xref: /freebsd/sys/dev/cxgbe/tom/t4_cpl_io.c (revision 8ef24a0d4b28fe230e20637f56869cc4148cd2ca)
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 #define TCPSTATES
49 #include <netinet/tcp_fsm.h>
50 #include <netinet/tcp_seq.h>
51 #include <netinet/tcp_var.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 = 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 	}
234 
235 	if (G_TCPOPT_SACK(opt))
236 		tp->t_flags |= TF_SACK_PERMIT;	/* should already be set */
237 	else
238 		tp->t_flags &= ~TF_SACK_PERMIT;	/* sack disallowed by peer */
239 
240 	if (G_TCPOPT_WSCALE_OK(opt))
241 		tp->t_flags |= TF_RCVD_SCALE;
242 
243 	/* Doing window scaling? */
244 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
245 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
246 		tp->rcv_scale = tp->request_r_scale;
247 		tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
248 	}
249 }
250 
251 /*
252  * Completes some final bits of initialization for just established connections
253  * and changes their state to TCPS_ESTABLISHED.
254  *
255  * The ISNs are from after the exchange of SYNs.  i.e., the true ISN + 1.
256  */
257 void
258 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
259     uint16_t opt)
260 {
261 	struct inpcb *inp = toep->inp;
262 	struct socket *so = inp->inp_socket;
263 	struct tcpcb *tp = intotcpcb(inp);
264 	long bufsize;
265 	uint32_t iss = be32toh(snd_isn) - 1;	/* true ISS */
266 	uint32_t irs = be32toh(rcv_isn) - 1;	/* true IRS */
267 	uint16_t tcpopt = be16toh(opt);
268 	struct flowc_tx_params ftxp;
269 
270 	CURVNET_SET(so->so_vnet);
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 	CURVNET_RESTORE();
322 }
323 
324 static int
325 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
326 {
327 	struct wrqe *wr;
328 	struct cpl_rx_data_ack *req;
329 	uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
330 
331 	KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
332 
333 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
334 	if (wr == NULL)
335 		return (0);
336 	req = wrtod(wr);
337 
338 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
339 	req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
340 
341 	t4_wrq_tx(sc, wr);
342 	return (credits);
343 }
344 
345 void
346 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp)
347 {
348 	struct adapter *sc = tod->tod_softc;
349 	struct inpcb *inp = tp->t_inpcb;
350 	struct socket *so = inp->inp_socket;
351 	struct sockbuf *sb = &so->so_rcv;
352 	struct toepcb *toep = tp->t_toe;
353 	int credits;
354 
355 	INP_WLOCK_ASSERT(inp);
356 
357 	SOCKBUF_LOCK_ASSERT(sb);
358 	KASSERT(toep->sb_cc >= sbused(sb),
359 	    ("%s: sb %p has more data (%d) than last time (%d).",
360 	    __func__, sb, sbused(sb), toep->sb_cc));
361 
362 	toep->rx_credits += toep->sb_cc - sbused(sb);
363 	toep->sb_cc = sbused(sb);
364 
365 	if (toep->rx_credits > 0 &&
366 	    (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
367 	    (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
368 	    toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
369 
370 		credits = send_rx_credits(sc, toep, toep->rx_credits);
371 		toep->rx_credits -= credits;
372 		tp->rcv_wnd += credits;
373 		tp->rcv_adv += credits;
374 	}
375 }
376 
377 void
378 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
379 {
380 	struct inpcb *inp = tp->t_inpcb;
381 	struct socket *so = inp->inp_socket;
382 	struct sockbuf *sb = &so->so_rcv;
383 
384 	SOCKBUF_LOCK(sb);
385 	t4_rcvd_locked(tod, tp);
386 	SOCKBUF_UNLOCK(sb);
387 }
388 
389 /*
390  * Close a connection by sending a CPL_CLOSE_CON_REQ message.
391  */
392 static int
393 close_conn(struct adapter *sc, struct toepcb *toep)
394 {
395 	struct wrqe *wr;
396 	struct cpl_close_con_req *req;
397 	unsigned int tid = toep->tid;
398 
399 	CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
400 	    toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
401 
402 	if (toep->flags & TPF_FIN_SENT)
403 		return (0);
404 
405 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
406 	    ("%s: flowc_wr not sent for tid %u.", __func__, tid));
407 
408 	wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
409 	if (wr == NULL) {
410 		/* XXX */
411 		panic("%s: allocation failure.", __func__);
412 	}
413 	req = wrtod(wr);
414 
415         req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
416 	    V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
417 	req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
418 	    V_FW_WR_FLOWID(tid));
419         req->wr.wr_lo = cpu_to_be64(0);
420         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
421 	req->rsvd = 0;
422 
423 	toep->flags |= TPF_FIN_SENT;
424 	toep->flags &= ~TPF_SEND_FIN;
425 	t4_l2t_send(sc, wr, toep->l2te);
426 
427 	return (0);
428 }
429 
430 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
431 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
432 
433 /* Maximum amount of immediate data we could stuff in a WR */
434 static inline int
435 max_imm_payload(int tx_credits)
436 {
437 	const int n = 2;	/* Use only up to 2 desc for imm. data WR */
438 
439 	KASSERT(tx_credits >= 0 &&
440 		tx_credits <= MAX_OFLD_TX_CREDITS,
441 		("%s: %d credits", __func__, tx_credits));
442 
443 	if (tx_credits < MIN_OFLD_TX_CREDITS)
444 		return (0);
445 
446 	if (tx_credits >= (n * EQ_ESIZE) / 16)
447 		return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
448 	else
449 		return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
450 }
451 
452 /* Maximum number of SGL entries we could stuff in a WR */
453 static inline int
454 max_dsgl_nsegs(int tx_credits)
455 {
456 	int nseg = 1;	/* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
457 	int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
458 
459 	KASSERT(tx_credits >= 0 &&
460 		tx_credits <= MAX_OFLD_TX_CREDITS,
461 		("%s: %d credits", __func__, tx_credits));
462 
463 	if (tx_credits < MIN_OFLD_TX_CREDITS)
464 		return (0);
465 
466 	nseg += 2 * (sge_pair_credits * 16 / 24);
467 	if ((sge_pair_credits * 16) % 24 == 16)
468 		nseg++;
469 
470 	return (nseg);
471 }
472 
473 static inline void
474 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
475     unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign)
476 {
477 	struct fw_ofld_tx_data_wr *txwr = dst;
478 
479 	txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
480 	    V_FW_WR_IMMDLEN(immdlen));
481 	txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
482 	    V_FW_WR_LEN16(credits));
483 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) |
484 	    V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
485 	txwr->plen = htobe32(plen);
486 
487 	if (txalign > 0) {
488 		struct tcpcb *tp = intotcpcb(toep->inp);
489 
490 		if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi))
491 			txwr->lsodisable_to_flags |=
492 			    htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
493 		else
494 			txwr->lsodisable_to_flags |=
495 			    htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
496 				(tp->t_flags & TF_NODELAY ? 0 :
497 				F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
498 	}
499 }
500 
501 /*
502  * Generate a DSGL from a starting mbuf.  The total number of segments and the
503  * maximum segments in any one mbuf are provided.
