xref: /freebsd/sys/dev/cxgbe/tom/t4_cpl_io.c (revision 89e3d5671ba13dceca272d5b159c9bd805f3f504)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2012, 2015 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_kern_tls.h"
36 #include "opt_ratelimit.h"
37 
38 #ifdef TCP_OFFLOAD
39 #include <sys/param.h>
40 #include <sys/aio.h>
41 #include <sys/file.h>
42 #include <sys/kernel.h>
43 #include <sys/ktr.h>
44 #include <sys/module.h>
45 #include <sys/proc.h>
46 #include <sys/protosw.h>
47 #include <sys/domain.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sglist.h>
51 #include <sys/taskqueue.h>
52 #include <netinet/in.h>
53 #include <netinet/in_pcb.h>
54 #include <netinet/ip.h>
55 #include <netinet/ip6.h>
56 #define TCPSTATES
57 #include <netinet/tcp_fsm.h>
58 #include <netinet/tcp_seq.h>
59 #include <netinet/tcp_var.h>
60 #include <netinet/toecore.h>
61 
62 #include <security/mac/mac_framework.h>
63 
64 #include <vm/vm.h>
65 #include <vm/vm_extern.h>
66 #include <vm/pmap.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_page.h>
69 
70 #include "common/common.h"
71 #include "common/t4_msg.h"
72 #include "common/t4_regs.h"
73 #include "common/t4_tcb.h"
74 #include "tom/t4_tom_l2t.h"
75 #include "tom/t4_tom.h"
76 
77 static void	t4_aiotx_cancel(struct kaiocb *job);
78 static void	t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep);
79 
80 void
81 send_flowc_wr(struct toepcb *toep, struct tcpcb *tp)
82 {
83 	struct wrqe *wr;
84 	struct fw_flowc_wr *flowc;
85 	unsigned int nparams, flowclen, paramidx;
86 	struct vi_info *vi = toep->vi;
87 	struct port_info *pi = vi->pi;
88 	struct adapter *sc = pi->adapter;
89 	unsigned int pfvf = sc->pf << S_FW_VIID_PFN;
90 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
91 
92 	KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
93 	    ("%s: flowc for tid %u sent already", __func__, toep->tid));
94 
95 	if (tp != NULL)
96 		nparams = 8;
97 	else
98 		nparams = 6;
99 	if (ulp_mode(toep) == ULP_MODE_TLS)
100 		nparams++;
101 	if (toep->tls.fcplenmax != 0)
102 		nparams++;
103 	if (toep->params.tc_idx != -1) {
104 		MPASS(toep->params.tc_idx >= 0 &&
105 		    toep->params.tc_idx < sc->chip_params->nsched_cls);
106 		nparams++;
107 	}
108 
109 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
110 
111 	wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
112 	if (wr == NULL) {
113 		/* XXX */
114 		panic("%s: allocation failure.", __func__);
115 	}
116 	flowc = wrtod(wr);
117 	memset(flowc, 0, wr->wr_len);
118 
119 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
120 	    V_FW_FLOWC_WR_NPARAMS(nparams));
121 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
122 	    V_FW_WR_FLOWID(toep->tid));
123 
124 #define FLOWC_PARAM(__m, __v) \
125 	do { \
126 		flowc->mnemval[paramidx].mnemonic = FW_FLOWC_MNEM_##__m; \
127 		flowc->mnemval[paramidx].val = htobe32(__v); \
128 		paramidx++; \
129 	} while (0)
130 
131 	paramidx = 0;
132 
133 	FLOWC_PARAM(PFNVFN, pfvf);
134 	FLOWC_PARAM(CH, pi->tx_chan);
135 	FLOWC_PARAM(PORT, pi->tx_chan);
136 	FLOWC_PARAM(IQID, toep->ofld_rxq->iq.abs_id);
137 	FLOWC_PARAM(SNDBUF, toep->params.sndbuf);
138 	if (tp) {
139 		FLOWC_PARAM(MSS, toep->params.emss);
140 		FLOWC_PARAM(SNDNXT, tp->snd_nxt);
141 		FLOWC_PARAM(RCVNXT, tp->rcv_nxt);
142 	} else
143 		FLOWC_PARAM(MSS, 512);
144 	CTR6(KTR_CXGBE,
145 	    "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x",
146 	    __func__, toep->tid, toep->params.emss, toep->params.sndbuf,
147 	    tp ? tp->snd_nxt : 0, tp ? tp->rcv_nxt : 0);
148 
149 	if (ulp_mode(toep) == ULP_MODE_TLS)
150 		FLOWC_PARAM(ULP_MODE, ulp_mode(toep));
151 	if (toep->tls.fcplenmax != 0)
152 		FLOWC_PARAM(TXDATAPLEN_MAX, toep->tls.fcplenmax);
153 	if (toep->params.tc_idx != -1)
154 		FLOWC_PARAM(SCHEDCLASS, toep->params.tc_idx);
155 #undef FLOWC_PARAM
156 
157 	KASSERT(paramidx == nparams, ("nparams mismatch"));
158 
159 	txsd->tx_credits = howmany(flowclen, 16);
160 	txsd->plen = 0;
161 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
162 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
163 	toep->tx_credits -= txsd->tx_credits;
164 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
165 		toep->txsd_pidx = 0;
166 	toep->txsd_avail--;
167 
168 	toep->flags |= TPF_FLOWC_WR_SENT;
169         t4_wrq_tx(sc, wr);
170 }
171 
172 #ifdef RATELIMIT
173 /*
174  * Input is Bytes/second (so_max_pacing_rate), chip counts in Kilobits/second.
175  */
176 static int
177 update_tx_rate_limit(struct adapter *sc, struct toepcb *toep, u_int Bps)
178 {
179 	int tc_idx, rc;
180 	const u_int kbps = (u_int) (uint64_t)Bps * 8ULL / 1000;
181 	const int port_id = toep->vi->pi->port_id;
182 
183 	CTR3(KTR_CXGBE, "%s: tid %u, rate %uKbps", __func__, toep->tid, kbps);
184 
185 	if (kbps == 0) {
186 		/* unbind */
187 		tc_idx = -1;
188 	} else {
189 		rc = t4_reserve_cl_rl_kbps(sc, port_id, kbps, &tc_idx);
190 		if (rc != 0)
191 			return (rc);
192 		MPASS(tc_idx >= 0 && tc_idx < sc->chip_params->nsched_cls);
193 	}
194 
195 	if (toep->params.tc_idx != tc_idx) {
196 		struct wrqe *wr;
197 		struct fw_flowc_wr *flowc;
198 		int nparams = 1, flowclen, flowclen16;
199 		struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
200 
201 		flowclen = sizeof(*flowc) + nparams * sizeof(struct
202 		    fw_flowc_mnemval);
203 		flowclen16 = howmany(flowclen, 16);
204 		if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0 ||
205 		    (wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq)) == NULL) {
206 			if (tc_idx >= 0)
207 				t4_release_cl_rl(sc, port_id, tc_idx);
208 			return (ENOMEM);
209 		}
210 
211 		flowc = wrtod(wr);
212 		memset(flowc, 0, wr->wr_len);
213 
214 		flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
215 		    V_FW_FLOWC_WR_NPARAMS(nparams));
216 		flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
217 		    V_FW_WR_FLOWID(toep->tid));
218 
219 		flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
220 		if (tc_idx == -1)
221 			flowc->mnemval[0].val = htobe32(0xff);
222 		else
223 			flowc->mnemval[0].val = htobe32(tc_idx);
224 
225 		txsd->tx_credits = flowclen16;
226 		txsd->plen = 0;
227 		toep->tx_credits -= txsd->tx_credits;
228 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
229 			toep->txsd_pidx = 0;
230 		toep->txsd_avail--;
231 		t4_wrq_tx(sc, wr);
232 	}
233 
234 	if (toep->params.tc_idx >= 0)
235 		t4_release_cl_rl(sc, port_id, toep->params.tc_idx);
236 	toep->params.tc_idx = tc_idx;
237 
238 	return (0);
239 }
240 #endif
241 
242 void
243 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
244 {
245 	struct wrqe *wr;
246 	struct cpl_abort_req *req;
247 	int tid = toep->tid;
248 	struct inpcb *inp = toep->inp;
249 	struct tcpcb *tp = intotcpcb(inp);	/* don't use if INP_DROPPED */
250 
251 	INP_WLOCK_ASSERT(inp);
252 
253 	CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
254 	    __func__, toep->tid,
255 	    inp->inp_flags & INP_DROPPED ? "inp dropped" :
256 	    tcpstates[tp->t_state],
257 	    toep->flags, inp->inp_flags,
258 	    toep->flags & TPF_ABORT_SHUTDOWN ?
