xref: /freebsd/sys/dev/cxgbe/tom/t4_cpl_io.c (revision 3fc9e2c36555140de248a0b4def91bbfa44d7c2c)
1 /*-
2  * Copyright (c) 2012 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/tcp_var.h>
48 #define TCPSTATES
49 #include <netinet/tcp_fsm.h>
50 #include <netinet/tcp_seq.h>
51 #include <netinet/toecore.h>
52 
53 #include "common/common.h"
54 #include "common/t4_msg.h"
55 #include "common/t4_regs.h"
56 #include "common/t4_tcb.h"
57 #include "tom/t4_tom_l2t.h"
58 #include "tom/t4_tom.h"
59 
60 VNET_DECLARE(int, tcp_do_autosndbuf);
61 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
62 VNET_DECLARE(int, tcp_autosndbuf_inc);
63 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
64 VNET_DECLARE(int, tcp_autosndbuf_max);
65 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
66 VNET_DECLARE(int, tcp_do_autorcvbuf);
67 #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf)
68 VNET_DECLARE(int, tcp_autorcvbuf_inc);
69 #define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc)
70 VNET_DECLARE(int, tcp_autorcvbuf_max);
71 #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max)
72 
73 void
74 send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp)
75 {
76 	struct wrqe *wr;
77 	struct fw_flowc_wr *flowc;
78 	unsigned int nparams = ftxp ? 8 : 6, flowclen;
79 	struct port_info *pi = toep->port;
80 	struct adapter *sc = pi->adapter;
81 	unsigned int pfvf = G_FW_VIID_PFN(pi->viid) << S_FW_VIID_PFN;
82 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
83 
84 	KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
85 	    ("%s: flowc for tid %u sent already", __func__, toep->tid));
86 
87 	CTR2(KTR_CXGBE, "%s: tid %u", __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 	} else {
124 		flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF;
125 		flowc->mnemval[4].val = htobe32(512);
126 		flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS;
127 		flowc->mnemval[5].val = htobe32(512);
128 	}
129 
130 	txsd->tx_credits = howmany(flowclen, 16);
131 	txsd->plen = 0;
132 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
133 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
134 	toep->tx_credits -= txsd->tx_credits;
135 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
136 		toep->txsd_pidx = 0;
137 	toep->txsd_avail--;
138 
139 	toep->flags |= TPF_FLOWC_WR_SENT;
140         t4_wrq_tx(sc, wr);
141 }
142 
143 void
144 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
145 {
146 	struct wrqe *wr;
147 	struct cpl_abort_req *req;
148 	int tid = toep->tid;
149 	struct inpcb *inp = toep->inp;
150 	struct tcpcb *tp = intotcpcb(inp);	/* don't use if INP_DROPPED */
151 
152 	INP_WLOCK_ASSERT(inp);
153 
154 	CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
155 	    __func__, toep->tid,
156 	    inp->inp_flags & INP_DROPPED ? "inp dropped" :
157 	    tcpstates[tp->t_state],
158 	    toep->flags, inp->inp_flags,
159 	    toep->flags & TPF_ABORT_SHUTDOWN ?
160 	    " (abort already in progress)" : "");
161 
162 	if (toep->flags & TPF_ABORT_SHUTDOWN)
163 		return;	/* abort already in progress */
164 
165 	toep->flags |= TPF_ABORT_SHUTDOWN;
166 
167 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
168 	    ("%s: flowc_wr not sent for tid %d.", __func__, tid));
169 
170 	wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
171 	if (wr == NULL) {
172 		/* XXX */
173 		panic("%s: allocation failure.", __func__);
174 	}
175 	req = wrtod(wr);
176 
177 	INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
178 	if (inp->inp_flags & INP_DROPPED)
179 		req->rsvd0 = htobe32(snd_nxt);
180 	else
181 		req->rsvd0 = htobe32(tp->snd_nxt);
182 	req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
183 	req->cmd = CPL_ABORT_SEND_RST;
184 
185 	/*
186 	 * XXX: What's the correct way to tell that the inp hasn't been detached
187 	 * from its socket?  Should I even be flushing the snd buffer here?
188 	 */
189 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
190 		struct socket *so = inp->inp_socket;
191 
192 		if (so != NULL)	/* because I'm not sure.  See comment above */
193 			sbflush(&so->so_snd);
194 	}
195 
196 	t4_l2t_send(sc, wr, toep->l2te);
197 }
198 
199 /*
200  * Called when a connection is established to translate the TCP options
201  * reported by HW to FreeBSD's native format.
202  */
203 static void
204 assign_rxopt(struct tcpcb *tp, unsigned int opt)
205 {
206 	struct toepcb *toep = tp->t_toe;
207 	struct adapter *sc = td_adapter(toep->td);
208 
209 	INP_LOCK_ASSERT(tp->t_inpcb);
210 
211 	tp->t_maxseg = tp->t_maxopd = sc->params.mtus[G_TCPOPT_MSS(opt)] - 40;
212 
213 	if (G_TCPOPT_TSTAMP(opt)) {
214 		tp->t_flags |= TF_RCVD_TSTMP;	/* timestamps ok */
215 		tp->ts_recent = 0;		/* hmmm */
216 		tp->ts_recent_age = tcp_ts_getticks();
217 		tp->t_maxseg -= TCPOLEN_TSTAMP_APPA;
218 	}
219 
220 	if (G_TCPOPT_SACK(opt))
221 		tp->t_flags |= TF_SACK_PERMIT;	/* should already be set */
222 	else
223 		tp->t_flags &= ~TF_SACK_PERMIT;	/* sack disallowed by peer */
224 
225 	if (G_TCPOPT_WSCALE_OK(opt))
226 		tp->t_flags |= TF_RCVD_SCALE;
227 
228 	/* Doing window scaling? */
229 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
230 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
231 		tp->rcv_scale = tp->request_r_scale;
232 		tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
233 	}
234 }
235 
236 /*
237  * Completes some final bits of initialization for just established connections
238  * and changes their state to TCPS_ESTABLISHED.