504  */
505 static void
506 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
507 {
508 	struct mbuf *m;
509 	struct ulptx_sgl *usgl = dst;
510 	int i, j, rc;
511 	struct sglist sg;
512 	struct sglist_seg segs[n];
513 
514 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
515 
516 	sglist_init(&sg, n, segs);
517 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
518 	    V_ULPTX_NSGE(nsegs));
519 
520 	i = -1;
521 	for (m = start; m != stop; m = m->m_next) {
522 		rc = sglist_append(&sg, mtod(m, void *), m->m_len);
523 		if (__predict_false(rc != 0))
524 			panic("%s: sglist_append %d", __func__, rc);
525 
526 		for (j = 0; j < sg.sg_nseg; i++, j++) {
527 			if (i < 0) {
528 				usgl->len0 = htobe32(segs[j].ss_len);
529 				usgl->addr0 = htobe64(segs[j].ss_paddr);
530 			} else {
531 				usgl->sge[i / 2].len[i & 1] =
532 				    htobe32(segs[j].ss_len);
533 				usgl->sge[i / 2].addr[i & 1] =
534 				    htobe64(segs[j].ss_paddr);
535 			}
536 #ifdef INVARIANTS
537 			nsegs--;
538 #endif
539 		}
540 		sglist_reset(&sg);
541 	}
542 	if (i & 1)
543 		usgl->sge[i / 2].len[1] = htobe32(0);
544 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
545 	    __func__, nsegs, start, stop));
546 }
547 
548 /*
549  * Max number of SGL entries an offload tx work request can have.  This is 41
550  * (1 + 40) for a full 512B work request.
551  * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
552  */
553 #define OFLD_SGL_LEN (41)
554 
555 /*
556  * Send data and/or a FIN to the peer.
557  *
558  * The socket's so_snd buffer consists of a stream of data starting with sb_mb
559  * and linked together with m_next.  sb_sndptr, if set, is the last mbuf that
560  * was transmitted.
561  *
562  * drop indicates the number of bytes that should be dropped from the head of
563  * the send buffer.  It is an optimization that lets do_fw4_ack avoid creating
564  * contention on the send buffer lock (before this change it used to do
565  * sowwakeup and then t4_push_frames right after that when recovering from tx
566  * stalls).  When drop is set this function MUST drop the bytes and wake up any
567  * writers.
568  */
569 void
570 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
571 {
572 	struct mbuf *sndptr, *m, *sb_sndptr;
573 	struct fw_ofld_tx_data_wr *txwr;
574 	struct wrqe *wr;
575 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
576 	struct inpcb *inp = toep->inp;
577 	struct tcpcb *tp = intotcpcb(inp);
578 	struct socket *so = inp->inp_socket;
579 	struct sockbuf *sb = &so->so_snd;
580 	int tx_credits, shove, compl, space, sowwakeup;
581 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
582 
583 	INP_WLOCK_ASSERT(inp);
584 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
585 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
586 
587 	KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
588 	    toep->ulp_mode == ULP_MODE_TCPDDP ||
589 	    toep->ulp_mode == ULP_MODE_RDMA,
590 	    ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
591 
592 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
593 		return;
594 
595 	/*
596 	 * This function doesn't resume by itself.  Someone else must clear the
597 	 * flag and call this function.
598 	 */
599 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
600 		KASSERT(drop == 0,
601 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
602 		return;
603 	}
604 
605 	do {
606 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
607 		max_imm = max_imm_payload(tx_credits);
608 		max_nsegs = max_dsgl_nsegs(tx_credits);
609 
610 		SOCKBUF_LOCK(sb);
611 		sowwakeup = drop;
612 		if (drop) {
613 			sbdrop_locked(sb, drop);
614 			drop = 0;
615 		}
616 		sb_sndptr = sb->sb_sndptr;
617 		sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
618 		plen = 0;
619 		nsegs = 0;
620 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
621 		for (m = sndptr; m != NULL; m = m->m_next) {
622 			int n = sglist_count(mtod(m, void *), m->m_len);
623 
624 			nsegs += n;
625 			plen += m->m_len;
626 
627 			/* This mbuf sent us _over_ the nsegs limit, back out */
628 			if (plen > max_imm && nsegs > max_nsegs) {
629 				nsegs -= n;
630 				plen -= m->m_len;
631 				if (plen == 0) {
632 					/* Too few credits */
633 					toep->flags |= TPF_TX_SUSPENDED;
634 					if (sowwakeup)
635 						sowwakeup_locked(so);
636 					else
637 						SOCKBUF_UNLOCK(sb);
638 					SOCKBUF_UNLOCK_ASSERT(sb);
639 					return;
640 				}
641 				break;
642 			}
643 
644 			if (max_nsegs_1mbuf < n)
645 				max_nsegs_1mbuf = n;
646 			sb_sndptr = m;	/* new sb->sb_sndptr if all goes well */
647 
648 			/* This mbuf put us right at the max_nsegs limit */
649 			if (plen > max_imm && nsegs == max_nsegs) {
650 				m = m->m_next;
651 				break;
652 			}
653 		}
654 
655 		space = sbspace(sb);
656 
657 		if (space <= sb->sb_hiwat * 3 / 8 &&
658 		    toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
659 			compl = 1;
660 		else
661 			compl = 0;
662 
663 		if (sb->sb_flags & SB_AUTOSIZE &&
664 		    V_tcp_do_autosndbuf &&
665 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
666 		    space < sb->sb_hiwat / 8) {
667 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
668 			    V_tcp_autosndbuf_max);
669 
670 			if (!sbreserve_locked(sb, newsize, so, NULL))
671 				sb->sb_flags &= ~SB_AUTOSIZE;
672 			else
673 				sowwakeup = 1;	/* room available */