259 	    " (abort already in progress)" : "");
260 
261 	if (toep->flags & TPF_ABORT_SHUTDOWN)
262 		return;	/* abort already in progress */
263 
264 	toep->flags |= TPF_ABORT_SHUTDOWN;
265 
266 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
267 	    ("%s: flowc_wr not sent for tid %d.", __func__, tid));
268 
269 	wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
270 	if (wr == NULL) {
271 		/* XXX */
272 		panic("%s: allocation failure.", __func__);
273 	}
274 	req = wrtod(wr);
275 
276 	INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
277 	if (inp->inp_flags & INP_DROPPED)
278 		req->rsvd0 = htobe32(snd_nxt);
279 	else
280 		req->rsvd0 = htobe32(tp->snd_nxt);
281 	req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
282 	req->cmd = CPL_ABORT_SEND_RST;
283 
284 	/*
285 	 * XXX: What's the correct way to tell that the inp hasn't been detached
286 	 * from its socket?  Should I even be flushing the snd buffer here?
287 	 */
288 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
289 		struct socket *so = inp->inp_socket;
290 
291 		if (so != NULL)	/* because I'm not sure.  See comment above */
292 			sbflush(&so->so_snd);
293 	}
294 
295 	t4_l2t_send(sc, wr, toep->l2te);
296 }
297 
298 /*
299  * Called when a connection is established to translate the TCP options
300  * reported by HW to FreeBSD's native format.
301  */
302 static void
303 assign_rxopt(struct tcpcb *tp, uint16_t opt)
304 {
305 	struct toepcb *toep = tp->t_toe;
306 	struct inpcb *inp = tp->t_inpcb;
307 	struct adapter *sc = td_adapter(toep->td);
308 
309 	INP_LOCK_ASSERT(inp);
310 
311 	toep->params.mtu_idx = G_TCPOPT_MSS(opt);
312 	tp->t_maxseg = sc->params.mtus[toep->params.mtu_idx];
313 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
314 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
315 	else
316 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
317 
318 	toep->params.emss = tp->t_maxseg;
319 	if (G_TCPOPT_TSTAMP(opt)) {
320 		toep->params.tstamp = 1;
321 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
322 		tp->t_flags |= TF_RCVD_TSTMP;	/* timestamps ok */
323 		tp->ts_recent = 0;		/* hmmm */
324 		tp->ts_recent_age = tcp_ts_getticks();
325 	} else
326 		toep->params.tstamp = 0;
327 
328 	if (G_TCPOPT_SACK(opt)) {
329 		toep->params.sack = 1;
330 		tp->t_flags |= TF_SACK_PERMIT;	/* should already be set */
331 	} else {
332 		toep->params.sack = 0;
333 		tp->t_flags &= ~TF_SACK_PERMIT;	/* sack disallowed by peer */
334 	}
335 
336 	if (G_TCPOPT_WSCALE_OK(opt))
337 		tp->t_flags |= TF_RCVD_SCALE;
338 
339 	/* Doing window scaling? */
340 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
341 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
342 		tp->rcv_scale = tp->request_r_scale;
343 		tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
344 	} else
345 		toep->params.wscale = 0;
346 
347 	CTR6(KTR_CXGBE,
348 	    "assign_rxopt: tid %d, mtu_idx %u, emss %u, ts %u, sack %u, wscale %u",
349 	    toep->tid, toep->params.mtu_idx, toep->params.emss,
350 	    toep->params.tstamp, toep->params.sack, toep->params.wscale);
351 }
352 
353 /*
354  * Completes some final bits of initialization for just established connections
355  * and changes their state to TCPS_ESTABLISHED.
356  *
357  * The ISNs are from the exchange of SYNs.
358  */
359 void
360 make_established(struct toepcb *toep, uint32_t iss, uint32_t irs, uint16_t opt)
361 {
362 	struct inpcb *inp = toep->inp;
363 	struct socket *so = inp->inp_socket;
364 	struct tcpcb *tp = intotcpcb(inp);
365 	uint16_t tcpopt = be16toh(opt);
366 
367 	INP_WLOCK_ASSERT(inp);
368 	KASSERT(tp->t_state == TCPS_SYN_SENT ||
369 	    tp->t_state == TCPS_SYN_RECEIVED,
370 	    ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
371 
372 	CTR6(KTR_CXGBE, "%s: tid %d, so %p, inp %p, tp %p, toep %p",
373 	    __func__, toep->tid, so, inp, tp, toep);
374 
375 	tcp_state_change(tp, TCPS_ESTABLISHED);
376 	tp->t_starttime = ticks;
377 	TCPSTAT_INC(tcps_connects);
378 
379 	tp->irs = irs;
380 	tcp_rcvseqinit(tp);
381 	tp->rcv_wnd = (u_int)toep->params.opt0_bufsize << 10;
382 	tp->rcv_adv += tp->rcv_wnd;
383 	tp->last_ack_sent = tp->rcv_nxt;
384 
385 	tp->iss = iss;
386 	tcp_sendseqinit(tp);
387 	tp->snd_una = iss + 1;
388 	tp->snd_nxt = iss + 1;
389 	tp->snd_max = iss + 1;
390 
391 	assign_rxopt(tp, tcpopt);
392 	send_flowc_wr(toep, tp);
393 
394 	soisconnected(so);
395 
396 	if (ulp_mode(toep) == ULP_MODE_TLS)
397 		tls_establish(toep);
398 }
399 
400 int
401 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
402 {
403 	struct wrqe *wr;
404 	struct cpl_rx_data_ack *req;
405 	uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
406 
407 	KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
408 
409 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
410 	if (wr == NULL)
411 		return (0);
412 	req = wrtod(wr);
413 
414 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
415 	req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
416 
417 	t4_wrq_tx(sc, wr);
418 	return (credits);
419 }
420 
421 void
422 send_rx_modulate(struct adapter *sc, struct toepcb *toep)
423 {
424 	struct wrqe *wr;
425 	struct cpl_rx_data_ack *req;
426 
427 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
428 	if (wr == NULL)
429 		return;
430 	req = wrtod(wr);
431 
432 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
433 	req->credit_dack = htobe32(F_RX_MODULATE_RX);
434 
435 	t4_wrq_tx(sc, wr);
436 }
437 
438 void
439 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp)
440 {
441 	struct adapter *sc = tod->tod_softc;
442 	struct inpcb *inp = tp->t_inpcb;
443 	struct socket *so = inp->inp_socket;
444 	struct sockbuf *sb = &so->so_rcv;
445 	struct toepcb *toep = tp->t_toe;
446 	int rx_credits;
447 
448 	INP_WLOCK_ASSERT(inp);
449 	SOCKBUF_LOCK_ASSERT(sb);
450 
451 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
452 	if (rx_credits > 0 &&
453 	    (tp->rcv_wnd <= 32 * 1024 || rx_credits >= 64 * 1024 ||
454 	    (rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
455 	    sbused(sb) + tp->rcv_wnd < sb->sb_lowat)) {
456 		rx_credits = send_rx_credits(sc, toep, rx_credits);
457 		tp->rcv_wnd += rx_credits;
458 		tp->rcv_adv += rx_credits;
459 	} else if (toep->flags & TPF_FORCE_CREDITS)
460 		send_rx_modulate(sc, toep);
461 }
462 
463 void
464 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
465 {
466 	struct inpcb *inp = tp->t_inpcb;
467 	struct socket *so = inp->inp_socket;
468 	struct sockbuf *sb = &so->so_rcv;
469 
470 	SOCKBUF_LOCK(sb);
471 	t4_rcvd_locked(tod, tp);
472 	SOCKBUF_UNLOCK(sb);
473 }
474 
475 /*
476  * Close a connection by sending a CPL_CLOSE_CON_REQ message.