239  *
240  * The ISNs are from after the exchange of SYNs.  i.e., the true ISN + 1.
241  */
242 void
243 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
244     uint16_t opt)
245 {
246 	struct inpcb *inp = toep->inp;
247 	struct socket *so = inp->inp_socket;
248 	struct tcpcb *tp = intotcpcb(inp);
249 	long bufsize;
250 	uint32_t iss = be32toh(snd_isn) - 1;	/* true ISS */
251 	uint32_t irs = be32toh(rcv_isn) - 1;	/* true IRS */
252 	uint16_t tcpopt = be16toh(opt);
253 	struct flowc_tx_params ftxp;
254 
255 	INP_WLOCK_ASSERT(inp);
256 	KASSERT(tp->t_state == TCPS_SYN_SENT ||
257 	    tp->t_state == TCPS_SYN_RECEIVED,
258 	    ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
259 
260 	CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p",
261 	    __func__, toep->tid, toep, inp);
262 
263 	tp->t_state = TCPS_ESTABLISHED;
264 	tp->t_starttime = ticks;
265 	TCPSTAT_INC(tcps_connects);
266 
267 	tp->irs = irs;
268 	tcp_rcvseqinit(tp);
269 	tp->rcv_wnd = toep->rx_credits << 10;
270 	tp->rcv_adv += tp->rcv_wnd;
271 	tp->last_ack_sent = tp->rcv_nxt;
272 
273 	/*
274 	 * If we were unable to send all rx credits via opt0, save the remainder
275 	 * in rx_credits so that they can be handed over with the next credit
276 	 * update.
277 	 */
278 	SOCKBUF_LOCK(&so->so_rcv);
279 	bufsize = select_rcv_wnd(so);
280 	SOCKBUF_UNLOCK(&so->so_rcv);
281 	toep->rx_credits = bufsize - tp->rcv_wnd;
282 
283 	tp->iss = iss;
284 	tcp_sendseqinit(tp);
285 	tp->snd_una = iss + 1;
286 	tp->snd_nxt = iss + 1;
287 	tp->snd_max = iss + 1;
288 
289 	assign_rxopt(tp, tcpopt);
290 
291 	SOCKBUF_LOCK(&so->so_snd);
292 	if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf)
293 		bufsize = V_tcp_autosndbuf_max;
294 	else
295 		bufsize = sbspace(&so->so_snd);
296 	SOCKBUF_UNLOCK(&so->so_snd);
297 
298 	ftxp.snd_nxt = tp->snd_nxt;
299 	ftxp.rcv_nxt = tp->rcv_nxt;
300 	ftxp.snd_space = bufsize;
301 	ftxp.mss = tp->t_maxseg;
302 	send_flowc_wr(toep, &ftxp);
303 
304 	soisconnected(so);
305 }
306 
307 static int
308 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
309 {
310 	struct wrqe *wr;
311 	struct cpl_rx_data_ack *req;
312 	uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
313 
314 	KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
315 
316 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
317 	if (wr == NULL)
318 		return (0);
319 	req = wrtod(wr);
320 
321 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
322 	req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
323 
324 	t4_wrq_tx(sc, wr);
325 	return (credits);
326 }
327 
328 void
329 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
330 {
331 	struct adapter *sc = tod->tod_softc;
332 	struct inpcb *inp = tp->t_inpcb;
333 	struct socket *so = inp->inp_socket;
334 	struct sockbuf *sb = &so->so_rcv;
335 	struct toepcb *toep = tp->t_toe;
336 	int credits;
337 
338 	INP_WLOCK_ASSERT(inp);
339 
340 	SOCKBUF_LOCK(sb);
341 	KASSERT(toep->sb_cc >= sb->sb_cc,
342 	    ("%s: sb %p has more data (%d) than last time (%d).",
343 	    __func__, sb, sb->sb_cc, toep->sb_cc));
344 	toep->rx_credits += toep->sb_cc - sb->sb_cc;
345 	toep->sb_cc = sb->sb_cc;
346 	credits = toep->rx_credits;
347 	SOCKBUF_UNLOCK(sb);
348 
349 	if (credits > 0 &&
350 	    (credits + 16384 >= tp->rcv_wnd || credits >= 15 * 1024)) {
351 
352 		credits = send_rx_credits(sc, toep, credits);
353 		SOCKBUF_LOCK(sb);
354 		toep->rx_credits -= credits;
355 		SOCKBUF_UNLOCK(sb);
356 		tp->rcv_wnd += credits;
357 		tp->rcv_adv += credits;
358 	}
359 }
360 
361 /*
362  * Close a connection by sending a CPL_CLOSE_CON_REQ message.
