xref: /freebsd/sys/dev/cxgbe/tom/t4_cpl_io.c (revision f7862a87d0e5e05024fe0399f091c616c748cccc)
1 /*-
2  * Copyright (c) 2012, 2015 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
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.
14  *
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
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 
33 #ifdef TCP_OFFLOAD
34 #include <sys/param.h>
35 #include <sys/types.h>
36 #include <sys/kernel.h>
37 #include <sys/ktr.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>
49 #define TCPSTATES
50 #include <netinet/tcp_fsm.h>
51 #include <netinet/tcp_seq.h>
52 #include <netinet/toecore.h>
53 
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"
60 
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)
73 
74 void
75 send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp)
76 {
77 	struct wrqe *wr;
78 	struct fw_flowc_wr *flowc;
79 	unsigned int nparams = ftxp ? 8 : 6, flowclen;
80 	struct vi_info *vi = toep->vi;
81 	struct port_info *pi = vi->pi;
82 	struct adapter *sc = pi->adapter;
83 	unsigned int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
84 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
85 
86 	KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
87 	    ("%s: flowc for tid %u sent already", __func__, toep->tid));
88 
89 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
90 
91 	wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
92 	if (wr == NULL) {
93 		/* XXX */
94 		panic("%s: allocation failure.", __func__);
95 	}
96 	flowc = wrtod(wr);
97 	memset(flowc, 0, wr->wr_len);
98 
99 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
100 	    V_FW_FLOWC_WR_NPARAMS(nparams));
101 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
102 	    V_FW_WR_FLOWID(toep->tid));
103 
104 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN;
105 	flowc->mnemval[0].val = htobe32(pfvf);
106 	flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH;
107 	flowc->mnemval[1].val = htobe32(pi->tx_chan);
108 	flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT;
109 	flowc->mnemval[2].val = htobe32(pi->tx_chan);
110 	flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID;
111 	flowc->mnemval[3].val = htobe32(toep->ofld_rxq->iq.abs_id);
112 	if (ftxp) {
113 		uint32_t sndbuf = min(ftxp->snd_space, sc->tt.sndbuf);
114 
115 		flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDNXT;
116 		flowc->mnemval[4].val = htobe32(ftxp->snd_nxt);
117 		flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_RCVNXT;
118 		flowc->mnemval[5].val = htobe32(ftxp->rcv_nxt);
119 		flowc->mnemval[6].mnemonic = FW_FLOWC_MNEM_SNDBUF;
120 		flowc->mnemval[6].val = htobe32(sndbuf);
121 		flowc->mnemval[7].mnemonic = FW_FLOWC_MNEM_MSS;
122 		flowc->mnemval[7].val = htobe32(ftxp->mss);
123 
124 		CTR6(KTR_CXGBE,
125 		    "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x",
126 		    __func__, toep->tid, ftxp->mss, sndbuf, ftxp->snd_nxt,
127 		    ftxp->rcv_nxt);
128 	} else {
129 		flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF;
130 		flowc->mnemval[4].val = htobe32(512);
131 		flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS;
132 		flowc->mnemval[5].val = htobe32(512);
133 
134 		CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid);
135 	}
136 
137 	txsd->tx_credits = howmany(flowclen, 16);
138 	txsd->plen = 0;
139 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
140 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
141 	toep->tx_credits -= txsd->tx_credits;
142 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
143 		toep->txsd_pidx = 0;
144 	toep->txsd_avail--;
145 
146 	toep->flags |= TPF_FLOWC_WR_SENT;
147         t4_wrq_tx(sc, wr);
148 }
149 
150 void
151 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
152 {
153 	struct wrqe *wr;
154 	struct cpl_abort_req *req;
155 	int tid = toep->tid;
156 	struct inpcb *inp = toep->inp;
157 	struct tcpcb *tp = intotcpcb(inp);	/* don't use if INP_DROPPED */
158 
159 	INP_WLOCK_ASSERT(inp);
160 
161 	CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
162 	    __func__, toep->tid,
163 	    inp->inp_flags & INP_DROPPED ? "inp dropped" :
164 	    tcpstates[tp->t_state],
165 	    toep->flags, inp->inp_flags,
166 	    toep->flags & TPF_ABORT_SHUTDOWN ?
167 	    " (abort already in progress)" : "");
168 
169 	if (toep->flags & TPF_ABORT_SHUTDOWN)
170 		return;	/* abort already in progress */
171 
172 	toep->flags |= TPF_ABORT_SHUTDOWN;
173 
174 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
175 	    ("%s: flowc_wr not sent for tid %d.", __func__, tid));
176 
177 	wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
178 	if (wr == NULL) {
179 		/* XXX */
180 		panic("%s: allocation failure.", __func__);
181 	}
182 	req = wrtod(wr);
183 
184 	INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
185 	if (inp->inp_flags & INP_DROPPED)
186 		req->rsvd0 = htobe32(snd_nxt);
187 	else
188 		req->rsvd0 = htobe32(tp->snd_nxt);
189 	req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
190 	req->cmd = CPL_ABORT_SEND_RST;
191 
192 	/*
193 	 * XXX: What's the correct way to tell that the inp hasn't been detached
194 	 * from its socket?  Should I even be flushing the snd buffer here?
195 	 */
196 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
197 		struct socket *so = inp->inp_socket;
198 
199 		if (so != NULL)	/* because I'm not sure.  See comment above */
200 			sbflush(&so->so_snd);
201 	}
202 
203 	t4_l2t_send(sc, wr, toep->l2te);
204 }
205 
206 /*
207  * Called when a connection is established to translate the TCP options
208  * reported by HW to FreeBSD's native format.
209  */
210 static void
211 assign_rxopt(struct tcpcb *tp, unsigned int opt)
212 {
213 	struct toepcb *toep = tp->t_toe;
214 	struct inpcb *inp = tp->t_inpcb;
215 	struct adapter *sc = td_adapter(toep->td);
216 	int n;
217 
218 	INP_LOCK_ASSERT(inp);
219 
220 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
221 		n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
222 	else
223 		n = sizeof(struct ip) + sizeof(struct tcphdr);
224 	tp->t_maxseg = tp->t_maxopd = sc->params.mtus[G_TCPOPT_MSS(opt)] - n;
225 
226 	CTR4(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u)", __func__, toep->tid,
227 	    G_TCPOPT_MSS(opt), sc->params.mtus[G_TCPOPT_MSS(opt)]);
228 
229 	if (G_TCPOPT_TSTAMP(opt)) {
230 		tp->t_flags |= TF_RCVD_TSTMP;	/* timestamps ok */
231 		tp->ts_recent = 0;		/* hmmm */
232 		tp->ts_recent_age = tcp_ts_getticks();
233 		tp->t_maxseg -= TCPOLEN_TSTAMP_APPA;
234 	}
235 
236 	if (G_TCPOPT_SACK(opt))
237 		tp->t_flags |= TF_SACK_PERMIT;	/* should already be set */
238 	else
239 		tp->t_flags &= ~TF_SACK_PERMIT;	/* sack disallowed by peer */
240 
241 	if (G_TCPOPT_WSCALE_OK(opt))
242 		tp->t_flags |= TF_RCVD_SCALE;
243 
244 	/* Doing window scaling? */
245 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
246 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
247 		tp->rcv_scale = tp->request_r_scale;
248 		tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
249 	}
250 }
251 
252 /*
253  * Completes some final bits of initialization for just established connections
254  * and changes their state to TCPS_ESTABLISHED.