477  */
478 int
479 t4_close_conn(struct adapter *sc, struct toepcb *toep)
480 {
481 	struct wrqe *wr;
482 	struct cpl_close_con_req *req;
483 	unsigned int tid = toep->tid;
484 
485 	CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
486 	    toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
487 
488 	if (toep->flags & TPF_FIN_SENT)
489 		return (0);
490 
491 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
492 	    ("%s: flowc_wr not sent for tid %u.", __func__, tid));
493 
494 	wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
495 	if (wr == NULL) {
496 		/* XXX */
497 		panic("%s: allocation failure.", __func__);
498 	}
499 	req = wrtod(wr);
500 
501         req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
502 	    V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
503 	req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
504 	    V_FW_WR_FLOWID(tid));
505         req->wr.wr_lo = cpu_to_be64(0);
506         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
507 	req->rsvd = 0;
508 
509 	toep->flags |= TPF_FIN_SENT;
510 	toep->flags &= ~TPF_SEND_FIN;
511 	t4_l2t_send(sc, wr, toep->l2te);
512 
513 	return (0);
514 }
515 
516 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
517 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
518 
519 /* Maximum amount of immediate data we could stuff in a WR */
520 static inline int
521 max_imm_payload(int tx_credits)
522 {
523 	const int n = 1;	/* Use no more than one desc for imm. data WR */
524 
525 	KASSERT(tx_credits >= 0 &&
526 		tx_credits <= MAX_OFLD_TX_CREDITS,
527 		("%s: %d credits", __func__, tx_credits));
528 
529 	if (tx_credits < MIN_OFLD_TX_CREDITS)
530 		return (0);
531 
532 	if (tx_credits >= (n * EQ_ESIZE) / 16)
533 		return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
534 	else
535 		return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
536 }
537 
538 /* Maximum number of SGL entries we could stuff in a WR */
539 static inline int
540 max_dsgl_nsegs(int tx_credits)
541 {
542 	int nseg = 1;	/* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
543 	int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
544 
545 	KASSERT(tx_credits >= 0 &&
546 		tx_credits <= MAX_OFLD_TX_CREDITS,
547 		("%s: %d credits", __func__, tx_credits));
548 
549 	if (tx_credits < MIN_OFLD_TX_CREDITS)
550 		return (0);
551 
552 	nseg += 2 * (sge_pair_credits * 16 / 24);
553 	if ((sge_pair_credits * 16) % 24 == 16)
554 		nseg++;
555 
556 	return (nseg);
557 }
558 
559 static inline void
560 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
561     unsigned int plen, uint8_t credits, int shove, int ulp_submode)
562 {
563 	struct fw_ofld_tx_data_wr *txwr = dst;
564 
565 	txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
566 	    V_FW_WR_IMMDLEN(immdlen));
567 	txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
568 	    V_FW_WR_LEN16(credits));
569 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ulp_mode(toep)) |
570 	    V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
571 	txwr->plen = htobe32(plen);
572 
573 	if (toep->params.tx_align > 0) {
574 		if (plen < 2 * toep->params.emss)
575 			txwr->lsodisable_to_flags |=
576 			    htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
577 		else
578 			txwr->lsodisable_to_flags |=
579 			    htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
580 				(toep->params.nagle == 0 ? 0 :
581 				F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
582 	}
583 }
584 
585 /*
586  * Generate a DSGL from a starting mbuf.  The total number of segments and the
587  * maximum segments in any one mbuf are provided.
588  */
589 static void
590 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
591 {
592 	struct mbuf *m;
593 	struct ulptx_sgl *usgl = dst;
594 	int i, j, rc;
595 	struct sglist sg;
596 	struct sglist_seg segs[n];
597 
598 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
599 
600 	sglist_init(&sg, n, segs);
601 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
602 	    V_ULPTX_NSGE(nsegs));
603 
604 	i = -1;
605 	for (m = start; m != stop; m = m->m_next) {
606 		if (m->m_flags & M_EXTPG)
607 			rc = sglist_append_mbuf_epg(&sg, m,
608 			    mtod(m, vm_offset_t), m->m_len);
609 		else
610 			rc = sglist_append(&sg, mtod(m, void *), m->m_len);
611 		if (__predict_false(rc != 0))
612 			panic("%s: sglist_append %d", __func__, rc);
613 
614 		for (j = 0; j < sg.sg_nseg; i++, j++) {
615 			if (i < 0) {
616 				usgl->len0 = htobe32(segs[j].ss_len);
617 				usgl->addr0 = htobe64(segs[j].ss_paddr);
618 			} else {
619 				usgl->sge[i / 2].len[i & 1] =
620 				    htobe32(segs[j].ss_len);
621 				usgl->sge[i / 2].addr[i & 1] =
622 				    htobe64(segs[j].ss_paddr);
623 			}
624 #ifdef INVARIANTS
625 			nsegs--;
626 #endif
627 		}
628 		sglist_reset(&sg);
629 	}
630 	if (i & 1)
631 		usgl->sge[i / 2].len[1] = htobe32(0);
632 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
633 	    __func__, nsegs, start, stop));
634 }
635 
636 /*
637  * Max number of SGL entries an offload tx work request can have.  This is 41
638  * (1 + 40) for a full 512B work request.
639  * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
640  */
641 #define OFLD_SGL_LEN (41)
642 
643 /*
644  * Send data and/or a FIN to the peer.
645  *
646  * The socket's so_snd buffer consists of a stream of data starting with sb_mb
647  * and linked together with m_next.  sb_sndptr, if set, is the last mbuf that
648  * was transmitted.