363  */
364 static int
365 close_conn(struct adapter *sc, struct toepcb *toep)
366 {
367 	struct wrqe *wr;
368 	struct cpl_close_con_req *req;
369 	unsigned int tid = toep->tid;
370 
371 	CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
372 	    toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
373 
374 	if (toep->flags & TPF_FIN_SENT)
375 		return (0);
376 
377 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
378 	    ("%s: flowc_wr not sent for tid %u.", __func__, tid));
379 
380 	wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
381 	if (wr == NULL) {
382 		/* XXX */
383 		panic("%s: allocation failure.", __func__);
384 	}
385 	req = wrtod(wr);
386 
387         req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
388 	    V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
389 	req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
390 	    V_FW_WR_FLOWID(tid));
391         req->wr.wr_lo = cpu_to_be64(0);
392         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
393 	req->rsvd = 0;
394 
395 	toep->flags |= TPF_FIN_SENT;
396 	toep->flags &= ~TPF_SEND_FIN;
397 	t4_l2t_send(sc, wr, toep->l2te);
398 
399 	return (0);
400 }
401 
402 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
403 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
404 
405 /* Maximum amount of immediate data we could stuff in a WR */
406 static inline int
407 max_imm_payload(int tx_credits)
408 {
409 	const int n = 2;	/* Use only up to 2 desc for imm. data WR */
410 
411 	KASSERT(tx_credits >= 0 &&
412 		tx_credits <= MAX_OFLD_TX_CREDITS,
413 		("%s: %d credits", __func__, tx_credits));
414 
415 	if (tx_credits < MIN_OFLD_TX_CREDITS)
416 		return (0);
417 
418 	if (tx_credits >= (n * EQ_ESIZE) / 16)
419 		return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
420 	else
421 		return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
422 }
423 
424 /* Maximum number of SGL entries we could stuff in a WR */
425 static inline int
426 max_dsgl_nsegs(int tx_credits)
427 {
428 	int nseg = 1;	/* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
429 	int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
430 
431 	KASSERT(tx_credits >= 0 &&
432 		tx_credits <= MAX_OFLD_TX_CREDITS,
433 		("%s: %d credits", __func__, tx_credits));
434 
435 	if (tx_credits < MIN_OFLD_TX_CREDITS)
436 		return (0);
437 
438 	nseg += 2 * (sge_pair_credits * 16 / 24);
439 	if ((sge_pair_credits * 16) % 24 == 16)
440 		nseg++;
441 
442 	return (nseg);
443 }
444 
445 static inline void
446 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
447     unsigned int plen, uint8_t credits, int more_to_come)
448 {
449 	struct fw_ofld_tx_data_wr *txwr = dst;
450 	int shove = !more_to_come;
451 	int compl = 1;
452 
453 	/*
454 	 * We always request completion notifications from the firmware.  The
455 	 * only exception is when we know we'll get more data to send shortly
456 	 * and that we'll have some tx credits remaining to transmit that data.
457 	 */
458 	if (more_to_come && toep->tx_credits - credits >= MIN_OFLD_TX_CREDITS)
459 		compl = 0;
460 
461 	txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
462 	    V_FW_WR_COMPL(compl) | V_FW_WR_IMMDLEN(immdlen));
463 	txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
464 	    V_FW_WR_LEN16(credits));
465 	txwr->tunnel_to_proxy =
466 	    htobe32(V_FW_OFLD_TX_DATA_WR_ULPMODE(toep->ulp_mode) |
467 		V_FW_OFLD_TX_DATA_WR_URGENT(0) |	/* XXX */
468 		V_FW_OFLD_TX_DATA_WR_SHOVE(shove));
469 	txwr->plen = htobe32(plen);
470 }
471 
472 /*
473  * Generate a DSGL from a starting mbuf.  The total number of segments and the
474  * maximum segments in any one mbuf are provided.
475  */
476 static void
477 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
478 {
479 	struct mbuf *m;
480 	struct ulptx_sgl *usgl = dst;
481 	int i, j, rc;
482 	struct sglist sg;
483 	struct sglist_seg segs[n];
484 
485 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
486 
487 	sglist_init(&sg, n, segs);
488 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
489 	    V_ULPTX_NSGE(nsegs));
490 
491 	i = -1;
492 	for (m = start; m != stop; m = m->m_next) {
493 		rc = sglist_append(&sg, mtod(m, void *), m->m_len);
494 		if (__predict_false(rc != 0))
495 			panic("%s: sglist_append %d", __func__, rc);
496 
497 		for (j = 0; j < sg.sg_nseg; i++, j++) {
498 			if (i < 0) {
499 				usgl->len0 = htobe32(segs[j].ss_len);
500 				usgl->addr0 = htobe64(segs[j].ss_paddr);
501 			} else {
502 				usgl->sge[i / 2].len[i & 1] =
503 				    htobe32(segs[j].ss_len);
504 				usgl->sge[i / 2].addr[i & 1] =
505 				    htobe64(segs[j].ss_paddr);
506 			}
507 #ifdef INVARIANTS
508 			nsegs--;
509 #endif
510 		}
511 		sglist_reset(&sg);
512 	}
513 	if (i & 1)
514 		usgl->sge[i / 2].len[1] = htobe32(0);
515 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
516 	    __func__, nsegs, start, stop));
517 }
518 
519 /*
520  * Max number of SGL entries an offload tx work request can have.  This is 41
521  * (1 + 40) for a full 512B work request.
522  * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
523  */
524 #define OFLD_SGL_LEN (41)
525 
526 /*
527  * Send data and/or a FIN to the peer.
528  *
529  * The socket's so_snd buffer consists of a stream of data starting with sb_mb
530  * and linked together with m_next.  sb_sndptr, if set, is the last mbuf that
531  * was transmitted.
532  */
533 static void
534 t4_push_frames(struct adapter *sc, struct toepcb *toep)
535 {
536 	struct mbuf *sndptr, *m, *sb_sndptr;
537 	struct fw_ofld_tx_data_wr *txwr;
538 	struct wrqe *wr;
539 	unsigned int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
540 	struct inpcb *inp = toep->inp;
541 	struct tcpcb *tp = intotcpcb(inp);
542 	struct socket *so = inp->inp_socket;
543 	struct sockbuf *sb = &so->so_snd;
544 	int tx_credits;
545 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
546 
547 	INP_WLOCK_ASSERT(inp);
548 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
549 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
550 
551 	KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
552 	    toep->ulp_mode == ULP_MODE_TCPDDP ||
553 	    toep->ulp_mode == ULP_MODE_RDMA,
554 	    ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
555 
556 	/*
557 	 * This function doesn't resume by itself.  Someone else must clear the
558 	 * flag and call this function.