255  *
256  * The ISNs are from after the exchange of SYNs.  i.e., the true ISN + 1.
257  */
258 void
259 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
260     uint16_t opt)
261 {
262 	struct inpcb *inp = toep->inp;
263 	struct socket *so = inp->inp_socket;
264 	struct tcpcb *tp = intotcpcb(inp);
265 	long bufsize;
266 	uint32_t iss = be32toh(snd_isn) - 1;	/* true ISS */
267 	uint32_t irs = be32toh(rcv_isn) - 1;	/* true IRS */
268 	uint16_t tcpopt = be16toh(opt);
269 	struct flowc_tx_params ftxp;
270 
271 	INP_WLOCK_ASSERT(inp);
272 	KASSERT(tp->t_state == TCPS_SYN_SENT ||
273 	    tp->t_state == TCPS_SYN_RECEIVED,
274 	    ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
275 
276 	CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p",
277 	    __func__, toep->tid, toep, inp);
278 
279 	tp->t_state = TCPS_ESTABLISHED;
280 	tp->t_starttime = ticks;
281 	TCPSTAT_INC(tcps_connects);
282 
283 	tp->irs = irs;
284 	tcp_rcvseqinit(tp);
285 	tp->rcv_wnd = toep->rx_credits << 10;
286 	tp->rcv_adv += tp->rcv_wnd;
287 	tp->last_ack_sent = tp->rcv_nxt;
288 
289 	/*
290 	 * If we were unable to send all rx credits via opt0, save the remainder
291 	 * in rx_credits so that they can be handed over with the next credit
292 	 * update.
293 	 */
294 	SOCKBUF_LOCK(&so->so_rcv);
295 	bufsize = select_rcv_wnd(so);
296 	SOCKBUF_UNLOCK(&so->so_rcv);
297 	toep->rx_credits = bufsize - tp->rcv_wnd;
298 
299 	tp->iss = iss;
300 	tcp_sendseqinit(tp);
301 	tp->snd_una = iss + 1;
302 	tp->snd_nxt = iss + 1;
303 	tp->snd_max = iss + 1;
304 
305 	assign_rxopt(tp, tcpopt);
306 
307 	SOCKBUF_LOCK(&so->so_snd);
308 	if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf)
309 		bufsize = V_tcp_autosndbuf_max;
310 	else
311 		bufsize = sbspace(&so->so_snd);
312 	SOCKBUF_UNLOCK(&so->so_snd);
313 
314 	ftxp.snd_nxt = tp->snd_nxt;
315 	ftxp.rcv_nxt = tp->rcv_nxt;
316 	ftxp.snd_space = bufsize;
317 	ftxp.mss = tp->t_maxseg;
318 	send_flowc_wr(toep, &ftxp);
319 
320 	soisconnected(so);
321 }
322 
323 static int
324 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
325 {
326 	struct wrqe *wr;
327 	struct cpl_rx_data_ack *req;
328 	uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
329 
330 	KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
331 
332 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
333 	if (wr == NULL)
334 		return (0);
335 	req = wrtod(wr);
336 
337 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
338 	req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
339 
340 	t4_wrq_tx(sc, wr);
341 	return (credits);
342 }
343 
344 void
345 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
346 {
347 	struct adapter *sc = tod->tod_softc;
348 	struct inpcb *inp = tp->t_inpcb;
349 	struct socket *so = inp->inp_socket;
350 	struct sockbuf *sb = &so->so_rcv;
351 	struct toepcb *toep = tp->t_toe;
352 	int credits;
353 
354 	INP_WLOCK_ASSERT(inp);
355 
356 	SOCKBUF_LOCK(sb);
357 	KASSERT(toep->sb_cc >= sbused(sb),
358 	    ("%s: sb %p has more data (%d) than last time (%d).",
359 	    __func__, sb, sbused(sb), toep->sb_cc));
360 
361 	toep->rx_credits += toep->sb_cc - sbused(sb);
362 	toep->sb_cc = sbused(sb);
363 
364 	if (toep->rx_credits > 0 &&
365 	    (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
366 	    (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
367 	    toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
368 
369 		credits = send_rx_credits(sc, toep, toep->rx_credits);
370 		toep->rx_credits -= credits;
371 		tp->rcv_wnd += credits;
372 		tp->rcv_adv += credits;
373 	}
374 	SOCKBUF_UNLOCK(sb);
375 }
376 
377 /*
378  * Close a connection by sending a CPL_CLOSE_CON_REQ message.
379  */
380 static int
381 close_conn(struct adapter *sc, struct toepcb *toep)
382 {
383 	struct wrqe *wr;
384 	struct cpl_close_con_req *req;
385 	unsigned int tid = toep->tid;
386 
387 	CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
388 	    toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
389 
390 	if (toep->flags & TPF_FIN_SENT)
391 		return (0);
392 
393 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
394 	    ("%s: flowc_wr not sent for tid %u.", __func__, tid));
395 
396 	wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
397 	if (wr == NULL) {
398 		/* XXX */
399 		panic("%s: allocation failure.", __func__);
400 	}
401 	req = wrtod(wr);
402 
403         req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
404 	    V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
405 	req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
406 	    V_FW_WR_FLOWID(tid));
407         req->wr.wr_lo = cpu_to_be64(0);
408         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
409 	req->rsvd = 0;
410 
411 	toep->flags |= TPF_FIN_SENT;
412 	toep->flags &= ~TPF_SEND_FIN;
413 	t4_l2t_send(sc, wr, toep->l2te);
414 
415 	return (0);
416 }
417 
418 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
419 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
420 
421 /* Maximum amount of immediate data we could stuff in a WR */
422 static inline int
423 max_imm_payload(int tx_credits)
424 {
425 	const int n = 2;	/* Use only up to 2 desc for imm. data WR */
426 
427 	KASSERT(tx_credits >= 0 &&
428 		tx_credits <= MAX_OFLD_TX_CREDITS,
429 		("%s: %d credits", __func__, tx_credits));
430 
431 	if (tx_credits < MIN_OFLD_TX_CREDITS)
432 		return (0);
433 
434 	if (tx_credits >= (n * EQ_ESIZE) / 16)
435 		return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
436 	else
437 		return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
438 }
439 
440 /* Maximum number of SGL entries we could stuff in a WR */
441 static inline int
442 max_dsgl_nsegs(int tx_credits)
443 {
444 	int nseg = 1;	/* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
445 	int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
446 
447 	KASSERT(tx_credits >= 0 &&
448 		tx_credits <= MAX_OFLD_TX_CREDITS,
449 		("%s: %d credits", __func__, tx_credits));
450 
451 	if (tx_credits < MIN_OFLD_TX_CREDITS)
452 		return (0);
453 
454 	nseg += 2 * (sge_pair_credits * 16 / 24);
455 	if ((sge_pair_credits * 16) % 24 == 16)
456 		nseg++;
457 
458 	return (nseg);
459 }
460 
461 static inline void
462 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
463     unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign)
464 {
465 	struct fw_ofld_tx_data_wr *txwr = dst;
466 
467 	txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
468 	    V_FW_WR_IMMDLEN(immdlen));
469 	txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
470 	    V_FW_WR_LEN16(credits));
471 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) |
472 	    V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
473 	txwr->plen = htobe32(plen);
474 
475 	if (txalign > 0) {
476 		struct tcpcb *tp = intotcpcb(toep->inp);
477 
478 		if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi))
479 			txwr->lsodisable_to_flags |=
480 			    htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
481 		else
482 			txwr->lsodisable_to_flags |=
483 			    htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
484 				(tp->t_flags & TF_NODELAY ? 0 :
485 				F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
486 	}
487 }
488 
489 /*
490  * Generate a DSGL from a starting mbuf.  The total number of segments and the
491  * maximum segments in any one mbuf are provided.