674 		}
675 		if (sowwakeup)
676 			sowwakeup_locked(so);
677 		else
678 			SOCKBUF_UNLOCK(sb);
679 		SOCKBUF_UNLOCK_ASSERT(sb);
680 
681 		/* nothing to send */
682 		if (plen == 0) {
683 			KASSERT(m == NULL,
684 			    ("%s: nothing to send, but m != NULL", __func__));
685 			break;
686 		}
687 
688 		if (__predict_false(toep->flags & TPF_FIN_SENT))
689 			panic("%s: excess tx.", __func__);
690 
691 		shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
692 		if (plen <= max_imm) {
693 
694 			/* Immediate data tx */
695 
696 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
697 					toep->ofld_txq);
698 			if (wr == NULL) {
699 				/* XXX: how will we recover from this? */
700 				toep->flags |= TPF_TX_SUSPENDED;
701 				return;
702 			}
703 			txwr = wrtod(wr);
704 			credits = howmany(wr->wr_len, 16);
705 			write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
706 			    sc->tt.tx_align);
707 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
708 			nsegs = 0;
709 		} else {
710 			int wr_len;
711 
712 			/* DSGL tx */
713 
714 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
715 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
716 			wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
717 			if (wr == NULL) {
718 				/* XXX: how will we recover from this? */
719 				toep->flags |= TPF_TX_SUSPENDED;
720 				return;
721 			}
722 			txwr = wrtod(wr);
723 			credits = howmany(wr_len, 16);
724 			write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
725 			    sc->tt.tx_align);
726 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
727 			    max_nsegs_1mbuf);
728 			if (wr_len & 0xf) {
729 				uint64_t *pad = (uint64_t *)
730 				    ((uintptr_t)txwr + wr_len);
731 				*pad = 0;
732 			}
733 		}
734 
735 		KASSERT(toep->tx_credits >= credits,
736 			("%s: not enough credits", __func__));
737 
738 		toep->tx_credits -= credits;
739 		toep->tx_nocompl += credits;
740 		toep->plen_nocompl += plen;
741 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
742 		    toep->tx_nocompl >= toep->tx_total / 4)
743 			compl = 1;
744 
745 		if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
746 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
747 			toep->tx_nocompl = 0;
748 			toep->plen_nocompl = 0;
749 		}
750 
751 		tp->snd_nxt += plen;
752 		tp->snd_max += plen;
753 
754 		SOCKBUF_LOCK(sb);
755 		KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
756 		sb->sb_sndptr = sb_sndptr;
757 		SOCKBUF_UNLOCK(sb);
758 
759 		toep->flags |= TPF_TX_DATA_SENT;
760 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
761 			toep->flags |= TPF_TX_SUSPENDED;
762 
763 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
764 		txsd->plen = plen;
765 		txsd->tx_credits = credits;
766 		txsd++;
767 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
768 			toep->txsd_pidx = 0;
769 			txsd = &toep->txsd[0];
770 		}
771 		toep->txsd_avail--;
772 
773 		t4_l2t_send(sc, wr, toep->l2te);
774 	} while (m != NULL);
775 
776 	/* Send a FIN if requested, but only if there's no more data to send */
777 	if (m == NULL && toep->flags & TPF_SEND_FIN)
778 		close_conn(sc, toep);
779 }
780 
781 static inline void
782 rqdrop_locked(struct mbufq *q, int plen)
783 {
784 	struct mbuf *m;
785 
786 	while (plen > 0) {
787 		m = mbufq_dequeue(q);
788 
789 		/* Too many credits. */
790 		MPASS(m != NULL);
791 		M_ASSERTPKTHDR(m);
792 
793 		/* Partial credits. */
794 		MPASS(plen >= m->m_pkthdr.len);
795 
796 		plen -= m->m_pkthdr.len;
797 		m_freem(m);
798 	}
799 }
800 
801 void
802 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
803 {
804 	struct mbuf *sndptr, *m;
805 	struct fw_ofld_tx_data_wr *txwr;
806 	struct wrqe *wr;
807 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
808 	u_int adjusted_plen, ulp_submode;
809 	struct inpcb *inp = toep->inp;
810 	struct tcpcb *tp = intotcpcb(inp);
811 	int tx_credits, shove;
812 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
813 	struct mbufq *pduq = &toep->ulp_pduq;
814 	static const u_int ulp_extra_len[] = {0, 4, 4, 8};
815 
816 	INP_WLOCK_ASSERT(inp);
817 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
818 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
819 	KASSERT(toep->ulp_mode == ULP_MODE_ISCSI,
820 	    ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
821 
822 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
823 		return;
824 
825 	/*
826 	 * This function doesn't resume by itself.  Someone else must clear the
827 	 * flag and call this function.
828 	 */
829 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
830 		KASSERT(drop == 0,
831 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
832 		return;
833 	}
834 
835 	if (drop)
836 		rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
837 
838 	while ((sndptr = mbufq_first(pduq)) != NULL) {
839 		M_ASSERTPKTHDR(sndptr);
840 
841 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
842 		max_imm = max_imm_payload(tx_credits);
843 		max_nsegs = max_dsgl_nsegs(tx_credits);
844 
845 		plen = 0;
846 		nsegs = 0;
847 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
848 		for (m = sndptr; m != NULL; m = m->m_next) {
849 			int n = sglist_count(mtod(m, void *), m->m_len);
850 
851 			nsegs += n;
852 			plen += m->m_len;
853 
854 			/*
855 			 * This mbuf would send us _over_ the nsegs limit.