649  *
650  * drop indicates the number of bytes that should be dropped from the head of
651  * the send buffer.  It is an optimization that lets do_fw4_ack avoid creating
652  * contention on the send buffer lock (before this change it used to do
653  * sowwakeup and then t4_push_frames right after that when recovering from tx
654  * stalls).  When drop is set this function MUST drop the bytes and wake up any
655  * writers.
656  */
657 void
658 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
659 {
660 	struct mbuf *sndptr, *m, *sb_sndptr;
661 	struct fw_ofld_tx_data_wr *txwr;
662 	struct wrqe *wr;
663 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
664 	struct inpcb *inp = toep->inp;
665 	struct tcpcb *tp = intotcpcb(inp);
666 	struct socket *so = inp->inp_socket;
667 	struct sockbuf *sb = &so->so_snd;
668 	int tx_credits, shove, compl, sowwakeup;
669 	struct ofld_tx_sdesc *txsd;
670 	bool nomap_mbuf_seen;
671 
672 	INP_WLOCK_ASSERT(inp);
673 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
674 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
675 
676 	KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
677 	    ulp_mode(toep) == ULP_MODE_TCPDDP ||
678 	    ulp_mode(toep) == ULP_MODE_TLS ||
679 	    ulp_mode(toep) == ULP_MODE_RDMA,
680 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
681 
682 #ifdef VERBOSE_TRACES
683 	CTR5(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
684 	    __func__, toep->tid, toep->flags, tp->t_flags, drop);
685 #endif
686 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
687 		return;
688 
689 #ifdef RATELIMIT
690 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
691 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
692 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
693 	}
694 #endif
695 
696 	/*
697 	 * This function doesn't resume by itself.  Someone else must clear the
698 	 * flag and call this function.
699 	 */
700 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
701 		KASSERT(drop == 0,
702 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
703 		return;
704 	}
705 
706 	txsd = &toep->txsd[toep->txsd_pidx];
707 	do {
708 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
709 		max_imm = max_imm_payload(tx_credits);
710 		max_nsegs = max_dsgl_nsegs(tx_credits);
711 
712 		SOCKBUF_LOCK(sb);
713 		sowwakeup = drop;
714 		if (drop) {
715 			sbdrop_locked(sb, drop);
716 			drop = 0;
717 		}
718 		sb_sndptr = sb->sb_sndptr;
719 		sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
720 		plen = 0;
721 		nsegs = 0;
722 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
723 		nomap_mbuf_seen = false;
724 		for (m = sndptr; m != NULL; m = m->m_next) {
725 			int n;
726 
727 			if ((m->m_flags & M_NOTAVAIL) != 0)
728 				break;
729 			if (m->m_flags & M_EXTPG) {
730 #ifdef KERN_TLS
731 				if (m->m_epg_tls != NULL) {
732 					toep->flags |= TPF_KTLS;
733 					if (plen == 0) {
734 						SOCKBUF_UNLOCK(sb);
735 						t4_push_ktls(sc, toep, 0);
736 						return;
737 					}
738 					break;
739 				}
740 #endif
741 				n = sglist_count_mbuf_epg(m,
742 				    mtod(m, vm_offset_t), m->m_len);
743 			} else
744 				n = sglist_count(mtod(m, void *), m->m_len);
745 
746 			nsegs += n;
747 			plen += m->m_len;
748 
749 			/* This mbuf sent us _over_ the nsegs limit, back out */
750 			if (plen > max_imm && nsegs > max_nsegs) {
751 				nsegs -= n;
752 				plen -= m->m_len;
753 				if (plen == 0) {
754 					/* Too few credits */
755 					toep->flags |= TPF_TX_SUSPENDED;
756 					if (sowwakeup) {
757 						if (!TAILQ_EMPTY(
758 						    &toep->aiotx_jobq))
759 							t4_aiotx_queue_toep(so,
760 							    toep);
761 						sowwakeup_locked(so);
762 					} else
763 						SOCKBUF_UNLOCK(sb);
764 					SOCKBUF_UNLOCK_ASSERT(sb);
765 					return;
766 				}
767 				break;
768 			}
769 
770 			if (m->m_flags & M_EXTPG)
771 				nomap_mbuf_seen = true;
772 			if (max_nsegs_1mbuf < n)
773 				max_nsegs_1mbuf = n;
774 			sb_sndptr = m;	/* new sb->sb_sndptr if all goes well */
775 
776 			/* This mbuf put us right at the max_nsegs limit */
777 			if (plen > max_imm && nsegs == max_nsegs) {
778 				m = m->m_next;
779 				break;
780 			}
781 		}
782 
783 		if (sbused(sb) > sb->sb_hiwat * 5 / 8 &&
784 		    toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
785 			compl = 1;
786 		else
787 			compl = 0;
788 
789 		if (sb->sb_flags & SB_AUTOSIZE &&
790 		    V_tcp_do_autosndbuf &&
791 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
792 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
793 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
794 			    V_tcp_autosndbuf_max);
795 
796 			if (!sbreserve_locked(sb, newsize, so, NULL))
797 				sb->sb_flags &= ~SB_AUTOSIZE;
798 			else
799 				sowwakeup = 1;	/* room available */
800 		}
801 		if (sowwakeup) {
802 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
803 				t4_aiotx_queue_toep(so, toep);
804 			sowwakeup_locked(so);
805 		} else
806 			SOCKBUF_UNLOCK(sb);
807 		SOCKBUF_UNLOCK_ASSERT(sb);
808 
809 		/* nothing to send */
810 		if (plen == 0) {
811 			KASSERT(m == NULL || (m->m_flags & M_NOTAVAIL) != 0,
812 			    ("%s: nothing to send, but m != NULL is ready",
813 			    __func__));
814 			break;
815 		}
816 
817 		if (__predict_false(toep->flags & TPF_FIN_SENT))
818 			panic("%s: excess tx.", __func__);
819 
820 		shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
821 		if (plen <= max_imm && !nomap_mbuf_seen) {
822 
823 			/* Immediate data tx */
824 
825 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
826 					toep->ofld_txq);
827 			if (wr == NULL) {
828 				/* XXX: how will we recover from this? */
829 				toep->flags |= TPF_TX_SUSPENDED;
830 				return;
831 			}
832 			txwr = wrtod(wr);
833 			credits = howmany(wr->wr_len, 16);
834 			write_tx_wr(txwr, toep, plen, plen, credits, shove, 0);
835 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
836 			nsegs = 0;
837 		} else {
838 			int wr_len;
839 
840 			/* DSGL tx */
841 
842 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
843 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
844 			wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
845 			if (wr == NULL) {
846 				/* XXX: how will we recover from this? */
847 				toep->flags |= TPF_TX_SUSPENDED;
848 				return;
849 			}
850 			txwr = wrtod(wr);
851 			credits = howmany(wr_len, 16);
852 			write_tx_wr(txwr, toep, 0, plen, credits, shove, 0);