559 	 */
560 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED))
561 		return;
562 
563 	do {
564 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
565 		max_imm = max_imm_payload(tx_credits);
566 		max_nsegs = max_dsgl_nsegs(tx_credits);
567 
568 		SOCKBUF_LOCK(sb);
569 		sb_sndptr = sb->sb_sndptr;
570 		sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
571 		plen = 0;
572 		nsegs = 0;
573 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
574 		for (m = sndptr; m != NULL; m = m->m_next) {
575 			int n = sglist_count(mtod(m, void *), m->m_len);
576 
577 			nsegs += n;
578 			plen += m->m_len;
579 
580 			/* This mbuf sent us _over_ the nsegs limit, back out */
581 			if (plen > max_imm && nsegs > max_nsegs) {
582 				nsegs -= n;
583 				plen -= m->m_len;
584 				if (plen == 0) {
585 					/* Too few credits */
586 					toep->flags |= TPF_TX_SUSPENDED;
587 					SOCKBUF_UNLOCK(sb);
588 					return;
589 				}
590 				break;
591 			}
592 
593 			if (max_nsegs_1mbuf < n)
594 				max_nsegs_1mbuf = n;
595 			sb_sndptr = m;	/* new sb->sb_sndptr if all goes well */
596 
597 			/* This mbuf put us right at the max_nsegs limit */
598 			if (plen > max_imm && nsegs == max_nsegs) {
599 				m = m->m_next;
600 				break;
601 			}
602 		}
603 
604 		if (sb->sb_flags & SB_AUTOSIZE &&
605 		    V_tcp_do_autosndbuf &&
606 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
607 		    sbspace(sb) < sb->sb_hiwat / 8 * 7) {
608 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
609 			    V_tcp_autosndbuf_max);
610 
611 			if (!sbreserve_locked(sb, newsize, so, NULL))
612 				sb->sb_flags &= ~SB_AUTOSIZE;
613 			else {
614 				sowwakeup_locked(so);	/* room available */
615 				SOCKBUF_UNLOCK_ASSERT(sb);
616 				goto unlocked;
617 			}
618 		}
619 		SOCKBUF_UNLOCK(sb);
620 unlocked:
621 
622 		/* nothing to send */
623 		if (plen == 0) {
624 			KASSERT(m == NULL,
625 			    ("%s: nothing to send, but m != NULL", __func__));
626 			break;
627 		}
628 
629 		if (__predict_false(toep->flags & TPF_FIN_SENT))
630 			panic("%s: excess tx.", __func__);
631 
632 		if (plen <= max_imm) {
633 
634 			/* Immediate data tx */
635 
636 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
637 					toep->ofld_txq);
638 			if (wr == NULL) {
639 				/* XXX: how will we recover from this? */
640 				toep->flags |= TPF_TX_SUSPENDED;
641 				return;
642 			}
643 			txwr = wrtod(wr);
644 			credits = howmany(wr->wr_len, 16);
645 			write_tx_wr(txwr, toep, plen, plen, credits,
646 			    tp->t_flags & TF_MORETOCOME);
647 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
648 		} else {
649 			int wr_len;
650 
651 			/* DSGL tx */
652 
653 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
654 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
655 			wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
656 			if (wr == NULL) {
657 				/* XXX: how will we recover from this? */
658 				toep->flags |= TPF_TX_SUSPENDED;
659 				return;
660 			}
661 			txwr = wrtod(wr);
662 			credits = howmany(wr_len, 16);
663 			write_tx_wr(txwr, toep, 0, plen, credits,
664 			    tp->t_flags & TF_MORETOCOME);
665 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
666 			    max_nsegs_1mbuf);
667 			if (wr_len & 0xf) {
668 				uint64_t *pad = (uint64_t *)
669 				    ((uintptr_t)txwr + wr_len);
670 				*pad = 0;
671 			}
672 		}
673 
674 		KASSERT(toep->tx_credits >= credits,
675 			("%s: not enough credits", __func__));
676 
677 		toep->tx_credits -= credits;
678 
679 		tp->snd_nxt += plen;
680 		tp->snd_max += plen;
681 
682 		SOCKBUF_LOCK(sb);
683 		KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
684 		sb->sb_sndptr = sb_sndptr;
685 		SOCKBUF_UNLOCK(sb);
686 
687 		toep->flags |= TPF_TX_DATA_SENT;
688 
689 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
690 		txsd->plen = plen;
691 		txsd->tx_credits = credits;
692 		txsd++;
693 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
694 			toep->txsd_pidx = 0;
695 			txsd = &toep->txsd[0];
696 		}
697 		toep->txsd_avail--;
698 
699 		t4_l2t_send(sc, wr, toep->l2te);
700 	} while (m != NULL);
701 
702 	/* Send a FIN if requested, but only if there's no more data to send */
703 	if (m == NULL && toep->flags & TPF_SEND_FIN)
704 		close_conn(sc, toep);
705 }
706 
707 int
708 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
709 {
710 	struct adapter *sc = tod->tod_softc;
711 #ifdef INVARIANTS
712 	struct inpcb *inp = tp->t_inpcb;
713 #endif
714 	struct toepcb *toep = tp->t_toe;
715 
716 	INP_WLOCK_ASSERT(inp);
717 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
718 	    ("%s: inp %p dropped.", __func__, inp));
719 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
720 
721 	t4_push_frames(sc, toep);
722 
723 	return (0);
724 }
725 
726 int
727 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
728 {
729 	struct adapter *sc = tod->tod_softc;
730 #ifdef INVARIANTS
731 	struct inpcb *inp = tp->t_inpcb;
732 #endif
733 	struct toepcb *toep = tp->t_toe;
734 
735 	INP_WLOCK_ASSERT(inp);
736 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
737 	    ("%s: inp %p dropped.", __func__, inp));
738 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
739 
740 	toep->flags |= TPF_SEND_FIN;
741 	t4_push_frames(sc, toep);
742 
743 	return (0);
744 }
745 
746 int
747 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
748 {
749 	struct adapter *sc = tod->tod_softc;
750 #if defined(INVARIANTS)
751 	struct inpcb *inp = tp->t_inpcb;
752 #endif
753 	struct toepcb *toep = tp->t_toe;
754 
755 	INP_WLOCK_ASSERT(inp);
756 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
757 	    ("%s: inp %p dropped.", __func__, inp));
758 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
759 
760 	/* hmmmm */
761 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
762 	    ("%s: flowc for tid %u [%s] not sent already",
763 	    __func__, toep->tid, tcpstates[tp->t_state]));
764 
765 	send_reset(sc, toep, 0);
766 	return (0);
767 }
768 
769 /*
770  * Peer has sent us a FIN.