856 			 * Suspend tx because the PDU can't be sent out.
857 			 */
858 			if (plen > max_imm && nsegs > max_nsegs) {
859 				toep->flags |= TPF_TX_SUSPENDED;
860 				return;
861 			}
862 
863 			if (max_nsegs_1mbuf < n)
864 				max_nsegs_1mbuf = n;
865 		}
866 
867 		if (__predict_false(toep->flags & TPF_FIN_SENT))
868 			panic("%s: excess tx.", __func__);
869 
870 		/*
871 		 * We have a PDU to send.  All of it goes out in one WR so 'm'
872 		 * is NULL.  A PDU's length is always a multiple of 4.
873 		 */
874 		MPASS(m == NULL);
875 		MPASS((plen & 3) == 0);
876 		MPASS(sndptr->m_pkthdr.len == plen);
877 
878 		shove = !(tp->t_flags & TF_MORETOCOME);
879 		ulp_submode = mbuf_ulp_submode(sndptr);
880 		MPASS(ulp_submode < nitems(ulp_extra_len));
881 
882 		/*
883 		 * plen doesn't include header and data digests, which are
884 		 * generated and inserted in the right places by the TOE, but
885 		 * they do occupy TCP sequence space and need to be accounted
886 		 * for.
887 		 */
888 		adjusted_plen = plen + ulp_extra_len[ulp_submode];
889 		if (plen <= max_imm) {
890 
891 			/* Immediate data tx */
892 
893 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
894 					toep->ofld_txq);
895 			if (wr == NULL) {
896 				/* XXX: how will we recover from this? */
897 				toep->flags |= TPF_TX_SUSPENDED;
898 				return;
899 			}
900 			txwr = wrtod(wr);
901 			credits = howmany(wr->wr_len, 16);
902 			write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
903 			    shove, ulp_submode, sc->tt.tx_align);
904 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
905 			nsegs = 0;
906 		} else {
907 			int wr_len;
908 
909 			/* DSGL tx */
910 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
911 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
912 			wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
913 			if (wr == NULL) {
914 				/* XXX: how will we recover from this? */
915 				toep->flags |= TPF_TX_SUSPENDED;
916 				return;
917 			}
918 			txwr = wrtod(wr);
919 			credits = howmany(wr_len, 16);
920 			write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
921 			    shove, ulp_submode, sc->tt.tx_align);
922 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
923 			    max_nsegs_1mbuf);
924 			if (wr_len & 0xf) {
925 				uint64_t *pad = (uint64_t *)
926 				    ((uintptr_t)txwr + wr_len);
927 				*pad = 0;
928 			}
929 		}
930 
931 		KASSERT(toep->tx_credits >= credits,
932 			("%s: not enough credits", __func__));
933 
934 		m = mbufq_dequeue(pduq);
935 		MPASS(m == sndptr);
936 		mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
937 
938 		toep->tx_credits -= credits;
939 		toep->tx_nocompl += credits;
940 		toep->plen_nocompl += plen;
941 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
942 		    toep->tx_nocompl >= toep->tx_total / 4) {
943 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
944 			toep->tx_nocompl = 0;
945 			toep->plen_nocompl = 0;
946 		}
947 
948 		tp->snd_nxt += adjusted_plen;
949 		tp->snd_max += adjusted_plen;
950 
951 		toep->flags |= TPF_TX_DATA_SENT;
952 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
953 			toep->flags |= TPF_TX_SUSPENDED;
954 
955 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
956 		txsd->plen = plen;
957 		txsd->tx_credits = credits;
958 		txsd++;
959 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
960 			toep->txsd_pidx = 0;
961 			txsd = &toep->txsd[0];
962 		}
963 		toep->txsd_avail--;
964 
965 		t4_l2t_send(sc, wr, toep->l2te);
966 	}
967 
968 	/* Send a FIN if requested, but only if there are no more PDUs to send */
969 	if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
970 		close_conn(sc, toep);
971 }
972 
973 int
974 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
975 {
976 	struct adapter *sc = tod->tod_softc;
977 #ifdef INVARIANTS
978 	struct inpcb *inp = tp->t_inpcb;
979 #endif
980 	struct toepcb *toep = tp->t_toe;
981 
982 	INP_WLOCK_ASSERT(inp);
983 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
984 	    ("%s: inp %p dropped.", __func__, inp));
985 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
986 
987 	if (toep->ulp_mode == ULP_MODE_ISCSI)
988 		t4_push_pdus(sc, toep, 0);
989 	else
990 		t4_push_frames(sc, toep, 0);
991 
992 	return (0);
993 }
994 
995 int
996 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
997 {
998 	struct adapter *sc = tod->tod_softc;
999 #ifdef INVARIANTS
1000 	struct inpcb *inp = tp->t_inpcb;
1001 #endif
1002 	struct toepcb *toep = tp->t_toe;
1003 
1004 	INP_WLOCK_ASSERT(inp);
1005 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1006 	    ("%s: inp %p dropped.", __func__, inp));
1007 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1008 
1009 	toep->flags |= TPF_SEND_FIN;
1010 	if (tp->t_state >= TCPS_ESTABLISHED) {
1011 		if (toep->ulp_mode == ULP_MODE_ISCSI)
1012 			t4_push_pdus(sc, toep, 0);
1013 		else
1014 			t4_push_frames(sc, toep, 0);
1015 	}
1016 
1017 	return (0);
1018 }
1019 
1020 int
1021 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1022 {
1023 	struct adapter *sc = tod->tod_softc;
1024 #if defined(INVARIANTS)
1025 	struct inpcb *inp = tp->t_inpcb;
1026 #endif
1027 	struct toepcb *toep = tp->t_toe;
1028 
1029 	INP_WLOCK_ASSERT(inp);
1030 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1031 	    ("%s: inp %p dropped.", __func__, inp));
1032 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1033 
1034 	/* hmmmm */
1035 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1036 	    ("%s: flowc for tid %u [%s] not sent already",
1037 	    __func__, toep->tid, tcpstates[tp->t_state]));
1038 
1039 	send_reset(sc, toep, 0);
1040 	return (0);
1041 }
1042 
1043 /*
1044  * Peer has sent us a FIN.