853 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
854 			    max_nsegs_1mbuf);
855 			if (wr_len & 0xf) {
856 				uint64_t *pad = (uint64_t *)
857 				    ((uintptr_t)txwr + wr_len);
858 				*pad = 0;
859 			}
860 		}
861 
862 		KASSERT(toep->tx_credits >= credits,
863 			("%s: not enough credits", __func__));
864 
865 		toep->tx_credits -= credits;
866 		toep->tx_nocompl += credits;
867 		toep->plen_nocompl += plen;
868 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
869 		    toep->tx_nocompl >= toep->tx_total / 4)
870 			compl = 1;
871 
872 		if (compl || ulp_mode(toep) == ULP_MODE_RDMA) {
873 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
874 			toep->tx_nocompl = 0;
875 			toep->plen_nocompl = 0;
876 		}
877 
878 		tp->snd_nxt += plen;
879 		tp->snd_max += plen;
880 
881 		SOCKBUF_LOCK(sb);
882 		KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
883 		sb->sb_sndptr = sb_sndptr;
884 		SOCKBUF_UNLOCK(sb);
885 
886 		toep->flags |= TPF_TX_DATA_SENT;
887 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
888 			toep->flags |= TPF_TX_SUSPENDED;
889 
890 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
891 		txsd->plen = plen;
892 		txsd->tx_credits = credits;
893 		txsd++;
894 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
895 			toep->txsd_pidx = 0;
896 			txsd = &toep->txsd[0];
897 		}
898 		toep->txsd_avail--;
899 
900 		t4_l2t_send(sc, wr, toep->l2te);
901 	} while (m != NULL && (m->m_flags & M_NOTAVAIL) == 0);
902 
903 	/* Send a FIN if requested, but only if there's no more data to send */
904 	if (m == NULL && toep->flags & TPF_SEND_FIN)
905 		t4_close_conn(sc, toep);
906 }
907 
908 static inline void
909 rqdrop_locked(struct mbufq *q, int plen)
910 {
911 	struct mbuf *m;
912 
913 	while (plen > 0) {
914 		m = mbufq_dequeue(q);
915 
916 		/* Too many credits. */
917 		MPASS(m != NULL);
918 		M_ASSERTPKTHDR(m);
919 
920 		/* Partial credits. */
921 		MPASS(plen >= m->m_pkthdr.len);
922 
923 		plen -= m->m_pkthdr.len;
924 		m_freem(m);
925 	}
926 }
927 
928 void
929 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
930 {
931 	struct mbuf *sndptr, *m;
932 	struct fw_ofld_tx_data_wr *txwr;
933 	struct wrqe *wr;
934 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
935 	u_int adjusted_plen, ulp_submode;
936 	struct inpcb *inp = toep->inp;
937 	struct tcpcb *tp = intotcpcb(inp);
938 	int tx_credits, shove;
939 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
940 	struct mbufq *pduq = &toep->ulp_pduq;
941 	static const u_int ulp_extra_len[] = {0, 4, 4, 8};
942 
943 	INP_WLOCK_ASSERT(inp);
944 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
945 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
946 	KASSERT(ulp_mode(toep) == ULP_MODE_ISCSI,
947 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
948 
949 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
950 		return;
951 
952 	/*
953 	 * This function doesn't resume by itself.  Someone else must clear the
954 	 * flag and call this function.
955 	 */
956 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
957 		KASSERT(drop == 0,
958 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
959 		return;
960 	}
961 
962 	if (drop)
963 		rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
964 
965 	while ((sndptr = mbufq_first(pduq)) != NULL) {
966 		M_ASSERTPKTHDR(sndptr);
967 
968 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
969 		max_imm = max_imm_payload(tx_credits);
970 		max_nsegs = max_dsgl_nsegs(tx_credits);
971 
972 		plen = 0;
973 		nsegs = 0;
974 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
975 		for (m = sndptr; m != NULL; m = m->m_next) {
976 			int n = sglist_count(mtod(m, void *), m->m_len);
977 
978 			nsegs += n;
979 			plen += m->m_len;
980 
981 			/*
982 			 * This mbuf would send us _over_ the nsegs limit.
983 			 * Suspend tx because the PDU can't be sent out.
984 			 */
985 			if (plen > max_imm && nsegs > max_nsegs) {
986 				toep->flags |= TPF_TX_SUSPENDED;
987 				return;
988 			}
989 
990 			if (max_nsegs_1mbuf < n)
991 				max_nsegs_1mbuf = n;
992 		}
993 
994 		if (__predict_false(toep->flags & TPF_FIN_SENT))
995 			panic("%s: excess tx.", __func__);
996 
997 		/*
998 		 * We have a PDU to send.  All of it goes out in one WR so 'm'
999 		 * is NULL.  A PDU's length is always a multiple of 4.
1000 		 */
1001 		MPASS(m == NULL);
1002 		MPASS((plen & 3) == 0);
1003 		MPASS(sndptr->m_pkthdr.len == plen);
1004 
1005 		shove = !(tp->t_flags & TF_MORETOCOME);
1006 		ulp_submode = mbuf_ulp_submode(sndptr);
1007 		MPASS(ulp_submode < nitems(ulp_extra_len));
1008 
1009 		/*
1010 		 * plen doesn't include header and data digests, which are
1011 		 * generated and inserted in the right places by the TOE, but
1012 		 * they do occupy TCP sequence space and need to be accounted
1013 		 * for.
1014 		 */
1015 		adjusted_plen = plen + ulp_extra_len[ulp_submode];
1016 		if (plen <= max_imm) {
1017 
1018 			/* Immediate data tx */
1019 
1020 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
1021 					toep->ofld_txq);
1022 			if (wr == NULL) {
1023 				/* XXX: how will we recover from this? */
1024 				toep->flags |= TPF_TX_SUSPENDED;
1025 				return;
1026 			}
1027 			txwr = wrtod(wr);
1028 			credits = howmany(wr->wr_len, 16);
1029 			write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
1030 			    shove, ulp_submode);
1031 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
1032 			nsegs = 0;
1033 		} else {
1034 			int wr_len;
1035 
1036 			/* DSGL tx */
1037 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
1038 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1039 			wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
1040 			if (wr == NULL) {
1041 				/* XXX: how will we recover from this? */
1042 				toep->flags |= TPF_TX_SUSPENDED;
1043 				return;
1044 			}
1045 			txwr = wrtod(wr);
1046 			credits = howmany(wr_len, 16);
1047 			write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
1048 			    shove, ulp_submode);
1049 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
1050 			    max_nsegs_1mbuf);
1051 			if (wr_len & 0xf) {
1052 				uint64_t *pad = (uint64_t *)
1053 				    ((uintptr_t)txwr + wr_len);
1054 				*pad = 0;
1055 			}
1056 		}
1057 
1058 		KASSERT(toep->tx_credits >= credits,
1059 			("%s: not enough credits", __func__));
1060 
1061 		m = mbufq_dequeue(pduq);
1062 		MPASS(m == sndptr);
1063 		mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