771  */
772 static int
773 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
774 {
775 	struct adapter *sc = iq->adapter;
776 	const struct cpl_peer_close *cpl = (const void *)(rss + 1);
777 	unsigned int tid = GET_TID(cpl);
778 	struct toepcb *toep = lookup_tid(sc, tid);
779 	struct inpcb *inp = toep->inp;
780 	struct tcpcb *tp = NULL;
781 	struct socket *so;
782 	struct sockbuf *sb;
783 #ifdef INVARIANTS
784 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
785 #endif
786 
787 	KASSERT(opcode == CPL_PEER_CLOSE,
788 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
789 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
790 
791 	if (__predict_false(toep->flags & TPF_SYNQE)) {
792 #ifdef INVARIANTS
793 		struct synq_entry *synqe = (void *)toep;
794 
795 		INP_WLOCK(synqe->lctx->inp);
796 		if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
797 			KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
798 			    ("%s: listen socket closed but tid %u not aborted.",
799 			    __func__, tid));
800 		} else {
801 			/*
802 			 * do_pass_accept_req is still running and will
803 			 * eventually take care of this tid.
804 			 */
805 		}
806 		INP_WUNLOCK(synqe->lctx->inp);
807 #endif
808 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
809 		    toep, toep->flags);
810 		return (0);
811 	}
812 
813 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
814 
815 	INP_INFO_WLOCK(&V_tcbinfo);
816 	INP_WLOCK(inp);
817 	tp = intotcpcb(inp);
818 
819 	CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
820 	    tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
821 
822 	if (toep->flags & TPF_ABORT_SHUTDOWN)
823 		goto done;
824 
825 	tp->rcv_nxt++;	/* FIN */
826 
827 	so = inp->inp_socket;
828 	sb = &so->so_rcv;
829 	SOCKBUF_LOCK(sb);
830 	if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) {
831 		m = get_ddp_mbuf(be32toh(cpl->rcv_nxt) - tp->rcv_nxt);
832 		tp->rcv_nxt = be32toh(cpl->rcv_nxt);
833 		toep->ddp_flags &= ~(DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE);
834 
835 		KASSERT(toep->sb_cc >= sb->sb_cc,
836 		    ("%s: sb %p has more data (%d) than last time (%d).",
837 		    __func__, sb, sb->sb_cc, toep->sb_cc));
838 		toep->rx_credits += toep->sb_cc - sb->sb_cc;
839 #ifdef USE_DDP_RX_FLOW_CONTROL
840 		toep->rx_credits -= m->m_len;	/* adjust for F_RX_FC_DDP */
841 #endif
842 		sbappendstream_locked(sb, m);
843 		toep->sb_cc = sb->sb_cc;
844 	}
845 	socantrcvmore_locked(so);	/* unlocks the sockbuf */
846 
847 	if (toep->ulp_mode != ULP_MODE_RDMA) {
848 		KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
849 	    		("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
850 	    		be32toh(cpl->rcv_nxt)));
851 	}
852 
853 	switch (tp->t_state) {
854 	case TCPS_SYN_RECEIVED:
855 		tp->t_starttime = ticks;
856 		/* FALLTHROUGH */
857 
858 	case TCPS_ESTABLISHED:
859 		tp->t_state = TCPS_CLOSE_WAIT;
860 		break;
861 
862 	case TCPS_FIN_WAIT_1:
863 		tp->t_state = TCPS_CLOSING;
864 		break;
865 
866 	case TCPS_FIN_WAIT_2:
867 		tcp_twstart(tp);
868 		INP_UNLOCK_ASSERT(inp);	 /* safe, we have a ref on the inp */
869 		INP_INFO_WUNLOCK(&V_tcbinfo);
870 
871 		INP_WLOCK(inp);
872 		final_cpl_received(toep);
873 		return (0);
874 
875 	default:
876 		log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
877 		    __func__, tid, tp->t_state);
878 	}
879 done:
880 	INP_WUNLOCK(inp);
881 	INP_INFO_WUNLOCK(&V_tcbinfo);
882 	return (0);
883 }
884 
885 /*
886  * Peer has ACK'd our FIN.