492  */
493 static void
494 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
495 {
496 	struct mbuf *m;
497 	struct ulptx_sgl *usgl = dst;
498 	int i, j, rc;
499 	struct sglist sg;
500 	struct sglist_seg segs[n];
501 
502 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
503 
504 	sglist_init(&sg, n, segs);
505 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
506 	    V_ULPTX_NSGE(nsegs));
507 
508 	i = -1;
509 	for (m = start; m != stop; m = m->m_next) {
510 		rc = sglist_append(&sg, mtod(m, void *), m->m_len);
511 		if (__predict_false(rc != 0))
512 			panic("%s: sglist_append %d", __func__, rc);
513 
514 		for (j = 0; j < sg.sg_nseg; i++, j++) {
515 			if (i < 0) {
516 				usgl->len0 = htobe32(segs[j].ss_len);
517 				usgl->addr0 = htobe64(segs[j].ss_paddr);
518 			} else {
519 				usgl->sge[i / 2].len[i & 1] =
520 				    htobe32(segs[j].ss_len);
521 				usgl->sge[i / 2].addr[i & 1] =
522 				    htobe64(segs[j].ss_paddr);
523 			}
524 #ifdef INVARIANTS
525 			nsegs--;
526 #endif
527 		}
528 		sglist_reset(&sg);
529 	}
530 	if (i & 1)
531 		usgl->sge[i / 2].len[1] = htobe32(0);
532 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
533 	    __func__, nsegs, start, stop));
534 }
535 
536 /*
537  * Max number of SGL entries an offload tx work request can have.  This is 41
538  * (1 + 40) for a full 512B work request.
539  * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
540  */
541 #define OFLD_SGL_LEN (41)
542 
543 /*
544  * Send data and/or a FIN to the peer.
545  *
546  * The socket's so_snd buffer consists of a stream of data starting with sb_mb
547  * and linked together with m_next.  sb_sndptr, if set, is the last mbuf that
548  * was transmitted.
549  *
550  * drop indicates the number of bytes that should be dropped from the head of
551  * the send buffer.  It is an optimization that lets do_fw4_ack avoid creating
552  * contention on the send buffer lock (before this change it used to do
553  * sowwakeup and then t4_push_frames right after that when recovering from tx
554  * stalls).  When drop is set this function MUST drop the bytes and wake up any
555  * writers.
556  */
557 void
558 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
559 {
560 	struct mbuf *sndptr, *m, *sb_sndptr;
561 	struct fw_ofld_tx_data_wr *txwr;
562 	struct wrqe *wr;
563 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
564 	struct inpcb *inp = toep->inp;
565 	struct tcpcb *tp = intotcpcb(inp);
566 	struct socket *so = inp->inp_socket;
567 	struct sockbuf *sb = &so->so_snd;
568 	int tx_credits, shove, compl, space, sowwakeup;
569 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
570 
571 	INP_WLOCK_ASSERT(inp);
572 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
573 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
574 
575 	KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
576 	    toep->ulp_mode == ULP_MODE_TCPDDP ||
577 	    toep->ulp_mode == ULP_MODE_RDMA,
578 	    ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
579 
580 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
581 		return;
582 
583 	/*
584 	 * This function doesn't resume by itself.  Someone else must clear the
585 	 * flag and call this function.
586 	 */
587 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
588 		KASSERT(drop == 0,
589 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
590 		return;
591 	}
592 
593 	do {
594 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
595 		max_imm = max_imm_payload(tx_credits);
596 		max_nsegs = max_dsgl_nsegs(tx_credits);
597 
598 		SOCKBUF_LOCK(sb);
599 		sowwakeup = drop;
600 		if (drop) {
601 			sbdrop_locked(sb, drop);
602 			drop = 0;
603 		}
604 		sb_sndptr = sb->sb_sndptr;
605 		sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
606 		plen = 0;
607 		nsegs = 0;
608 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
609 		for (m = sndptr; m != NULL; m = m->m_next) {
610 			int n = sglist_count(mtod(m, void *), m->m_len);
611 
612 			nsegs += n;
613 			plen += m->m_len;
614 
615 			/* This mbuf sent us _over_ the nsegs limit, back out */
616 			if (plen > max_imm && nsegs > max_nsegs) {
617 				nsegs -= n;
618 				plen -= m->m_len;
619 				if (plen == 0) {
620 					/* Too few credits */
621 					toep->flags |= TPF_TX_SUSPENDED;
622 					if (sowwakeup)
623 						sowwakeup_locked(so);
624 					else
625 						SOCKBUF_UNLOCK(sb);
626 					SOCKBUF_UNLOCK_ASSERT(sb);
627 					return;
628 				}
629 				break;
630 			}
631 
632 			if (max_nsegs_1mbuf < n)
633 				max_nsegs_1mbuf = n;
634 			sb_sndptr = m;	/* new sb->sb_sndptr if all goes well */
635 
636 			/* This mbuf put us right at the max_nsegs limit */
637 			if (plen > max_imm && nsegs == max_nsegs) {
638 				m = m->m_next;
639 				break;
640 			}
641 		}
642 
643 		space = sbspace(sb);
644 
645 		if (space <= sb->sb_hiwat * 3 / 8 &&
646 		    toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
647 			compl = 1;
648 		else
649 			compl = 0;
650 
651 		if (sb->sb_flags & SB_AUTOSIZE &&
652 		    V_tcp_do_autosndbuf &&
653 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
654 		    space < sb->sb_hiwat / 8) {
655 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
656 			    V_tcp_autosndbuf_max);
657 
658 			if (!sbreserve_locked(sb, newsize, so, NULL))
659 				sb->sb_flags &= ~SB_AUTOSIZE;
660 			else
661 				sowwakeup = 1;	/* room available */
662 		}
663 		if (sowwakeup)
664 			sowwakeup_locked(so);
665 		else
666 			SOCKBUF_UNLOCK(sb);
667 		SOCKBUF_UNLOCK_ASSERT(sb);
668 
669 		/* nothing to send */
670 		if (plen == 0) {
671 			KASSERT(m == NULL,
672 			    ("%s: nothing to send, but m != NULL", __func__));
673 			break;
674 		}
675 
676 		if (__predict_false(toep->flags & TPF_FIN_SENT))
677 			panic("%s: excess tx.", __func__);
678 
679 		shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