1064 
1065 		toep->tx_credits -= credits;
1066 		toep->tx_nocompl += credits;
1067 		toep->plen_nocompl += plen;
1068 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
1069 		    toep->tx_nocompl >= toep->tx_total / 4) {
1070 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
1071 			toep->tx_nocompl = 0;
1072 			toep->plen_nocompl = 0;
1073 		}
1074 
1075 		tp->snd_nxt += adjusted_plen;
1076 		tp->snd_max += adjusted_plen;
1077 
1078 		toep->flags |= TPF_TX_DATA_SENT;
1079 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
1080 			toep->flags |= TPF_TX_SUSPENDED;
1081 
1082 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1083 		txsd->plen = plen;
1084 		txsd->tx_credits = credits;
1085 		txsd++;
1086 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1087 			toep->txsd_pidx = 0;
1088 			txsd = &toep->txsd[0];
1089 		}
1090 		toep->txsd_avail--;
1091 
1092 		t4_l2t_send(sc, wr, toep->l2te);
1093 	}
1094 
1095 	/* Send a FIN if requested, but only if there are no more PDUs to send */
1096 	if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
1097 		t4_close_conn(sc, toep);
1098 }
1099 
1100 static inline void
1101 t4_push_data(struct adapter *sc, struct toepcb *toep, int drop)
1102 {
1103 
1104 	if (ulp_mode(toep) == ULP_MODE_ISCSI)
1105 		t4_push_pdus(sc, toep, drop);
1106 	else if (tls_tx_key(toep) && toep->tls.mode == TLS_MODE_TLSOM)
1107 		t4_push_tls_records(sc, toep, drop);
1108 #ifdef KERN_TLS
1109 	else if (toep->flags & TPF_KTLS)
1110 		t4_push_ktls(sc, toep, drop);
1111 #endif
1112 	else
1113 		t4_push_frames(sc, toep, drop);
1114 }
1115 
1116 int
1117 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
1118 {
1119 	struct adapter *sc = tod->tod_softc;
1120 #ifdef INVARIANTS
1121 	struct inpcb *inp = tp->t_inpcb;
1122 #endif
1123 	struct toepcb *toep = tp->t_toe;
1124 
1125 	INP_WLOCK_ASSERT(inp);
1126 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1127 	    ("%s: inp %p dropped.", __func__, inp));
1128 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1129 
1130 	t4_push_data(sc, toep, 0);
1131 
1132 	return (0);
1133 }
1134 
1135 int
1136 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
1137 {
1138 	struct adapter *sc = tod->tod_softc;
1139 #ifdef INVARIANTS
1140 	struct inpcb *inp = tp->t_inpcb;
1141 #endif
1142 	struct toepcb *toep = tp->t_toe;
1143 
1144 	INP_WLOCK_ASSERT(inp);
1145 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1146 	    ("%s: inp %p dropped.", __func__, inp));
1147 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1148 
1149 	toep->flags |= TPF_SEND_FIN;
1150 	if (tp->t_state >= TCPS_ESTABLISHED)
1151 		t4_push_data(sc, toep, 0);
1152 
1153 	return (0);
1154 }
1155 
1156 int
1157 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1158 {
1159 	struct adapter *sc = tod->tod_softc;
1160 #if defined(INVARIANTS)
1161 	struct inpcb *inp = tp->t_inpcb;
1162 #endif
1163 	struct toepcb *toep = tp->t_toe;
1164 
1165 	INP_WLOCK_ASSERT(inp);
1166 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1167 	    ("%s: inp %p dropped.", __func__, inp));
1168 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1169 
1170 	/* hmmmm */
1171 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1172 	    ("%s: flowc for tid %u [%s] not sent already",
1173 	    __func__, toep->tid, tcpstates[tp->t_state]));
1174 
1175 	send_reset(sc, toep, 0);
1176 	return (0);
1177 }
1178 
1179 /*
1180  * Peer has sent us a FIN.
1181  */
1182 static int
1183 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1184 {
1185 	struct adapter *sc = iq->adapter;
1186 	const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1187 	unsigned int tid = GET_TID(cpl);
1188 	struct toepcb *toep = lookup_tid(sc, tid);
1189 	struct inpcb *inp = toep->inp;
1190 	struct tcpcb *tp = NULL;
1191 	struct socket *so;
1192 	struct epoch_tracker et;
1193 #ifdef INVARIANTS
1194 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1195 #endif
1196 
1197 	KASSERT(opcode == CPL_PEER_CLOSE,
1198 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1199 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1200 
1201 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1202 		/*
1203 		 * do_pass_establish must have run before do_peer_close and if
1204 		 * this is still a synqe instead of a toepcb then the connection
1205 		 * must be getting aborted.
1206 		 */
1207 		MPASS(toep->flags & TPF_ABORT_SHUTDOWN);
1208 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1209 		    toep, toep->flags);
1210 		return (0);
1211 	}
1212 
1213 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1214 
1215 	CURVNET_SET(toep->vnet);
1216 	NET_EPOCH_ENTER(et);
1217 	INP_WLOCK(inp);
1218 	tp = intotcpcb(inp);
1219 
1220 	CTR6(KTR_CXGBE,
1221 	    "%s: tid %u (%s), toep_flags 0x%x, ddp_flags 0x%x, inp %p",
1222 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1223 	    toep->ddp.flags, inp);
1224 
1225 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1226 		goto done;
1227 
1228 	tp->rcv_nxt++;	/* FIN */
1229 
1230 	so = inp->inp_socket;
1231 	socantrcvmore(so);
1232 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1233 		DDP_LOCK(toep);
1234 		if (__predict_false(toep->ddp.flags &
1235 		    (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)))
1236 			handle_ddp_close(toep, tp, cpl->rcv_nxt);
1237 		DDP_UNLOCK(toep);
1238 	}
1239 
1240 	if (ulp_mode(toep) != ULP_MODE_RDMA) {
1241 		KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1242 	    		("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1243 	    		be32toh(cpl->rcv_nxt)));
1244 	}
1245 
1246 	switch (tp->t_state) {
1247 	case TCPS_SYN_RECEIVED:
1248 		tp->t_starttime = ticks;
1249 		/* FALLTHROUGH */
1250 
1251 	case TCPS_ESTABLISHED:
1252 		tcp_state_change(tp, TCPS_CLOSE_WAIT);
1253 		break;
1254 
1255 	case TCPS_FIN_WAIT_1:
1256 		tcp_state_change(tp, TCPS_CLOSING);
1257 		break;
1258 
1259 	case TCPS_FIN_WAIT_2:
1260 		tcp_twstart(tp);
1261 		INP_UNLOCK_ASSERT(inp);	 /* safe, we have a ref on the inp */
1262 		NET_EPOCH_EXIT(et);
1263 		CURVNET_RESTORE();
1264 
1265 		INP_WLOCK(inp);
1266 		final_cpl_received(toep);
1267 		return (0);
1268 
1269 	default:
1270 		log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1271 		    __func__, tid, tp->t_state);
1272 	}
1273 done:
1274 	INP_WUNLOCK(inp);
1275 	NET_EPOCH_EXIT(et);
1276 	CURVNET_RESTORE();
1277 	return (0);
1278 }
1279 
1280 /*
1281  * Peer has ACK'd our FIN.