887  */
888 static int
889 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
890     struct mbuf *m)
891 {
892 	struct adapter *sc = iq->adapter;
893 	const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
894 	unsigned int tid = GET_TID(cpl);
895 	struct toepcb *toep = lookup_tid(sc, tid);
896 	struct inpcb *inp = toep->inp;
897 	struct tcpcb *tp = NULL;
898 	struct socket *so = NULL;
899 #ifdef INVARIANTS
900 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
901 #endif
902 
903 	KASSERT(opcode == CPL_CLOSE_CON_RPL,
904 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
905 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
906 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
907 
908 	INP_INFO_WLOCK(&V_tcbinfo);
909 	INP_WLOCK(inp);
910 	tp = intotcpcb(inp);
911 
912 	CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
913 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
914 
915 	if (toep->flags & TPF_ABORT_SHUTDOWN)
916 		goto done;
917 
918 	so = inp->inp_socket;
919 	tp->snd_una = be32toh(cpl->snd_nxt) - 1;	/* exclude FIN */
920 
921 	switch (tp->t_state) {
922 	case TCPS_CLOSING:	/* see TCPS_FIN_WAIT_2 in do_peer_close too */
923 		tcp_twstart(tp);
924 release:
925 		INP_UNLOCK_ASSERT(inp);	/* safe, we have a ref on the  inp */
926 		INP_INFO_WUNLOCK(&V_tcbinfo);
927 
928 		INP_WLOCK(inp);
929 		final_cpl_received(toep);	/* no more CPLs expected */
930 
931 		return (0);
932 	case TCPS_LAST_ACK:
933 		if (tcp_close(tp))
934 			INP_WUNLOCK(inp);
935 		goto release;
936 
937 	case TCPS_FIN_WAIT_1:
938 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
939 			soisdisconnected(so);
940 		tp->t_state = TCPS_FIN_WAIT_2;
941 		break;
942 
943 	default:
944 		log(LOG_ERR,
945 		    "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
946 		    __func__, tid, tcpstates[tp->t_state]);
947 	}
948 done:
949 	INP_WUNLOCK(inp);
950 	INP_INFO_WUNLOCK(&V_tcbinfo);
951 	return (0);
952 }
953 
954 void
955 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
956     int rst_status)
957 {
958 	struct wrqe *wr;
959 	struct cpl_abort_rpl *cpl;
960 
961 	wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
962 	if (wr == NULL) {
963 		/* XXX */
964 		panic("%s: allocation failure.", __func__);
965 	}
966 	cpl = wrtod(wr);
967 
968 	INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
969 	cpl->cmd = rst_status;
970 
971 	t4_wrq_tx(sc, wr);
972 }
973 
974 static int
975 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
976 {
977 	switch (abort_reason) {
978 	case CPL_ERR_BAD_SYN:
979 	case CPL_ERR_CONN_RESET:
980 		return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
981 	case CPL_ERR_XMIT_TIMEDOUT:
982 	case CPL_ERR_PERSIST_TIMEDOUT:
983 	case CPL_ERR_FINWAIT2_TIMEDOUT:
984 	case CPL_ERR_KEEPALIVE_TIMEDOUT:
985 		return (ETIMEDOUT);
986 	default:
987 		return (EIO);
988 	}
989 }
990 
991 /*
992  * TCP RST from the peer, timeout, or some other such critical error.
993  */
994 static int
995 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
996 {
997 	struct adapter *sc = iq->adapter;
998 	const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
999 	unsigned int tid = GET_TID(cpl);
1000 	struct toepcb *toep = lookup_tid(sc, tid);
1001 	struct sge_wrq *ofld_txq = toep->ofld_txq;
1002 	struct inpcb *inp;
1003 	struct tcpcb *tp;
1004 #ifdef INVARIANTS
1005 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1006 #endif
1007 
1008 	KASSERT(opcode == CPL_ABORT_REQ_RSS,
1009 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1010 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1011 
1012 	if (toep->flags & TPF_SYNQE)
1013 		return (do_abort_req_synqe(iq, rss, m));
1014 
1015 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1016 
1017 	if (negative_advice(cpl->status)) {
1018 		CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1019 		    __func__, cpl->status, tid, toep->flags);
1020 		return (0);	/* Ignore negative advice */
1021 	}
1022 
1023 	inp = toep->inp;
1024 	INP_INFO_WLOCK(&V_tcbinfo);	/* for tcp_close */
1025 	INP_WLOCK(inp);
1026 
1027 	tp = intotcpcb(inp);
1028 
1029 	CTR6(KTR_CXGBE,
1030 	    "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1031 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1032 	    inp->inp_flags, cpl->status);
1033 
1034 	/*
1035 	 * If we'd initiated an abort earlier the reply to it is responsible for
1036 	 * cleaning up resources.  Otherwise we tear everything down right here
1037 	 * right now.  We owe the T4 a CPL_ABORT_RPL no matter what.
1038 	 */
1039 	if (toep->flags & TPF_ABORT_SHUTDOWN) {
1040 		INP_WUNLOCK(inp);
1041 		goto done;
1042 	}
1043 	toep->flags |= TPF_ABORT_SHUTDOWN;
1044 
1045 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1046 		struct socket *so = inp->inp_socket;
1047 
1048 		if (so != NULL)
1049 			so_error_set(so, abort_status_to_errno(tp,
1050 			    cpl->status));
1051 		tp = tcp_close(tp);
1052 		if (tp == NULL)
1053 			INP_WLOCK(inp);	/* re-acquire */
1054 	}
1055 
1056 	final_cpl_received(toep);
1057 done:
1058 	INP_INFO_WUNLOCK(&V_tcbinfo);
1059 	send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1060 	return (0);
1061 }
1062 
1063 /*
1064  * Reply to the CPL_ABORT_REQ (send_reset)
1065  */
1066 static int
1067 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1068 {
1069 	struct adapter *sc = iq->adapter;
1070 	const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1071 	unsigned int tid = GET_TID(cpl);
1072 	struct toepcb *toep = lookup_tid(sc, tid);
1073 	struct inpcb *inp = toep->inp;
1074 #ifdef INVARIANTS
1075 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1076 #endif
1077 
1078 	KASSERT(opcode == CPL_ABORT_RPL_RSS,
1079 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1080 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1081 
1082 	if (toep->flags & TPF_SYNQE)
1083 		return (do_abort_rpl_synqe(iq, rss, m));
1084 
1085 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1086 
1087 	CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1088 	    __func__, tid, toep, inp, cpl->status);
1089 
1090 	KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1091 	    ("%s: wasn't expecting abort reply", __func__));
1092 
1093 	INP_WLOCK(inp);
1094 	final_cpl_received(toep);
1095 
1096 	return (0);
1097 }
1098 
1099 static int
1100 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1101 {
1102 	struct adapter *sc = iq->adapter;
1103 	const struct cpl_rx_data *cpl = mtod(m, const void *);
1104 	unsigned int tid = GET_TID(cpl);
1105 	struct toepcb *toep = lookup_tid(sc, tid);
1106 	struct inpcb *inp = toep->inp;
1107 	struct tcpcb *tp;
1108 	struct socket *so;
1109 	struct sockbuf *sb;
1110 	int len;
1111 	uint32_t ddp_placed = 0;
1112 
1113 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1114 #ifdef INVARIANTS
1115 		struct synq_entry *synqe = (void *)toep;
1116 
1117 		INP_WLOCK(synqe->lctx->inp);
1118 		if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1119 			KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1120 			    ("%s: listen socket closed but tid %u not aborted.",
1121 			    __func__, tid));
1122 		} else {
1123 			/*
1124 			 * do_pass_accept_req is still running and will
1125 			 * eventually take care of this tid.