1045  */
1046 static int
1047 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1048 {
1049 	struct adapter *sc = iq->adapter;
1050 	const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1051 	unsigned int tid = GET_TID(cpl);
1052 	struct toepcb *toep = lookup_tid(sc, tid);
1053 	struct inpcb *inp = toep->inp;
1054 	struct tcpcb *tp = NULL;
1055 	struct socket *so;
1056 #ifdef INVARIANTS
1057 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1058 #endif
1059 
1060 	KASSERT(opcode == CPL_PEER_CLOSE,
1061 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1062 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1063 
1064 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1065 #ifdef INVARIANTS
1066 		struct synq_entry *synqe = (void *)toep;
1067 
1068 		INP_WLOCK(synqe->lctx->inp);
1069 		if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1070 			KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1071 			    ("%s: listen socket closed but tid %u not aborted.",
1072 			    __func__, tid));
1073 		} else {
1074 			/*
1075 			 * do_pass_accept_req is still running and will
1076 			 * eventually take care of this tid.
1077 			 */
1078 		}
1079 		INP_WUNLOCK(synqe->lctx->inp);
1080 #endif
1081 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1082 		    toep, toep->flags);
1083 		return (0);
1084 	}
1085 
1086 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1087 
1088 	INP_INFO_RLOCK(&V_tcbinfo);
1089 	INP_WLOCK(inp);
1090 	tp = intotcpcb(inp);
1091 
1092 	CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1093 	    tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1094 
1095 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1096 		goto done;
1097 
1098 	tp->rcv_nxt++;	/* FIN */
1099 
1100 	so = inp->inp_socket;
1101 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1102 		DDP_LOCK(toep);
1103 		if (__predict_false(toep->ddp_flags &
1104 		    (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)))
1105 			handle_ddp_close(toep, tp, cpl->rcv_nxt);
1106 		DDP_UNLOCK(toep);
1107 	}
1108 	socantrcvmore(so);
1109 
1110 	if (toep->ulp_mode != ULP_MODE_RDMA) {
1111 		KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1112 	    		("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1113 	    		be32toh(cpl->rcv_nxt)));
1114 	}
1115 
1116 	switch (tp->t_state) {
1117 	case TCPS_SYN_RECEIVED:
1118 		tp->t_starttime = ticks;
1119 		/* FALLTHROUGH */
1120 
1121 	case TCPS_ESTABLISHED:
1122 		tp->t_state = TCPS_CLOSE_WAIT;
1123 		break;
1124 
1125 	case TCPS_FIN_WAIT_1:
1126 		tp->t_state = TCPS_CLOSING;
1127 		break;
1128 
1129 	case TCPS_FIN_WAIT_2:
1130 		tcp_twstart(tp);
1131 		INP_UNLOCK_ASSERT(inp);	 /* safe, we have a ref on the inp */
1132 		INP_INFO_RUNLOCK(&V_tcbinfo);
1133 
1134 		INP_WLOCK(inp);
1135 		final_cpl_received(toep);
1136 		return (0);
1137 
1138 	default:
1139 		log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1140 		    __func__, tid, tp->t_state);
1141 	}
1142 done:
1143 	INP_WUNLOCK(inp);
1144 	INP_INFO_RUNLOCK(&V_tcbinfo);
1145 	return (0);
1146 }
1147 
1148 /*
1149  * Peer has ACK'd our FIN.
1150  */
1151 static int
1152 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1153     struct mbuf *m)
1154 {
1155 	struct adapter *sc = iq->adapter;
1156 	const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1157 	unsigned int tid = GET_TID(cpl);
1158 	struct toepcb *toep = lookup_tid(sc, tid);
1159 	struct inpcb *inp = toep->inp;
1160 	struct tcpcb *tp = NULL;
1161 	struct socket *so = NULL;
1162 #ifdef INVARIANTS
1163 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1164 #endif
1165 
1166 	KASSERT(opcode == CPL_CLOSE_CON_RPL,
1167 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1168 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1169 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1170 
1171 	INP_INFO_RLOCK(&V_tcbinfo);
1172 	INP_WLOCK(inp);
1173 	tp = intotcpcb(inp);
1174 
1175 	CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1176 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1177 
1178 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1179 		goto done;
1180 
1181 	so = inp->inp_socket;
1182 	tp->snd_una = be32toh(cpl->snd_nxt) - 1;	/* exclude FIN */
1183 
1184 	switch (tp->t_state) {
1185 	case TCPS_CLOSING:	/* see TCPS_FIN_WAIT_2 in do_peer_close too */
1186 		tcp_twstart(tp);
1187 release:
1188 		INP_UNLOCK_ASSERT(inp);	/* safe, we have a ref on the  inp */
1189 		INP_INFO_RUNLOCK(&V_tcbinfo);
1190 
1191 		INP_WLOCK(inp);
1192 		final_cpl_received(toep);	/* no more CPLs expected */
1193 
1194 		return (0);
1195 	case TCPS_LAST_ACK:
1196 		if (tcp_close(tp))
1197 			INP_WUNLOCK(inp);
1198 		goto release;
1199 
1200 	case TCPS_FIN_WAIT_1:
1201 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1202 			soisdisconnected(so);
1203 		tp->t_state = TCPS_FIN_WAIT_2;
1204 		break;
1205 
1206 	default:
1207 		log(LOG_ERR,
1208 		    "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1209 		    __func__, tid, tcpstates[tp->t_state]);
1210 	}
1211 done:
1212 	INP_WUNLOCK(inp);
1213 	INP_INFO_RUNLOCK(&V_tcbinfo);
1214 	return (0);
1215 }
1216 
1217 void
1218 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1219     int rst_status)
1220 {
1221 	struct wrqe *wr;
1222 	struct cpl_abort_rpl *cpl;
1223 
1224 	wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1225 	if (wr == NULL) {
1226 		/* XXX */
1227 		panic("%s: allocation failure.", __func__);
1228 	}
1229 	cpl = wrtod(wr);
1230 
1231 	INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1232 	cpl->cmd = rst_status;
1233 
1234 	t4_wrq_tx(sc, wr);
1235 }
1236 
1237 static int
1238 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1239 {
1240 	switch (abort_reason) {
1241 	case CPL_ERR_BAD_SYN:
1242 	case CPL_ERR_CONN_RESET:
1243 		return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1244 	case CPL_ERR_XMIT_TIMEDOUT:
1245 	case CPL_ERR_PERSIST_TIMEDOUT:
1246 	case CPL_ERR_FINWAIT2_TIMEDOUT:
1247 	case CPL_ERR_KEEPALIVE_TIMEDOUT:
1248 		return (ETIMEDOUT);
1249 	default:
1250 		return (EIO);
1251 	}
1252 }
1253 
1254 /*
1255  * TCP RST from the peer, timeout, or some other such critical error.