680 		if (plen <= max_imm) {
681 
682 			/* Immediate data tx */
683 
684 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
685 					toep->ofld_txq);
686 			if (wr == NULL) {
687 				/* XXX: how will we recover from this? */
688 				toep->flags |= TPF_TX_SUSPENDED;
689 				return;
690 			}
691 			txwr = wrtod(wr);
692 			credits = howmany(wr->wr_len, 16);
693 			write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
694 			    sc->tt.tx_align);
695 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
696 			nsegs = 0;
697 		} else {
698 			int wr_len;
699 
700 			/* DSGL tx */
701 
702 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
703 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
704 			wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
705 			if (wr == NULL) {
706 				/* XXX: how will we recover from this? */
707 				toep->flags |= TPF_TX_SUSPENDED;
708 				return;
709 			}
710 			txwr = wrtod(wr);
711 			credits = howmany(wr_len, 16);
712 			write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
713 			    sc->tt.tx_align);
714 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
715 			    max_nsegs_1mbuf);
716 			if (wr_len & 0xf) {
717 				uint64_t *pad = (uint64_t *)
718 				    ((uintptr_t)txwr + wr_len);
719 				*pad = 0;
720 			}
721 		}
722 
723 		KASSERT(toep->tx_credits >= credits,
724 			("%s: not enough credits", __func__));
725 
726 		toep->tx_credits -= credits;
727 		toep->tx_nocompl += credits;
728 		toep->plen_nocompl += plen;
729 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
730 		    toep->tx_nocompl >= toep->tx_total / 4)
731 			compl = 1;
732 
733 		if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
734 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
735 			toep->tx_nocompl = 0;
736 			toep->plen_nocompl = 0;
737 		}
738 
739 		tp->snd_nxt += plen;
740 		tp->snd_max += plen;
741 
742 		SOCKBUF_LOCK(sb);
743 		KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
744 		sb->sb_sndptr = sb_sndptr;
745 		SOCKBUF_UNLOCK(sb);
746 
747 		toep->flags |= TPF_TX_DATA_SENT;
748 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
749 			toep->flags |= TPF_TX_SUSPENDED;
750 
751 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
752 		txsd->plen = plen;
753 		txsd->tx_credits = credits;
754 		txsd++;
755 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
756 			toep->txsd_pidx = 0;
757 			txsd = &toep->txsd[0];
758 		}
759 		toep->txsd_avail--;
760 
761 		t4_l2t_send(sc, wr, toep->l2te);
762 	} while (m != NULL);
763 
764 	/* Send a FIN if requested, but only if there's no more data to send */
765 	if (m == NULL && toep->flags & TPF_SEND_FIN)
766 		close_conn(sc, toep);
767 }
768 
769 static inline void
770 rqdrop_locked(struct mbufq *q, int plen)
771 {
772 	struct mbuf *m;
773 
774 	while (plen > 0) {
775 		m = mbufq_dequeue(q);
776 
777 		/* Too many credits. */
778 		MPASS(m != NULL);
779 		M_ASSERTPKTHDR(m);
780 
781 		/* Partial credits. */
782 		MPASS(plen >= m->m_pkthdr.len);
783 
784 		plen -= m->m_pkthdr.len;
785 		m_freem(m);
786 	}
787 }
788 
789 void
790 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
791 {
792 	struct mbuf *sndptr, *m;
793 	struct fw_ofld_tx_data_wr *txwr;
794 	struct wrqe *wr;
795 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
796 	u_int adjusted_plen, ulp_submode;
797 	struct inpcb *inp = toep->inp;
798 	struct tcpcb *tp = intotcpcb(inp);
799 	int tx_credits, shove;
800 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
801 	struct mbufq *pduq = &toep->ulp_pduq;
802 	static const u_int ulp_extra_len[] = {0, 4, 4, 8};
803 
804 	INP_WLOCK_ASSERT(inp);
805 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
806 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
807 	KASSERT(toep->ulp_mode == ULP_MODE_ISCSI,
808 	    ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
809 
810 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
811 		return;
812 
813 	/*
814 	 * This function doesn't resume by itself.  Someone else must clear the
815 	 * flag and call this function.
816 	 */
817 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
818 		KASSERT(drop == 0,
819 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
820 		return;
821 	}
822 
823 	if (drop)
824 		rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
825 
826 	while ((sndptr = mbufq_first(pduq)) != NULL) {
827 		M_ASSERTPKTHDR(sndptr);
828 
829 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
830 		max_imm = max_imm_payload(tx_credits);
831 		max_nsegs = max_dsgl_nsegs(tx_credits);
832 
833 		plen = 0;
834 		nsegs = 0;
835 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
836 		for (m = sndptr; m != NULL; m = m->m_next) {
837 			int n = sglist_count(mtod(m, void *), m->m_len);
838 
839 			nsegs += n;
840 			plen += m->m_len;
841 
842 			/*
843 			 * This mbuf would send us _over_ the nsegs limit.
844 			 * Suspend tx because the PDU can't be sent out.
845 			 */
846 			if (plen > max_imm && nsegs > max_nsegs) {
847 				toep->flags |= TPF_TX_SUSPENDED;
848 				return;
849 			}
850 
851 			if (max_nsegs_1mbuf < n)
852 				max_nsegs_1mbuf = n;
853 		}
854 
855 		if (__predict_false(toep->flags & TPF_FIN_SENT))
856 			panic("%s: excess tx.", __func__);
857 
858 		/*
859 		 * We have a PDU to send.  All of it goes out in one WR so 'm'
860 		 * is NULL.  A PDU's length is always a multiple of 4.
861 		 */
862 		MPASS(m == NULL);
863 		MPASS((plen & 3) == 0);
864 		MPASS(sndptr->m_pkthdr.len == plen);
865 
866 		shove = !(tp->t_flags & TF_MORETOCOME);
867 		ulp_submode = mbuf_ulp_submode(sndptr);
868 		MPASS(ulp_submode < nitems(ulp_extra_len));
869 
870 		/*
871 		 * plen doesn't include header and data digests, which are
872 		 * generated and inserted in the right places by the TOE, but
873 		 * they do occupy TCP sequence space and need to be accounted
874 		 * for.