1126 			 */
1127 		}
1128 		INP_WUNLOCK(synqe->lctx->inp);
1129 #endif
1130 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1131 		    toep, toep->flags);
1132 		m_freem(m);
1133 		return (0);
1134 	}
1135 
1136 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1137 
1138 	/* strip off CPL header */
1139 	m_adj(m, sizeof(*cpl));
1140 	len = m->m_pkthdr.len;
1141 
1142 	INP_WLOCK(inp);
1143 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1144 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1145 		    __func__, tid, len, inp->inp_flags);
1146 		INP_WUNLOCK(inp);
1147 		m_freem(m);
1148 		return (0);
1149 	}
1150 
1151 	tp = intotcpcb(inp);
1152 
1153 	if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1154 		ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1155 
1156 	tp->rcv_nxt += len;
1157 	KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__));
1158 	tp->rcv_wnd -= len;
1159 	tp->t_rcvtime = ticks;
1160 
1161 	so = inp_inpcbtosocket(inp);
1162 	sb = &so->so_rcv;
1163 	SOCKBUF_LOCK(sb);
1164 
1165 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1166 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1167 		    __func__, tid, len);
1168 		m_freem(m);
1169 		SOCKBUF_UNLOCK(sb);
1170 		INP_WUNLOCK(inp);
1171 
1172 		INP_INFO_WLOCK(&V_tcbinfo);
1173 		INP_WLOCK(inp);
1174 		tp = tcp_drop(tp, ECONNRESET);
1175 		if (tp)
1176 			INP_WUNLOCK(inp);
1177 		INP_INFO_WUNLOCK(&V_tcbinfo);
1178 
1179 		return (0);
1180 	}
1181 
1182 	/* receive buffer autosize */
1183 	if (sb->sb_flags & SB_AUTOSIZE &&
1184 	    V_tcp_do_autorcvbuf &&
1185 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1186 	    len > (sbspace(sb) / 8 * 7)) {
1187 		unsigned int hiwat = sb->sb_hiwat;
1188 		unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1189 		    V_tcp_autorcvbuf_max);
1190 
1191 		if (!sbreserve_locked(sb, newsize, so, NULL))
1192 			sb->sb_flags &= ~SB_AUTOSIZE;
1193 		else
1194 			toep->rx_credits += newsize - hiwat;
1195 	}
1196 
1197 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1198 		int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off;
1199 
1200 		if (changed) {
1201 			if (toep->ddp_flags & DDP_SC_REQ)
1202 				toep->ddp_flags ^= DDP_ON | DDP_SC_REQ;
1203 			else {
1204 				KASSERT(cpl->ddp_off == 1,
1205 				    ("%s: DDP switched on by itself.",
1206 				    __func__));
1207 
1208 				/* Fell out of DDP mode */
1209 				toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE |
1210 				    DDP_BUF1_ACTIVE);
1211 
1212 				if (ddp_placed)
1213 					insert_ddp_data(toep, ddp_placed);
1214 			}
1215 		}
1216 
1217 		if ((toep->ddp_flags & DDP_OK) == 0 &&
1218 		    time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) {
1219 			toep->ddp_score = DDP_LOW_SCORE;
1220 			toep->ddp_flags |= DDP_OK;
1221 			CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u",
1222 			    __func__, tid, time_uptime);
1223 		}
1224 
1225 		if (toep->ddp_flags & DDP_ON) {
1226 
1227 			/*
1228 			 * CPL_RX_DATA with DDP on can only be an indicate.  Ask
1229 			 * soreceive to post a buffer or disable DDP.  The
1230 			 * payload that arrived in this indicate is appended to
1231 			 * the socket buffer as usual.
1232 			 */
1233 
1234 #if 0
1235 			CTR5(KTR_CXGBE,
1236 			    "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)",
1237 			    __func__, tid, toep->flags, be32toh(cpl->seq), len);
1238 #endif
1239 			sb->sb_flags |= SB_DDP_INDICATE;
1240 		} else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK &&
1241 		    tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) {
1242 
1243 			/*
1244 			 * DDP allowed but isn't on (and a request to switch it
1245 			 * on isn't pending either), and conditions are ripe for
1246 			 * it to work.  Switch it on.