1282  */
1283 static int
1284 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1285     struct mbuf *m)
1286 {
1287 	struct adapter *sc = iq->adapter;
1288 	const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1289 	unsigned int tid = GET_TID(cpl);
1290 	struct toepcb *toep = lookup_tid(sc, tid);
1291 	struct inpcb *inp = toep->inp;
1292 	struct tcpcb *tp = NULL;
1293 	struct socket *so = NULL;
1294 	struct epoch_tracker et;
1295 #ifdef INVARIANTS
1296 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1297 #endif
1298 
1299 	KASSERT(opcode == CPL_CLOSE_CON_RPL,
1300 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1301 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1302 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1303 
1304 	CURVNET_SET(toep->vnet);
1305 	NET_EPOCH_ENTER(et);
1306 	INP_WLOCK(inp);
1307 	tp = intotcpcb(inp);
1308 
1309 	CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1310 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1311 
1312 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1313 		goto done;
1314 
1315 	so = inp->inp_socket;
1316 	tp->snd_una = be32toh(cpl->snd_nxt) - 1;	/* exclude FIN */
1317 
1318 	switch (tp->t_state) {
1319 	case TCPS_CLOSING:	/* see TCPS_FIN_WAIT_2 in do_peer_close too */
1320 		tcp_twstart(tp);
1321 release:
1322 		INP_UNLOCK_ASSERT(inp);	/* safe, we have a ref on the  inp */
1323 		NET_EPOCH_EXIT(et);
1324 		CURVNET_RESTORE();
1325 
1326 		INP_WLOCK(inp);
1327 		final_cpl_received(toep);	/* no more CPLs expected */
1328 
1329 		return (0);
1330 	case TCPS_LAST_ACK:
1331 		if (tcp_close(tp))
1332 			INP_WUNLOCK(inp);
1333 		goto release;
1334 
1335 	case TCPS_FIN_WAIT_1:
1336 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1337 			soisdisconnected(so);
1338 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
1339 		break;
1340 
1341 	default:
1342 		log(LOG_ERR,
1343 		    "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1344 		    __func__, tid, tcpstates[tp->t_state]);
1345 	}
1346 done:
1347 	INP_WUNLOCK(inp);
1348 	NET_EPOCH_EXIT(et);
1349 	CURVNET_RESTORE();
1350 	return (0);
1351 }
1352 
1353 void
1354 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1355     int rst_status)
1356 {
1357 	struct wrqe *wr;
1358 	struct cpl_abort_rpl *cpl;
1359 
1360 	wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1361 	if (wr == NULL) {
1362 		/* XXX */
1363 		panic("%s: allocation failure.", __func__);
1364 	}
1365 	cpl = wrtod(wr);
1366 
1367 	INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1368 	cpl->cmd = rst_status;
1369 
1370 	t4_wrq_tx(sc, wr);
1371 }
1372 
1373 static int
1374 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1375 {
1376 	switch (abort_reason) {
1377 	case CPL_ERR_BAD_SYN:
1378 	case CPL_ERR_CONN_RESET:
1379 		return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1380 	case CPL_ERR_XMIT_TIMEDOUT:
1381 	case CPL_ERR_PERSIST_TIMEDOUT:
1382 	case CPL_ERR_FINWAIT2_TIMEDOUT:
1383 	case CPL_ERR_KEEPALIVE_TIMEDOUT:
1384 		return (ETIMEDOUT);
1385 	default:
1386 		return (EIO);
1387 	}
1388 }
1389 
1390 /*
1391  * TCP RST from the peer, timeout, or some other such critical error.
1392  */
1393 static int
1394 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1395 {
1396 	struct adapter *sc = iq->adapter;
1397 	const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1398 	unsigned int tid = GET_TID(cpl);
1399 	struct toepcb *toep = lookup_tid(sc, tid);
1400 	struct sge_wrq *ofld_txq = toep->ofld_txq;
1401 	struct inpcb *inp;
1402 	struct tcpcb *tp;
1403 	struct epoch_tracker et;
1404 #ifdef INVARIANTS
1405 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1406 #endif
1407 
1408 	KASSERT(opcode == CPL_ABORT_REQ_RSS,
1409 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1410 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1411 
1412 	if (toep->flags & TPF_SYNQE)
1413 		return (do_abort_req_synqe(iq, rss, m));
1414 
1415 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1416 
1417 	if (negative_advice(cpl->status)) {
1418 		CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1419 		    __func__, cpl->status, tid, toep->flags);
1420 		return (0);	/* Ignore negative advice */
1421 	}
1422 
1423 	inp = toep->inp;
1424 	CURVNET_SET(toep->vnet);
1425 	NET_EPOCH_ENTER(et);	/* for tcp_close */
1426 	INP_WLOCK(inp);
1427 
1428 	tp = intotcpcb(inp);
1429 
1430 	CTR6(KTR_CXGBE,
1431 	    "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1432 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1433 	    inp->inp_flags, cpl->status);
1434 
1435 	/*
1436 	 * If we'd initiated an abort earlier the reply to it is responsible for
1437 	 * cleaning up resources.  Otherwise we tear everything down right here
1438 	 * right now.  We owe the T4 a CPL_ABORT_RPL no matter what.
1439 	 */
1440 	if (toep->flags & TPF_ABORT_SHUTDOWN) {
1441 		INP_WUNLOCK(inp);
1442 		goto done;
1443 	}
1444 	toep->flags |= TPF_ABORT_SHUTDOWN;
1445 
1446 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1447 		struct socket *so = inp->inp_socket;
1448 
1449 		if (so != NULL)
1450 			so_error_set(so, abort_status_to_errno(tp,
1451 			    cpl->status));
1452 		tp = tcp_close(tp);
1453 		if (tp == NULL)
1454 			INP_WLOCK(inp);	/* re-acquire */
1455 	}
1456 
1457 	final_cpl_received(toep);
1458 done:
1459 	NET_EPOCH_EXIT(et);
1460 	CURVNET_RESTORE();
1461 	send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1462 	return (0);
1463 }
1464 
1465 /*
1466  * Reply to the CPL_ABORT_REQ (send_reset)
1467  */
1468 static int
1469 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1470 {
1471 	struct adapter *sc = iq->adapter;
1472 	const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1473 	unsigned int tid = GET_TID(cpl);
1474 	struct toepcb *toep = lookup_tid(sc, tid);
1475 	struct inpcb *inp = toep->inp;
1476 #ifdef INVARIANTS
1477 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1478 #endif
1479 
1480 	KASSERT(opcode == CPL_ABORT_RPL_RSS,
1481 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1482 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1483 
1484 	if (toep->flags & TPF_SYNQE)
1485 		return (do_abort_rpl_synqe(iq, rss, m));
1486 
1487 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1488 
1489 	CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1490 	    __func__, tid, toep, inp, cpl->status);
1491 
1492 	KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1493 	    ("%s: wasn't expecting abort reply", __func__));
1494 
1495 	INP_WLOCK(inp);
1496 	final_cpl_received(toep);
1497 
1498 	return (0);
1499 }
1500 
1501 static int
1502 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1503 {
1504 	struct adapter *sc = iq->adapter;
1505 	const struct cpl_rx_data *cpl = mtod(m, const void *);
1506 	unsigned int tid = GET_TID(cpl);
1507 	struct toepcb *toep = lookup_tid(sc, tid);
1508 	struct inpcb *inp = toep->inp;
1509 	struct tcpcb *tp;
1510 	struct socket *so;
1511 	struct sockbuf *sb;
1512 	struct epoch_tracker et;
1513 	int len, rx_credits;
1514 	uint32_t ddp_placed = 0;
1515 
1516 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1517 		/*
1518 		 * do_pass_establish must have run before do_rx_data and if this
1519 		 * is still a synqe instead of a toepcb then the connection must
1520 		 * be getting aborted.