875 		 */
876 		adjusted_plen = plen + ulp_extra_len[ulp_submode];
877 		if (plen <= max_imm) {
878 
879 			/* Immediate data tx */
880 
881 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
882 					toep->ofld_txq);
883 			if (wr == NULL) {
884 				/* XXX: how will we recover from this? */
885 				toep->flags |= TPF_TX_SUSPENDED;
886 				return;
887 			}
888 			txwr = wrtod(wr);
889 			credits = howmany(wr->wr_len, 16);
890 			write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
891 			    shove, ulp_submode, sc->tt.tx_align);
892 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
893 			nsegs = 0;
894 		} else {
895 			int wr_len;
896 
897 			/* DSGL tx */
898 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
899 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
900 			wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
901 			if (wr == NULL) {
902 				/* XXX: how will we recover from this? */
903 				toep->flags |= TPF_TX_SUSPENDED;
904 				return;
905 			}
906 			txwr = wrtod(wr);
907 			credits = howmany(wr_len, 16);
908 			write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
909 			    shove, ulp_submode, sc->tt.tx_align);
910 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
911 			    max_nsegs_1mbuf);
912 			if (wr_len & 0xf) {
913 				uint64_t *pad = (uint64_t *)
914 				    ((uintptr_t)txwr + wr_len);
915 				*pad = 0;
916 			}
917 		}
918 
919 		KASSERT(toep->tx_credits >= credits,
920 			("%s: not enough credits", __func__));
921 
922 		m = mbufq_dequeue(pduq);
923 		MPASS(m == sndptr);
924 		mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
925 
926 		toep->tx_credits -= credits;
927 		toep->tx_nocompl += credits;
928 		toep->plen_nocompl += plen;
929 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
930 		    toep->tx_nocompl >= toep->tx_total / 4) {
931 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
932 			toep->tx_nocompl = 0;
933 			toep->plen_nocompl = 0;
934 		}
935 
936 		tp->snd_nxt += adjusted_plen;
937 		tp->snd_max += adjusted_plen;
938 
939 		toep->flags |= TPF_TX_DATA_SENT;
940 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
941 			toep->flags |= TPF_TX_SUSPENDED;
942 
943 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
944 		txsd->plen = plen;
945 		txsd->tx_credits = credits;
946 		txsd++;
947 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
948 			toep->txsd_pidx = 0;
949 			txsd = &toep->txsd[0];
950 		}
951 		toep->txsd_avail--;
952 
953 		t4_l2t_send(sc, wr, toep->l2te);
954 	}
955 
956 	/* Send a FIN if requested, but only if there are no more PDUs to send */
957 	if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
958 		close_conn(sc, toep);
959 }
960 
961 int
962 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
963 {
964 	struct adapter *sc = tod->tod_softc;
965 #ifdef INVARIANTS
966 	struct inpcb *inp = tp->t_inpcb;
967 #endif
968 	struct toepcb *toep = tp->t_toe;
969 
970 	INP_WLOCK_ASSERT(inp);
971 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
972 	    ("%s: inp %p dropped.", __func__, inp));
973 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
974 
975 	if (toep->ulp_mode == ULP_MODE_ISCSI)
976 		t4_push_pdus(sc, toep, 0);
977 	else
978 		t4_push_frames(sc, toep, 0);
979 
980 	return (0);
981 }
982 
983 int
984 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
985 {
986 	struct adapter *sc = tod->tod_softc;
987 #ifdef INVARIANTS
988 	struct inpcb *inp = tp->t_inpcb;
989 #endif
990 	struct toepcb *toep = tp->t_toe;
991 
992 	INP_WLOCK_ASSERT(inp);
993 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
994 	    ("%s: inp %p dropped.", __func__, inp));
995 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
996 
997 	toep->flags |= TPF_SEND_FIN;
998 	if (tp->t_state >= TCPS_ESTABLISHED) {
999 		if (toep->ulp_mode == ULP_MODE_ISCSI)
1000 			t4_push_pdus(sc, toep, 0);
1001 		else
1002 			t4_push_frames(sc, toep, 0);
1003 	}
1004 
1005 	return (0);
1006 }
1007 
1008 int
1009 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1010 {
1011 	struct adapter *sc = tod->tod_softc;
1012 #if defined(INVARIANTS)
1013 	struct inpcb *inp = tp->t_inpcb;
1014 #endif
1015 	struct toepcb *toep = tp->t_toe;
1016 
1017 	INP_WLOCK_ASSERT(inp);
1018 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1019 	    ("%s: inp %p dropped.", __func__, inp));
1020 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1021 
1022 	/* hmmmm */
1023 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1024 	    ("%s: flowc for tid %u [%s] not sent already",
1025 	    __func__, toep->tid, tcpstates[tp->t_state]));
1026 
1027 	send_reset(sc, toep, 0);
1028 	return (0);
1029 }
1030 
1031 /*
1032  * Peer has sent us a FIN.
1033  */
1034 static int
1035 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1036 {
1037 	struct adapter *sc = iq->adapter;
1038 	const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1039 	unsigned int tid = GET_TID(cpl);
1040 	struct toepcb *toep = lookup_tid(sc, tid);
1041 	struct inpcb *inp = toep->inp;
1042 	struct tcpcb *tp = NULL;
1043 	struct socket *so;
1044 	struct sockbuf *sb;
1045 #ifdef INVARIANTS
1046 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1047 #endif
1048 
1049 	KASSERT(opcode == CPL_PEER_CLOSE,
1050 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1051 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1052 
1053 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1054 #ifdef INVARIANTS
1055 		struct synq_entry *synqe = (void *)toep;
1056 
1057 		INP_WLOCK(synqe->lctx->inp);
1058 		if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1059 			KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1060 			    ("%s: listen socket closed but tid %u not aborted.",
1061 			    __func__, tid));
1062 		} else {
1063 			/*
1064 			 * do_pass_accept_req is still running and will
1065 			 * eventually take care of this tid.
1066 			 */
1067 		}
1068 		INP_WUNLOCK(synqe->lctx->inp);
1069 #endif
1070 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1071 		    toep, toep->flags);
1072 		return (0);
1073 	}
1074 
1075 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1076 
1077 	INP_INFO_RLOCK(&V_tcbinfo);
1078 	INP_WLOCK(inp);
1079 	tp = intotcpcb(inp);
1080 
1081 	CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1082 	    tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1083 
1084 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1085 		goto done;
1086 
1087 	tp->rcv_nxt++;	/* FIN */
1088 
1089 	so = inp->inp_socket;
1090 	sb = &so->so_rcv;
1091 	SOCKBUF_LOCK(sb);
1092 	if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
1093 		handle_ddp_close(toep, tp, sb, cpl->rcv_nxt);
1094 	}
1095 	socantrcvmore_locked(so);	/* unlocks the sockbuf */
1096 
1097 	if (toep->ulp_mode != ULP_MODE_RDMA) {
1098 		KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1099 	    		("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1100 	    		be32toh(cpl->rcv_nxt)));
1101 	}
1102 
1103 	switch (tp->t_state) {
1104 	case TCPS_SYN_RECEIVED:
1105 		tp->t_starttime = ticks;
1106 		/* FALLTHROUGH */
1107 
1108 	case TCPS_ESTABLISHED:
1109 		tp->t_state = TCPS_CLOSE_WAIT;
1110 		break;
1111 
1112 	case TCPS_FIN_WAIT_1:
1113 		tp->t_state = TCPS_CLOSING;
1114 		break;
1115 
1116 	case TCPS_FIN_WAIT_2:
1117 		tcp_twstart(tp);
1118 		INP_UNLOCK_ASSERT(inp);	 /* safe, we have a ref on the inp */
1119 		INP_INFO_RUNLOCK(&V_tcbinfo);
1120 
1121 		INP_WLOCK(inp);
1122 		final_cpl_received(toep);
1123 		return (0);
1124 
1125 	default:
1126 		log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1127 		    __func__, tid, tp->t_state);
1128 	}
1129 done:
1130 	INP_WUNLOCK(inp);
1131 	INP_INFO_RUNLOCK(&V_tcbinfo);
1132 	return (0);
1133 }
1134 
1135 /*
1136  * Peer has ACK'd our FIN.