1256  */
1257 static int
1258 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1259 {
1260 	struct adapter *sc = iq->adapter;
1261 	const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1262 	unsigned int tid = GET_TID(cpl);
1263 	struct toepcb *toep = lookup_tid(sc, tid);
1264 	struct sge_wrq *ofld_txq = toep->ofld_txq;
1265 	struct inpcb *inp;
1266 	struct tcpcb *tp;
1267 #ifdef INVARIANTS
1268 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1269 #endif
1270 
1271 	KASSERT(opcode == CPL_ABORT_REQ_RSS,
1272 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1273 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1274 
1275 	if (toep->flags & TPF_SYNQE)
1276 		return (do_abort_req_synqe(iq, rss, m));
1277 
1278 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1279 
1280 	if (negative_advice(cpl->status)) {
1281 		CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1282 		    __func__, cpl->status, tid, toep->flags);
1283 		return (0);	/* Ignore negative advice */
1284 	}
1285 
1286 	inp = toep->inp;
1287 	INP_INFO_RLOCK(&V_tcbinfo);	/* for tcp_close */
1288 	INP_WLOCK(inp);
1289 
1290 	tp = intotcpcb(inp);
1291 
1292 	CTR6(KTR_CXGBE,
1293 	    "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1294 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1295 	    inp->inp_flags, cpl->status);
1296 
1297 	/*
1298 	 * If we'd initiated an abort earlier the reply to it is responsible for
1299 	 * cleaning up resources.  Otherwise we tear everything down right here
1300 	 * right now.  We owe the T4 a CPL_ABORT_RPL no matter what.
1301 	 */
1302 	if (toep->flags & TPF_ABORT_SHUTDOWN) {
1303 		INP_WUNLOCK(inp);
1304 		goto done;
1305 	}
1306 	toep->flags |= TPF_ABORT_SHUTDOWN;
1307 
1308 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1309 		struct socket *so = inp->inp_socket;
1310 
1311 		if (so != NULL)
1312 			so_error_set(so, abort_status_to_errno(tp,
1313 			    cpl->status));
1314 		tp = tcp_close(tp);
1315 		if (tp == NULL)
1316 			INP_WLOCK(inp);	/* re-acquire */
1317 	}
1318 
1319 	final_cpl_received(toep);
1320 done:
1321 	INP_INFO_RUNLOCK(&V_tcbinfo);
1322 	send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1323 	return (0);
1324 }
1325 
1326 /*
1327  * Reply to the CPL_ABORT_REQ (send_reset)
1328  */
1329 static int
1330 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1331 {
1332 	struct adapter *sc = iq->adapter;
1333 	const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1334 	unsigned int tid = GET_TID(cpl);
1335 	struct toepcb *toep = lookup_tid(sc, tid);
1336 	struct inpcb *inp = toep->inp;
1337 #ifdef INVARIANTS
1338 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1339 #endif
1340 
1341 	KASSERT(opcode == CPL_ABORT_RPL_RSS,
1342 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1343 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1344 
1345 	if (toep->flags & TPF_SYNQE)
1346 		return (do_abort_rpl_synqe(iq, rss, m));
1347 
1348 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1349 
1350 	CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1351 	    __func__, tid, toep, inp, cpl->status);
1352 
1353 	KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1354 	    ("%s: wasn't expecting abort reply", __func__));
1355 
1356 	INP_WLOCK(inp);
1357 	final_cpl_received(toep);
1358 
1359 	return (0);
1360 }
1361 
1362 static int
1363 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1364 {
1365 	struct adapter *sc = iq->adapter;
1366 	const struct cpl_rx_data *cpl = mtod(m, const void *);
1367 	unsigned int tid = GET_TID(cpl);
1368 	struct toepcb *toep = lookup_tid(sc, tid);
1369 	struct inpcb *inp = toep->inp;
1370 	struct tcpcb *tp;
1371 	struct socket *so;
1372 	struct sockbuf *sb;
1373 	int len;
1374 	uint32_t ddp_placed = 0;
1375 
1376 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1377 #ifdef INVARIANTS
1378 		struct synq_entry *synqe = (void *)toep;
1379 
1380 		INP_WLOCK(synqe->lctx->inp);
1381 		if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1382 			KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1383 			    ("%s: listen socket closed but tid %u not aborted.",
1384 			    __func__, tid));
1385 		} else {
1386 			/*
1387 			 * do_pass_accept_req is still running and will
1388 			 * eventually take care of this tid.