1247 			 */
1248 
1249 			enable_ddp(sc, toep);
1250 		}
1251 	}
1252 
1253 	KASSERT(toep->sb_cc >= sb->sb_cc,
1254 	    ("%s: sb %p has more data (%d) than last time (%d).",
1255 	    __func__, sb, sb->sb_cc, toep->sb_cc));
1256 	toep->rx_credits += toep->sb_cc - sb->sb_cc;
1257 	sbappendstream_locked(sb, m);
1258 	toep->sb_cc = sb->sb_cc;
1259 	sorwakeup_locked(so);
1260 	SOCKBUF_UNLOCK_ASSERT(sb);
1261 
1262 	INP_WUNLOCK(inp);
1263 	return (0);
1264 }
1265 
1266 #define S_CPL_FW4_ACK_OPCODE    24
1267 #define M_CPL_FW4_ACK_OPCODE    0xff
1268 #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE)
1269 #define G_CPL_FW4_ACK_OPCODE(x) \
1270     (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE)
1271 
1272 #define S_CPL_FW4_ACK_FLOWID    0
1273 #define M_CPL_FW4_ACK_FLOWID    0xffffff
1274 #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID)
1275 #define G_CPL_FW4_ACK_FLOWID(x) \
1276     (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID)
1277 
1278 #define S_CPL_FW4_ACK_CR        24
1279 #define M_CPL_FW4_ACK_CR        0xff
1280 #define V_CPL_FW4_ACK_CR(x)     ((x) << S_CPL_FW4_ACK_CR)
1281 #define G_CPL_FW4_ACK_CR(x)     (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR)
1282 
1283 #define S_CPL_FW4_ACK_SEQVAL    0
1284 #define M_CPL_FW4_ACK_SEQVAL    0x1
1285 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL)
1286 #define G_CPL_FW4_ACK_SEQVAL(x) \
1287     (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL)
1288 #define F_CPL_FW4_ACK_SEQVAL    V_CPL_FW4_ACK_SEQVAL(1U)
1289 
1290 static int
1291 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1292 {
1293 	struct adapter *sc = iq->adapter;
1294 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1295 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1296 	struct toepcb *toep = lookup_tid(sc, tid);
1297 	struct inpcb *inp;
1298 	struct tcpcb *tp;
1299 	struct socket *so;
1300 	uint8_t credits = cpl->credits;
1301 	struct ofld_tx_sdesc *txsd;
1302 	int plen;
1303 #ifdef INVARIANTS
1304 	unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1305 #endif
1306 
1307 	/*
1308 	 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1309 	 * now this comes back carrying the credits for the flowc.
1310 	 */
1311 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1312 		KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1313 		    ("%s: credits for a synq entry %p", __func__, toep));
1314 		return (0);
1315 	}
1316 
1317 	inp = toep->inp;
1318 
1319 	KASSERT(opcode == CPL_FW4_ACK,
1320 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1321 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1322 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1323 
1324 	INP_WLOCK(inp);
1325 
1326 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1327 		INP_WUNLOCK(inp);
1328 		return (0);
1329 	}
1330 
1331 	KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1332 	    ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1333 
1334 	tp = intotcpcb(inp);
1335 
1336 	if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1337 		tcp_seq snd_una = be32toh(cpl->snd_una);
1338 
1339 #ifdef INVARIANTS
1340 		if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1341 			log(LOG_ERR,
1342 			    "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1343 			    __func__, snd_una, toep->tid, tp->snd_una);
1344 		}
1345 #endif
1346 
1347 		if (tp->snd_una != snd_una) {
1348 			tp->snd_una = snd_una;
1349 			tp->ts_recent_age = tcp_ts_getticks();
1350 		}
1351 	}
1352 
1353 	so = inp->inp_socket;
1354 	txsd = &toep->txsd[toep->txsd_cidx];
1355 	plen = 0;
1356 	while (credits) {
1357 		KASSERT(credits >= txsd->tx_credits,
1358 		    ("%s: too many (or partial) credits", __func__));
1359 		credits -= txsd->tx_credits;
1360 		toep->tx_credits += txsd->tx_credits;
1361 		plen += txsd->plen;
1362 		txsd++;
1363 		toep->txsd_avail++;
1364 		KASSERT(toep->txsd_avail <= toep->txsd_total,
1365 		    ("%s: txsd avail > total", __func__));
1366 		if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1367 			txsd = &toep->txsd[0];
1368 			toep->txsd_cidx = 0;
1369 		}
1370 	}
1371 
1372 	if (plen > 0) {
1373 		struct sockbuf *sb = &so->so_snd;
1374 
1375 		SOCKBUF_LOCK(sb);
1376 		sbdrop_locked(sb, plen);
1377 		sowwakeup_locked(so);
1378 		SOCKBUF_UNLOCK_ASSERT(sb);
1379 	}
1380 
1381 	/* XXX */
1382 	if ((toep->flags & TPF_TX_SUSPENDED &&
1383 	    toep->tx_credits >= MIN_OFLD_TX_CREDITS) ||
1384 	    toep->tx_credits == toep->txsd_total *
1385 	    howmany((sizeof(struct fw_ofld_tx_data_wr) + 1), 16)) {
1386 		toep->flags &= ~TPF_TX_SUSPENDED;
1387 		t4_push_frames(sc, toep);
1388 	}
1389 	INP_WUNLOCK(inp);
1390 
1391 	return (0);
1392 }
1393 
1394 static int
1395 do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1396 {
1397 	struct adapter *sc = iq->adapter;
1398 	const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1);
1399 	unsigned int tid = GET_TID(cpl);
1400 #ifdef INVARIANTS
1401 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1402 #endif
1403 
1404 	KASSERT(opcode == CPL_SET_TCB_RPL,
1405 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1406 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1407 
1408 	if (tid >= sc->tids.ftid_base &&
1409 	    tid < sc->tids.ftid_base + sc->tids.nftids)
1410 		return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */
1411 
1412 	CXGBE_UNIMPLEMENTED(__func__);
1413 }
1414 
1415 void
1416 t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl,
1417     uint16_t word, uint64_t mask, uint64_t val)
1418 {
1419 	struct wrqe *wr;
1420 	struct cpl_set_tcb_field *req;
1421 
1422 	wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq);
1423 	if (wr == NULL) {
1424 		/* XXX */
1425 		panic("%s: allocation failure.", __func__);
1426 	}
1427 	req = wrtod(wr);
1428 
1429 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1430 	req->reply_ctrl = htobe16(V_NO_REPLY(1) |
1431 	    V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1432 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
1433 	req->mask = htobe64(mask);
1434 	req->val = htobe64(val);
1435 
1436 	t4_wrq_tx(sc, wr);
1437 }
1438 
1439 void
1440 t4_init_cpl_io_handlers(struct adapter *sc)
1441 {
1442 
1443 	t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close);
1444 	t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl);
1445 	t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req);
1446 	t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl);
1447 	t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data);
1448 	t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack);
1449 	t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl);
1450 }
1451 
1452 void
1453 t4_uninit_cpl_io_handlers(struct adapter *sc)
1454 {
1455 
1456 	t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl);
1457 }
1458 #endif
1459