1137  */
1138 static int
1139 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1140     struct mbuf *m)
1141 {
1142 	struct adapter *sc = iq->adapter;
1143 	const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1144 	unsigned int tid = GET_TID(cpl);
1145 	struct toepcb *toep = lookup_tid(sc, tid);
1146 	struct inpcb *inp = toep->inp;
1147 	struct tcpcb *tp = NULL;
1148 	struct socket *so = NULL;
1149 #ifdef INVARIANTS
1150 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1151 #endif
1152 
1153 	KASSERT(opcode == CPL_CLOSE_CON_RPL,
1154 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1155 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1156 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1157 
1158 	INP_INFO_RLOCK(&V_tcbinfo);
1159 	INP_WLOCK(inp);
1160 	tp = intotcpcb(inp);
1161 
1162 	CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1163 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1164 
1165 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1166 		goto done;
1167 
1168 	so = inp->inp_socket;
1169 	tp->snd_una = be32toh(cpl->snd_nxt) - 1;	/* exclude FIN */
1170 
1171 	switch (tp->t_state) {
1172 	case TCPS_CLOSING:	/* see TCPS_FIN_WAIT_2 in do_peer_close too */
1173 		tcp_twstart(tp);
1174 release:
1175 		INP_UNLOCK_ASSERT(inp);	/* safe, we have a ref on the  inp */
1176 		INP_INFO_RUNLOCK(&V_tcbinfo);
1177 
1178 		INP_WLOCK(inp);
1179 		final_cpl_received(toep);	/* no more CPLs expected */
1180 
1181 		return (0);
1182 	case TCPS_LAST_ACK:
1183 		if (tcp_close(tp))
1184 			INP_WUNLOCK(inp);
1185 		goto release;
1186 
1187 	case TCPS_FIN_WAIT_1:
1188 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1189 			soisdisconnected(so);
1190 		tp->t_state = TCPS_FIN_WAIT_2;
1191 		break;
1192 
1193 	default:
1194 		log(LOG_ERR,
1195 		    "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1196 		    __func__, tid, tcpstates[tp->t_state]);
1197 	}
1198 done:
1199 	INP_WUNLOCK(inp);
1200 	INP_INFO_RUNLOCK(&V_tcbinfo);
1201 	return (0);
1202 }
1203 
1204 void
1205 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1206     int rst_status)
1207 {
1208 	struct wrqe *wr;
1209 	struct cpl_abort_rpl *cpl;
1210 
1211 	wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1212 	if (wr == NULL) {
1213 		/* XXX */
1214 		panic("%s: allocation failure.", __func__);
1215 	}
1216 	cpl = wrtod(wr);
1217 
1218 	INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1219 	cpl->cmd = rst_status;
1220 
1221 	t4_wrq_tx(sc, wr);
1222 }
1223 
1224 static int
1225 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1226 {
1227 	switch (abort_reason) {
1228 	case CPL_ERR_BAD_SYN:
1229 	case CPL_ERR_CONN_RESET:
1230 		return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1231 	case CPL_ERR_XMIT_TIMEDOUT:
1232 	case CPL_ERR_PERSIST_TIMEDOUT:
1233 	case CPL_ERR_FINWAIT2_TIMEDOUT:
1234 	case CPL_ERR_KEEPALIVE_TIMEDOUT:
1235 		return (ETIMEDOUT);
1236 	default:
1237 		return (EIO);
1238 	}
1239 }
1240 
1241 /*
1242  * TCP RST from the peer, timeout, or some other such critical error.
1243  */
1244 static int
1245 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1246 {
1247 	struct adapter *sc = iq->adapter;
1248 	const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1249 	unsigned int tid = GET_TID(cpl);
1250 	struct toepcb *toep = lookup_tid(sc, tid);
1251 	struct sge_wrq *ofld_txq = toep->ofld_txq;
1252 	struct inpcb *inp;
1253 	struct tcpcb *tp;
1254 #ifdef INVARIANTS
1255 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1256 #endif
1257 
1258 	KASSERT(opcode == CPL_ABORT_REQ_RSS,
1259 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1260 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1261 
1262 	if (toep->flags & TPF_SYNQE)
1263 		return (do_abort_req_synqe(iq, rss, m));
1264 
1265 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1266 
1267 	if (negative_advice(cpl->status)) {
1268 		CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1269 		    __func__, cpl->status, tid, toep->flags);
1270 		return (0);	/* Ignore negative advice */
1271 	}
1272 
1273 	inp = toep->inp;
1274 	INP_INFO_RLOCK(&V_tcbinfo);	/* for tcp_close */
1275 	INP_WLOCK(inp);
1276 
1277 	tp = intotcpcb(inp);
1278 
1279 	CTR6(KTR_CXGBE,
1280 	    "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1281 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1282 	    inp->inp_flags, cpl->status);
1283 
1284 	/*
1285 	 * If we'd initiated an abort earlier the reply to it is responsible for
1286 	 * cleaning up resources.  Otherwise we tear everything down right here
1287 	 * right now.  We owe the T4 a CPL_ABORT_RPL no matter what.
