KASSERT(toep->sb_cc >= sb->sb_cc,
1614 ("%s: sb %p has more data (%d) than last time (%d).",
1615 __func__, sb, sb->sb_cc, toep->sb_cc));
1616 toep->rx_credits += toep->sb_cc - sb->sb_cc;
1617 sbappendstream_locked(sb, m);
1618 toep->sb_cc = sb->sb_cc;
1619 if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1622 credits = send_rx_credits(sc, toep, toep->rx_credits);
1623 toep->rx_credits -= credits;
1624 tp->rcv_wnd += credits;
1625 tp->rcv_adv += credits;
1627 sorwakeup_locked(so);
1628 SOCKBUF_UNLOCK_ASSERT(sb);
1634 #define S_CPL_FW4_ACK_OPCODE 24
1635 #define M_CPL_FW4_ACK_OPCODE 0xff
1636 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1637 #define G_CPL_FW4_ACK_OPCODE(x) \
1638 (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1640 #define S_CPL_FW4_ACK_FLOWID 0
1641 #define M_CPL_FW4_ACK_FLOWID 0xffffff
1642 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1643 #define G_CPL_FW4_ACK_FLOWID(x) \
1644 (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1646 #define S_CPL_FW4_ACK_CR 24
1647 #define M_CPL_FW4_ACK_CR 0xff
1648 #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR)
1649 #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1651 #define S_CPL_FW4_ACK_SEQVAL 0
1652 #define M_CPL_FW4_ACK_SEQVAL 0x1
1653 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1654 #define G_CPL_FW4_ACK_SEQVAL(x) \
1655 (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1656 #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U)
1659 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1661 struct adapter *sc = iq->adapter;
1662 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1663 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1664 struct toepcb *toep = lookup_tid(sc, tid);
1668 uint8_t credits = cpl->credits;
1669 struct ofld_tx_sdesc *txsd;
1672 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1676 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1677 * now this comes back carrying the credits for the flowc.
1679 if (__predict_false(toep->flags & TPF_SYNQE)) {
1680 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1681 ("%s: credits for a synq entry %p", __func__, toep));
1687 KASSERT(opcode == CPL_FW4_ACK,
1688 ("%s: unexpected opcode 0x%x", __func__, opcode));
1689 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1690 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1694 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1699 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1700 ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1702 tp = intotcpcb(inp);
1704 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1705 tcp_seq snd_una = be32toh(cpl->snd_una);
1708 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1710 "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1711 __func__, snd_una, toep->tid, tp->snd_una);
1715 if (tp->snd_una != snd_una) {
1716 tp->snd_una = snd_una;
1717 tp->ts_recent_age = tcp_ts_getticks();
1721 so = inp->inp_socket;
1722 txsd = &toep->txsd[toep->txsd_cidx];
1725 KASSERT(credits >= txsd->tx_credits,
1726 ("%s: too many (or partial) credits", __func__));
1727 credits -= txsd->tx_credits;
1728 toep->tx_credits += txsd->tx_credits;
1732 KASSERT(toep->txsd_avail <= toep->txsd_total,
1733 ("%s: txsd avail > total", __func__));
1734 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1735 txsd = &toep->txsd[0];
1736 toep->txsd_cidx = 0;
1740 if (toep->tx_credits == toep->tx_total) {
1741 toep->tx_nocompl = 0;
1742 toep->plen_nocompl = 0;
1745 if (toep->flags & TPF_TX_SUSPENDED &&
1746 toep->tx_credits >= toep->tx_total / 4) {
1747 toep->flags &= ~TPF_TX_SUSPENDED;
1748 if (toep->ulp_mode == ULP_MODE_ISCSI)
1749 t4_ulp_push_frames(sc, toep, plen);
1751 t4_push_frames(sc, toep, plen);
1752 } else if (plen > 0) {
1753 struct sockbuf *sb = &so->so_snd;
1755 if (toep->ulp_mode == ULP_MODE_ISCSI)
1756 t4_cpl_iscsi_callback(toep->td, toep, &plen,
1760 sbdrop_locked(sb, plen);
1761 sowwakeup_locked(so);
1762 SOCKBUF_UNLOCK_ASSERT(sb);
1772 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1774 struct adapter *sc = iq->adapter;
1775 const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1776 unsigned int tid = GET_TID(cpl);
1778 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1781 KASSERT(opcode == CPL_SET_TCB_RPL,
1782 ("%s: unexpected opcode 0x%x", __func__, opcode));
1783 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1785 if (is_ftid(sc, tid))
1786 return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1788 struct toepcb *toep = lookup_tid(sc, tid);
1790 t4_cpl_iscsi_callback(toep->td, toep, m, CPL_SET_TCB_RPL);
1794 CXGBE_UNIMPLEMENTED(__func__);
1798 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1799 uint16_t word, uint64_t mask, uint64_t val)
1802 struct cpl_set_tcb_field *req;
1804 wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1807 panic("%s: allocation failure.", __func__);
1811 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1812 req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1813 V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1814 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1815 req->mask = htobe64(mask);
1816 req->val = htobe64(val);
1822 t4_init_cpl_io_handlers(struct adapter *sc)
1825 t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1826 t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1827 t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1828 t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1829 t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1830 t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1831 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1835 t4_uninit_cpl_io_handlers(struct adapter *sc)
1838 t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);