1389 			 */
1390 		}
1391 		INP_WUNLOCK(synqe->lctx->inp);
1392 #endif
1393 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1394 		    toep, toep->flags);
1395 		m_freem(m);
1396 		return (0);
1397 	}
1398 
1399 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1400 
1401 	/* strip off CPL header */
1402 	m_adj(m, sizeof(*cpl));
1403 	len = m->m_pkthdr.len;
1404 
1405 	INP_WLOCK(inp);
1406 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1407 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1408 		    __func__, tid, len, inp->inp_flags);
1409 		INP_WUNLOCK(inp);
1410 		m_freem(m);
1411 		return (0);
1412 	}
1413 
1414 	tp = intotcpcb(inp);
1415 
1416 	if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1417 		ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1418 
1419 	tp->rcv_nxt += len;
1420 	KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
1421 	tp->rcv_wnd -= len;
1422 	tp->t_rcvtime = ticks;
1423 
1424 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
1425 		DDP_LOCK(toep);
1426 	so = inp_inpcbtosocket(inp);
1427 	sb = &so->so_rcv;
1428 	SOCKBUF_LOCK(sb);
1429 
1430 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1431 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1432 		    __func__, tid, len);
1433 		m_freem(m);
1434 		SOCKBUF_UNLOCK(sb);
1435 		if (toep->ulp_mode == ULP_MODE_TCPDDP)
1436 			DDP_UNLOCK(toep);
1437 		INP_WUNLOCK(inp);
1438 
1439 		INP_INFO_RLOCK(&V_tcbinfo);
1440 		INP_WLOCK(inp);
1441 		tp = tcp_drop(tp, ECONNRESET);
1442 		if (tp)
1443 			INP_WUNLOCK(inp);
1444 		INP_INFO_RUNLOCK(&V_tcbinfo);
1445 
1446 		return (0);
1447 	}
1448 
1449 	/* receive buffer autosize */
1450 	CURVNET_SET(so->so_vnet);
1451 	if (sb->sb_flags & SB_AUTOSIZE &&
1452 	    V_tcp_do_autorcvbuf &&
1453 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1454 	    len > (sbspace(sb) / 8 * 7)) {
1455 		unsigned int hiwat = sb->sb_hiwat;
1456 		unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1457 		    V_tcp_autorcvbuf_max);
1458 
1459 		if (!sbreserve_locked(sb, newsize, so, NULL))
1460 			sb->sb_flags &= ~SB_AUTOSIZE;
1461 		else
1462 			toep->rx_credits += newsize - hiwat;
1463 	}
1464 
1465 	if (toep->ddp_waiting_count != 0 || toep->ddp_active_count != 0)
1466 		CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)", __func__,
1467 		    tid, len);
1468 
1469 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1470 		int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1471 
1472 		if (changed) {
1473 			if (toep->ddp_flags & DDP_SC_REQ)
1474 				toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1475 			else {
1476 				KASSERT(cpl->ddp_off == 1,
1477 				    ("%s: DDP switched on by itself.",
1478 				    __func__));
1479 
1480 				/* Fell out of DDP mode */
1481 				toep->ddp_flags &= ~DDP_ON;
1482 				CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
1483 				    __func__);
1484 
1485 				insert_ddp_data(toep, ddp_placed);
1486 			}
1487 		}
1488 
1489 		if (toep->ddp_flags & DDP_ON) {
1490 			/*
1491 			 * CPL_RX_DATA with DDP on can only be an indicate.
1492 			 * Start posting queued AIO requests via DDP.  The
1493 			 * payload that arrived in this indicate is appended
1494 			 * to the socket buffer as usual.
1495 			 */
1496 			handle_ddp_indicate(toep);
1497 		}
1498 	}
1499 
1500 	KASSERT(toep->sb_cc >= sbused(sb),
1501 	    ("%s: sb %p has more data (%d) than last time (%d).",
1502 	    __func__, sb, sbused(sb), toep->sb_cc));
1503 	toep->rx_credits += toep->sb_cc - sbused(sb);
1504 	sbappendstream_locked(sb, m, 0);
1505 	toep->sb_cc = sbused(sb);
1506 	if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1507 		int credits;
1508 
1509 		credits = send_rx_credits(sc, toep, toep->rx_credits);
1510 		toep->rx_credits -= credits;
1511 		tp->rcv_wnd += credits;
1512 		tp->rcv_adv += credits;
1513 	}
1514 
1515 	if (toep->ddp_waiting_count > 0 && sbavail(sb) != 0) {
1516 		CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__,
1517 		    tid);
1518 		ddp_queue_toep(toep);
1519 	}
1520 	sorwakeup_locked(so);
1521 	SOCKBUF_UNLOCK_ASSERT(sb);
1522 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
1523 		DDP_UNLOCK(toep);
1524 
1525 	INP_WUNLOCK(inp);
1526 	CURVNET_RESTORE();
1527 	return (0);
1528 }
1529 
1530 #define S_CPL_FW4_ACK_OPCODE    24
1531 #define M_CPL_FW4_ACK_OPCODE    0xff
1532 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1533 #define G_CPL_FW4_ACK_OPCODE(x) \
1534     (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1535 
1536 #define S_CPL_FW4_ACK_FLOWID    0
1537 #define M_CPL_FW4_ACK_FLOWID    0xffffff
1538 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1539 #define G_CPL_FW4_ACK_FLOWID(x) \
1540     (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1541 
1542 #define S_CPL_FW4_ACK_CR        24
1543 #define M_CPL_FW4_ACK_CR        0xff
1544 #define V_CPL_FW4_ACK_CR(x)     ((x) << S_CPL_FW4_ACK_CR)
1545 #define G_CPL_FW4_ACK_CR(x)     (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1546 
1547 #define S_CPL_FW4_ACK_SEQVAL    0
1548 #define M_CPL_FW4_ACK_SEQVAL    0x1
1549 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1550 #define G_CPL_FW4_ACK_SEQVAL(x) \
1551     (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1552 #define F_CPL_FW4_ACK_SEQVAL    V_CPL_FW4_ACK_SEQVAL(1U)
1553 
1554 static int
1555 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1556 {
1557 	struct adapter *sc = iq->adapter;
1558 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1559 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1560 	struct toepcb *toep = lookup_tid(sc, tid);
1561 	struct inpcb *inp;
1562 	struct tcpcb *tp;
1563 	struct socket *so;
1564 	uint8_t credits = cpl->credits;
1565 	struct ofld_tx_sdesc *txsd;
1566 	int plen;
1567 #ifdef INVARIANTS
1568 	unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1569 #endif
1570 
1571 	/*
1572 	 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1573 	 * now this comes back carrying the credits for the flowc.