1288 	 */
1289 	if (toep->flags & TPF_ABORT_SHUTDOWN) {
1290 		INP_WUNLOCK(inp);
1291 		goto done;
1292 	}
1293 	toep->flags |= TPF_ABORT_SHUTDOWN;
1294 
1295 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1296 		struct socket *so = inp->inp_socket;
1297 
1298 		if (so != NULL)
1299 			so_error_set(so, abort_status_to_errno(tp,
1300 			    cpl->status));
1301 		tp = tcp_close(tp);
1302 		if (tp == NULL)
1303 			INP_WLOCK(inp);	/* re-acquire */
1304 	}
1305 
1306 	final_cpl_received(toep);
1307 done:
1308 	INP_INFO_RUNLOCK(&V_tcbinfo);
1309 	send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1310 	return (0);
1311 }
1312 
1313 /*
1314  * Reply to the CPL_ABORT_REQ (send_reset)
1315  */
1316 static int
1317 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1318 {
1319 	struct adapter *sc = iq->adapter;
1320 	const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1321 	unsigned int tid = GET_TID(cpl);
1322 	struct toepcb *toep = lookup_tid(sc, tid);
1323 	struct inpcb *inp = toep->inp;
1324 #ifdef INVARIANTS
1325 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1326 #endif
1327 
1328 	KASSERT(opcode == CPL_ABORT_RPL_RSS,
1329 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1330 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1331 
1332 	if (toep->flags & TPF_SYNQE)
1333 		return (do_abort_rpl_synqe(iq, rss, m));
1334 
1335 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1336 
1337 	CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1338 	    __func__, tid, toep, inp, cpl->status);
1339 
1340 	KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1341 	    ("%s: wasn't expecting abort reply", __func__));
1342 
1343 	INP_WLOCK(inp);
1344 	final_cpl_received(toep);
1345 
1346 	return (0);
1347 }
1348 
1349 static int
1350 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1351 {
1352 	struct adapter *sc = iq->adapter;
1353 	const struct cpl_rx_data *cpl = mtod(m, const void *);
1354 	unsigned int tid = GET_TID(cpl);
1355 	struct toepcb *toep = lookup_tid(sc, tid);
1356 	struct inpcb *inp = toep->inp;
1357 	struct tcpcb *tp;
1358 	struct socket *so;
1359 	struct sockbuf *sb;
1360 	int len;
1361 	uint32_t ddp_placed = 0;
1362 
1363 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1364 #ifdef INVARIANTS
1365 		struct synq_entry *synqe = (void *)toep;
1366 
1367 		INP_WLOCK(synqe->lctx->inp);
1368 		if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1369 			KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1370 			    ("%s: listen socket closed but tid %u not aborted.",
1371 			    __func__, tid));
1372 		} else {
1373 			/*
1374 			 * do_pass_accept_req is still running and will
1375 			 * eventually take care of this tid.
1376 			 */
1377 		}
1378 		INP_WUNLOCK(synqe->lctx->inp);
1379 #endif
1380 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1381 		    toep, toep->flags);
1382 		m_freem(m);
1383 		return (0);
1384 	}
1385 
1386 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1387 
1388 	/* strip off CPL header */
1389 	m_adj(m, sizeof(*cpl));
1390 	len = m->m_pkthdr.len;
1391 
1392 	INP_WLOCK(inp);
1393 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1394 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1395 		    __func__, tid, len, inp->inp_flags);
1396 		INP_WUNLOCK(inp);
1397 		m_freem(m);
1398 		return (0);
1399 	}
1400 
1401 	tp = intotcpcb(inp);
1402 
1403 	if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1404 		ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1405 
1406 	tp->rcv_nxt += len;
1407 	KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
1408 	tp->rcv_wnd -= len;
1409 	tp->t_rcvtime = ticks;
1410 
1411 	so = inp_inpcbtosocket(inp);
1412 	sb = &so->so_rcv;
1413 	SOCKBUF_LOCK(sb);
1414 
1415 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1416 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1417 		    __func__, tid, len);
1418 		m_freem(m);
1419 		SOCKBUF_UNLOCK(sb);
1420 		INP_WUNLOCK(inp);
1421 
1422 		INP_INFO_RLOCK(&V_tcbinfo);
1423 		INP_WLOCK(inp);
1424 		tp = tcp_drop(tp, ECONNRESET);
1425 		if (tp)
1426 			INP_WUNLOCK(inp);
1427 		INP_INFO_RUNLOCK(&V_tcbinfo);
1428 
1429 		return (0);
1430 	}
1431 
1432 	/* receive buffer autosize */
1433 	if (sb->sb_flags & SB_AUTOSIZE &&
1434 	    V_tcp_do_autorcvbuf &&
1435 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1436 	    len > (sbspace(sb) / 8 * 7)) {
1437 		unsigned int hiwat = sb->sb_hiwat;
1438 		unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1439 		    V_tcp_autorcvbuf_max);
1440 
1441 		if (!sbreserve_locked(sb, newsize, so, NULL))
1442 			sb->sb_flags &= ~SB_AUTOSIZE;
1443 		else
1444 			toep->rx_credits += newsize - hiwat;
1445 	}
1446 
1447 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1448 		int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1449 
1450 		if (changed) {
1451 			if (toep->ddp_flags & DDP_SC_REQ)
1452 				toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1453 			else {
1454 				KASSERT(cpl->ddp_off == 1,
1455 				    ("%s: DDP switched on by itself.",
1456 				    __func__));
1457 
1458 				/* Fell out of DDP mode */
1459 				toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1460 				    DDP_BUF1_ACTIVE);
1461 
1462 				if (ddp_placed)
1463 					insert_ddp_data(toep, ddp_placed);
1464 			}
1465 		}
1466 
1467 		if ((toep->ddp_flags & DDP_OK) == 0 &&
1468 		    time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1469 			toep->ddp_score = DDP_LOW_SCORE;
1470 			toep->ddp_flags |= DDP_OK;
1471 			CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1472 			    __func__, tid, time_uptime);
1473 		}
1474 
1475 		if (toep->ddp_flags & DDP_ON) {
1476 
1477 			/*
1478 			 * CPL_RX_DATA with DDP on can only be an indicate.  Ask
1479 			 * soreceive to post a buffer or disable DDP.  The
1480 			 * payload that arrived in this indicate is appended to
1481 			 * the socket buffer as usual.
1482 			 */
1483 
1484 #if 0
1485 			CTR5(KTR_CXGBE,
1486 			    "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1487 			    __func__, tid, toep->flags, be32toh(cpl->seq), len);
1488 #endif
1489 			sb->sb_flags |= SB_DDP_INDICATE;
1490 		} else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1491 		    tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1492 
1493 			/*
1494 			 * DDP allowed but isn't on (and a request to switch it
1495 			 * on isn't pending either), and conditions are ripe for
1496 			 * it to work.  Switch it on.
1497 			 */
1498 
1499 			enable_ddp(sc, toep);
1500 		}
1501 	}
1502 
1503 	KASSERT(toep->sb_cc >= sbused(sb),
1504 	    ("%s: sb %p has more data (%d) than last time (%d).",
1505 	    __func__, sb, sbused(sb), toep->sb_cc));
1506 	toep->rx_credits += toep->sb_cc - sbused(sb);
1507 	sbappendstream_locked(sb, m, 0);
1508 	toep->sb_cc = sbused(sb);
1509 	if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1510 		int credits;
1511 
1512 		credits = send_rx_credits(sc, toep, toep->rx_credits);
1513 		toep->rx_credits -= credits;
1514 		tp->rcv_wnd += credits;
1515 		tp->rcv_adv += credits;
1516 	}
1517 	sorwakeup_locked(so);
1518 	SOCKBUF_UNLOCK_ASSERT(sb);
1519 
1520 	INP_WUNLOCK(inp);
1521 	return (0);
1522 }
1523 
1524 #define S_CPL_FW4_ACK_OPCODE    24
1525 #define M_CPL_FW4_ACK_OPCODE    0xff
1526 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1527 #define G_CPL_FW4_ACK_OPCODE(x) \
1528     (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1529 
1530 #define S_CPL_FW4_ACK_FLOWID    0
1531 #define M_CPL_FW4_ACK_FLOWID    0xffffff
1532 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1533 #define G_CPL_FW4_ACK_FLOWID(x) \
1534     (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1535 
1536 #define S_CPL_FW4_ACK_CR        24
1537 #define M_CPL_FW4_ACK_CR        0xff
1538 #define V_CPL_FW4_ACK_CR(x)     ((x) << S_CPL_FW4_ACK_CR)
1539 #define G_CPL_FW4_ACK_CR(x)     (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1540 
1541 #define S_CPL_FW4_ACK_SEQVAL    0
1542 #define M_CPL_FW4_ACK_SEQVAL    0x1
1543 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1544 #define G_CPL_FW4_ACK_SEQVAL(x) \
1545     (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1546 #define F_CPL_FW4_ACK_SEQVAL    V_CPL_FW4_ACK_SEQVAL(1U)
1547 
1548 static int
1549 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1550 {
1551 	struct adapter *sc = iq->adapter;
1552 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1553 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1554 	struct toepcb *toep = lookup_tid(sc, tid);
1555 	struct inpcb *inp;
1556 	struct tcpcb *tp;
1557 	struct socket *so;
1558 	uint8_t credits = cpl->credits;
1559 	struct ofld_tx_sdesc *txsd;
1560 	int plen;
1561 #ifdef INVARIANTS
1562 	unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1563 #endif
1564 
1565 	/*
1566 	 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1567 	 * now this comes back carrying the credits for the flowc.
1568 	 */
1569 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1570 		KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1571 		    ("%s: credits for a synq entry %p", __func__, toep));
1572 		return (0);
1573 	}
1574 
1575 	inp = toep->inp;
1576 
1577 	KASSERT(opcode == CPL_FW4_ACK,
1578 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1579 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1580 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1581 
1582 	INP_WLOCK(inp);
1583 
1584 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1585 		INP_WUNLOCK(inp);
1586 		return (0);
1587 	}
1588 
1589 	KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1590 	    ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1591 
1592 	tp = intotcpcb(inp);
1593 
1594 	if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1595 		tcp_seq snd_una = be32toh(cpl->snd_una);
1596 
1597 #ifdef INVARIANTS
1598 		if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1599 			log(LOG_ERR,
1600 			    "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1601 			    __func__, snd_una, toep->tid, tp->snd_una);
1602 		}
1603 #endif
1604 
1605 		if (tp->snd_una != snd_una) {
1606 			tp->snd_una = snd_una;
1607 			tp->ts_recent_age = tcp_ts_getticks();
1608 		}
1609 	}
1610 
1611 	so = inp->inp_socket;
1612 	txsd = &toep->txsd[toep->txsd_cidx];
1613 	plen = 0;
1614 	while (credits) {
1615 		KASSERT(credits >= txsd->tx_credits,
1616 		    ("%s: too many (or partial) credits", __func__));
1617 		credits -= txsd->tx_credits;
1618 		toep->tx_credits += txsd->tx_credits;
1619 		plen += txsd->plen;
1620 		txsd++;
1621 		toep->txsd_avail++;
1622 		KASSERT(toep->txsd_avail <= toep->txsd_total,
1623 		    ("%s: txsd avail > total", __func__));
1624 		if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1625 			txsd = &toep->txsd[0];
1626 			toep->txsd_cidx = 0;
1627 		}
1628 	}
1629 
1630 	if (toep->tx_credits == toep->tx_total) {
1631 		toep->tx_nocompl = 0;
1632 		toep->plen_nocompl = 0;
1633 	}
1634 
1635 	if (toep->flags & TPF_TX_SUSPENDED &&
1636 	    toep->tx_credits >= toep->tx_total / 4) {
1637 		toep->flags &= ~TPF_TX_SUSPENDED;
1638 		if (toep->ulp_mode == ULP_MODE_ISCSI)
1639 			t4_push_pdus(sc, toep, plen);
1640 		else
1641 			t4_push_frames(sc, toep, plen);
1642 	} else if (plen > 0) {
1643 		struct sockbuf *sb = &so->so_snd;
1644 		int sbu;
1645 
1646 		SOCKBUF_LOCK(sb);
1647 		sbu = sbused(sb);
1648 		if (toep->ulp_mode == ULP_MODE_ISCSI) {
1649 
1650 			if (__predict_false(sbu > 0)) {
1651 				/*
1652 				 * The data trasmitted before the tid's ULP mode
1653 				 * changed to ISCSI is still in so_snd.
1654 				 * Incoming credits should account for so_snd
1655 				 * first.
1656 				 */
1657 				sbdrop_locked(sb, min(sbu, plen));
1658 				plen -= min(sbu, plen);
1659 			}
1660 			sowwakeup_locked(so);	/* unlocks so_snd */
1661 			rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1662 		} else {
1663 			sbdrop_locked(sb, plen);
1664 			sowwakeup_locked(so);	/* unlocks so_snd */
1665 		}
1666 		SOCKBUF_UNLOCK_ASSERT(sb);
1667 	}
1668 
1669 	INP_WUNLOCK(inp);
1670 
1671 	return (0);
1672 }
1673 
1674 static int
1675 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1676 {
1677 	struct adapter *sc = iq->adapter;
1678 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1679 	unsigned int tid = GET_TID(cpl);
1680 #ifdef INVARIANTS
1681 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1682 #endif
1683 
1684 	KASSERT(opcode == CPL_SET_TCB_RPL,
1685 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1686 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1687 
1688 	if (is_ftid(sc, tid))
1689 		return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1690 
1691 	/*
1692 	 * TOM and/or other ULPs don't request replies for CPL_SET_TCB or
1693 	 * CPL_SET_TCB_FIELD requests.  This can easily change and when it does
1694 	 * the dispatch code will go here.
1695 	 */
1696 #ifdef INVARIANTS
1697 	panic("%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p", __func__,
1698 	    tid, iq);
1699 #else
1700 	log(LOG_ERR, "%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p\n",
1701 	    __func__, tid, iq);
1702 #endif
1703 
1704 	return (0);
1705 }
1706 
1707 void
1708 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1709     uint16_t word, uint64_t mask, uint64_t val)
1710 {
1711 	struct wrqe *wr;
1712 	struct cpl_set_tcb_field *req;
1713 
1714 	wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1715 	if (wr == NULL) {
1716 		/* XXX */
1717 		panic("%s: allocation failure.", __func__);
1718 	}
1719 	req = wrtod(wr);
1720 
1721 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1722 	req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1723 	    V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1724 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1725 	req->mask = htobe64(mask);
1726 	req->val = htobe64(val);
1727 
1728 	t4_wrq_tx(sc, wr);
1729 }
1730 
1731 void
1732 t4_init_cpl_io_handlers(struct adapter *sc)
1733 {
1734 
1735 	t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1736 	t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1737 	t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1738 	t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1739 	t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1740 	t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1741 	t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1742 }
1743 
1744 void
1745 t4_uninit_cpl_io_handlers(struct adapter *sc)
1746 {
1747 
1748 	t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);
1749 }
1750 #endif
1751