1574 	 */
1575 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1576 		KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1577 		    ("%s: credits for a synq entry %p", __func__, toep));
1578 		return (0);
1579 	}
1580 
1581 	inp = toep->inp;
1582 
1583 	KASSERT(opcode == CPL_FW4_ACK,
1584 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1585 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1586 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1587 
1588 	INP_WLOCK(inp);
1589 
1590 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1591 		INP_WUNLOCK(inp);
1592 		return (0);
1593 	}
1594 
1595 	KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1596 	    ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1597 
1598 	tp = intotcpcb(inp);
1599 
1600 	if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1601 		tcp_seq snd_una = be32toh(cpl->snd_una);
1602 
1603 #ifdef INVARIANTS
1604 		if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1605 			log(LOG_ERR,
1606 			    "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1607 			    __func__, snd_una, toep->tid, tp->snd_una);
1608 		}
1609 #endif
1610 
1611 		if (tp->snd_una != snd_una) {
1612 			tp->snd_una = snd_una;
1613 			tp->ts_recent_age = tcp_ts_getticks();
1614 		}
1615 	}
1616 
1617 	so = inp->inp_socket;
1618 	txsd = &toep->txsd[toep->txsd_cidx];
1619 	plen = 0;
1620 	while (credits) {
1621 		KASSERT(credits >= txsd->tx_credits,
1622 		    ("%s: too many (or partial) credits", __func__));
1623 		credits -= txsd->tx_credits;
1624 		toep->tx_credits += txsd->tx_credits;
1625 		plen += txsd->plen;
1626 		txsd++;
1627 		toep->txsd_avail++;
1628 		KASSERT(toep->txsd_avail <= toep->txsd_total,
1629 		    ("%s: txsd avail > total", __func__));
1630 		if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1631 			txsd = &toep->txsd[0];
1632 			toep->txsd_cidx = 0;
1633 		}
1634 	}
1635 
1636 	if (toep->tx_credits == toep->tx_total) {
1637 		toep->tx_nocompl = 0;
1638 		toep->plen_nocompl = 0;
1639 	}
1640 
1641 	if (toep->flags & TPF_TX_SUSPENDED &&
1642 	    toep->tx_credits >= toep->tx_total / 4) {
1643 		toep->flags &= ~TPF_TX_SUSPENDED;
1644 		if (toep->ulp_mode == ULP_MODE_ISCSI)
1645 			t4_push_pdus(sc, toep, plen);
1646 		else
1647 			t4_push_frames(sc, toep, plen);
1648 	} else if (plen > 0) {
1649 		struct sockbuf *sb = &so->so_snd;
1650 		int sbu;
1651 
1652 		SOCKBUF_LOCK(sb);
1653 		sbu = sbused(sb);
1654 		if (toep->ulp_mode == ULP_MODE_ISCSI) {
1655 
1656 			if (__predict_false(sbu > 0)) {
1657 				/*
1658 				 * The data trasmitted before the tid's ULP mode
1659 				 * changed to ISCSI is still in so_snd.
1660 				 * Incoming credits should account for so_snd
1661 				 * first.
1662 				 */
1663 				sbdrop_locked(sb, min(sbu, plen));
1664 				plen -= min(sbu, plen);
1665 			}
1666 			sowwakeup_locked(so);	/* unlocks so_snd */
1667 			rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1668 		} else {
1669 			sbdrop_locked(sb, plen);
1670 			sowwakeup_locked(so);	/* unlocks so_snd */
1671 		}
1672 		SOCKBUF_UNLOCK_ASSERT(sb);
1673 	}
1674 
1675 	INP_WUNLOCK(inp);
1676 
1677 	return (0);
1678 }
1679 
1680 static int
1681 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1682 {
1683 	struct adapter *sc = iq->adapter;
1684 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1685 	unsigned int tid = GET_TID(cpl);
1686 	struct toepcb *toep;
1687 #ifdef INVARIANTS
1688 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1689 #endif
1690 
1691 	KASSERT(opcode == CPL_SET_TCB_RPL,
1692 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1693 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1694 
1695 	if (is_ftid(sc, tid))
1696 		return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1697 
1698 	toep = lookup_tid(sc, tid);
1699 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1700 		handle_ddp_tcb_rpl(toep, cpl);
1701 		return (0);
1702 	}
1703 
1704 	/*
1705 	 * TOM and/or other ULPs don't request replies for CPL_SET_TCB or
1706 	 * CPL_SET_TCB_FIELD requests.  This can easily change and when it does
1707 	 * the dispatch code will go here.
1708 	 */
1709 #ifdef INVARIANTS
1710 	panic("%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p", __func__,
1711 	    tid, iq);
1712 #else
1713 	log(LOG_ERR, "%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p\n",
1714 	    __func__, tid, iq);
1715 #endif
1716 
1717 	return (0);
1718 }
1719 
1720 void
1721 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1722     uint16_t word, uint64_t mask, uint64_t val)
1723 {
1724 	struct wrqe *wr;
1725 	struct cpl_set_tcb_field *req;
1726 
1727 	wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1728 	if (wr == NULL) {
1729 		/* XXX */
1730 		panic("%s: allocation failure.", __func__);
1731 	}
1732 	req = wrtod(wr);
1733 
1734 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1735 	req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1736 	    V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1737 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1738 	req->mask = htobe64(mask);
1739 	req->val = htobe64(val);
1740 
1741 	t4_wrq_tx(sc, wr);
1742 }
1743 
1744 void
1745 t4_set_tcb_field_rpl(struct adapter *sc, struct toepcb *toep, int ctrl,
1746     uint16_t word, uint64_t mask, uint64_t val, uint8_t cookie)
1747 {
1748 	struct wrqe *wr;
1749 	struct cpl_set_tcb_field *req;
1750 
1751 	KASSERT((cookie & ~M_COOKIE) == 0, ("%s: invalid cookie %#x", __func__,
1752 	    cookie));
1753 	wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1754 	if (wr == NULL) {
1755 		/* XXX */
1756 		panic("%s: allocation failure.", __func__);
1757 	}
1758 	req = wrtod(wr);
1759 
1760 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1761 	req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1762 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie));
1763 	req->mask = htobe64(mask);
1764 	req->val = htobe64(val);
1765 
1766 	t4_wrq_tx(sc, wr);
1767 }
1768 
1769 void
1770 t4_init_cpl_io_handlers(struct adapter *sc)
1771 {
1772 
1773 	t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1774 	t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1775 	t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1776 	t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1777 	t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1778 	t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1779 	t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1780 }
1781 
1782 void
1783 t4_uninit_cpl_io_handlers(struct adapter *sc)
1784 {
1785 
1786 	t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);
1787 }
1788 #endif
1789