1521 		 */
1522 		MPASS(toep->flags & TPF_ABORT_SHUTDOWN);
1523 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1524 		    toep, toep->flags);
1525 		m_freem(m);
1526 		return (0);
1527 	}
1528 
1529 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1530 
1531 	/* strip off CPL header */
1532 	m_adj(m, sizeof(*cpl));
1533 	len = m->m_pkthdr.len;
1534 
1535 	INP_WLOCK(inp);
1536 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1537 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1538 		    __func__, tid, len, inp->inp_flags);
1539 		INP_WUNLOCK(inp);
1540 		m_freem(m);
1541 		return (0);
1542 	}
1543 
1544 	tp = intotcpcb(inp);
1545 
1546 	if (__predict_false(ulp_mode(toep) == ULP_MODE_TLS &&
1547 	   toep->flags & TPF_TLS_RECEIVE)) {
1548 		/* Received "raw" data on a TLS socket. */
1549 		CTR3(KTR_CXGBE, "%s: tid %u, raw TLS data (%d bytes)",
1550 		    __func__, tid, len);
1551 		do_rx_data_tls(cpl, toep, m);
1552 		return (0);
1553 	}
1554 
1555 	if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1556 		ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1557 
1558 	tp->rcv_nxt += len;
1559 	if (tp->rcv_wnd < len) {
1560 		KASSERT(ulp_mode(toep) == ULP_MODE_RDMA,
1561 				("%s: negative window size", __func__));
1562 	}
1563 
1564 	tp->rcv_wnd -= len;
1565 	tp->t_rcvtime = ticks;
1566 
1567 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1568 		DDP_LOCK(toep);
1569 	so = inp_inpcbtosocket(inp);
1570 	sb = &so->so_rcv;
1571 	SOCKBUF_LOCK(sb);
1572 
1573 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1574 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1575 		    __func__, tid, len);
1576 		m_freem(m);
1577 		SOCKBUF_UNLOCK(sb);
1578 		if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1579 			DDP_UNLOCK(toep);
1580 		INP_WUNLOCK(inp);
1581 
1582 		CURVNET_SET(toep->vnet);
1583 		NET_EPOCH_ENTER(et);
1584 		INP_WLOCK(inp);
1585 		tp = tcp_drop(tp, ECONNRESET);
1586 		if (tp)
1587 			INP_WUNLOCK(inp);
1588 		NET_EPOCH_EXIT(et);
1589 		CURVNET_RESTORE();
1590 
1591 		return (0);
1592 	}
1593 
1594 	/* receive buffer autosize */
1595 	MPASS(toep->vnet == so->so_vnet);
1596 	CURVNET_SET(toep->vnet);
1597 	if (sb->sb_flags & SB_AUTOSIZE &&
1598 	    V_tcp_do_autorcvbuf &&
1599 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1600 	    len > (sbspace(sb) / 8 * 7)) {
1601 		unsigned int hiwat = sb->sb_hiwat;
1602 		unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
1603 		    V_tcp_autorcvbuf_max);
1604 
1605 		if (!sbreserve_locked(sb, newsize, so, NULL))
1606 			sb->sb_flags &= ~SB_AUTOSIZE;
1607 	}
1608 
1609 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1610 		int changed = !(toep->ddp.flags & DDP_ON) ^ cpl->ddp_off;
1611 
1612 		if (toep->ddp.waiting_count != 0 || toep->ddp.active_count != 0)
1613 			CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)",
1614 			    __func__, tid, len);
1615 
1616 		if (changed) {
1617 			if (toep->ddp.flags & DDP_SC_REQ)
1618 				toep->ddp.flags ^= DDP_ON | DDP_SC_REQ;
1619 			else {
1620 				KASSERT(cpl->ddp_off == 1,
1621 				    ("%s: DDP switched on by itself.",
1622 				    __func__));
1623 
1624 				/* Fell out of DDP mode */
1625 				toep->ddp.flags &= ~DDP_ON;
1626 				CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
1627 				    __func__);
1628 
1629 				insert_ddp_data(toep, ddp_placed);
1630 			}
1631 		}
1632 
1633 		if (toep->ddp.flags & DDP_ON) {
1634 			/*
1635 			 * CPL_RX_DATA with DDP on can only be an indicate.
1636 			 * Start posting queued AIO requests via DDP.  The
1637 			 * payload that arrived in this indicate is appended
1638 			 * to the socket buffer as usual.
1639 			 */
1640 			handle_ddp_indicate(toep);
1641 		}
1642 	}
1643 
1644 	sbappendstream_locked(sb, m, 0);
1645 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
1646 	if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) {
1647 		rx_credits = send_rx_credits(sc, toep, rx_credits);
1648 		tp->rcv_wnd += rx_credits;
1649 		tp->rcv_adv += rx_credits;
1650 	}
1651 
1652 	if (ulp_mode(toep) == ULP_MODE_TCPDDP && toep->ddp.waiting_count > 0 &&
1653 	    sbavail(sb) != 0) {
1654 		CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__,
1655 		    tid);
1656 		ddp_queue_toep(toep);
1657 	}
1658 	sorwakeup_locked(so);
1659 	SOCKBUF_UNLOCK_ASSERT(sb);
1660 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1661 		DDP_UNLOCK(toep);
1662 
1663 	INP_WUNLOCK(inp);
1664 	CURVNET_RESTORE();
1665 	return (0);
1666 }
1667 
1668 static int
1669 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1670 {
1671 	struct adapter *sc = iq->adapter;
1672 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1673 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1674 	struct toepcb *toep = lookup_tid(sc, tid);
1675 	struct inpcb *inp;
1676 	struct tcpcb *tp;
1677 	struct socket *so;
1678 	uint8_t credits = cpl->credits;
1679 	struct ofld_tx_sdesc *txsd;
1680 	int plen;
1681 #ifdef INVARIANTS
1682 	unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1683 #endif
1684 
1685 	/*
1686 	 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1687 	 * now this comes back carrying the credits for the flowc.
1688 	 */
1689 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1690 		KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1691 		    ("%s: credits for a synq entry %p", __func__, toep));
1692 		return (0);
1693 	}
1694 
1695 	inp = toep->inp;
1696 
1697 	KASSERT(opcode == CPL_FW4_ACK,
1698 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1699 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1700 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1701 
1702 	INP_WLOCK(inp);
1703 
1704 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1705 		INP_WUNLOCK(inp);
1706 		return (0);
1707 	}
1708 
1709 	KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1710 	    ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1711 
1712 	tp = intotcpcb(inp);
1713 
1714 	if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1715 		tcp_seq snd_una = be32toh(cpl->snd_una);
1716 
1717 #ifdef INVARIANTS
1718 		if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1719 			log(LOG_ERR,
1720 			    "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1721 			    __func__, snd_una, toep->tid, tp->snd_una);
1722 		}
1723 #endif
1724 
1725 		if (tp->snd_una != snd_una) {
1726 			tp->snd_una = snd_una;
1727 			tp->ts_recent_age = tcp_ts_getticks();
1728 		}
1729 	}
1730 
1731 #ifdef VERBOSE_TRACES
1732 	CTR3(KTR_CXGBE, "%s: tid %d credits %u", __func__, tid, credits);
1733 #endif
1734 	so = inp->inp_socket;
1735 	txsd = &toep->txsd[toep->txsd_cidx];
1736 	plen = 0;
1737 	while (credits) {
1738 		KASSERT(credits >= txsd->tx_credits,
1739 		    ("%s: too many (or partial) credits", __func__));
1740 		credits -= txsd->tx_credits;
1741 		toep->tx_credits += txsd->tx_credits;
1742 		plen += txsd->plen;
1743 		if (txsd->iv_buffer) {
1744 			free(txsd->iv_buffer, M_CXGBE);
1745 			txsd->iv_buffer = NULL;
1746 		}
1747 		txsd++;
1748 		toep->txsd_avail++;
1749 		KASSERT(toep->txsd_avail <= toep->txsd_total,
1750 		    ("%s: txsd avail > total", __func__));
1751 		if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1752 			txsd = &toep->txsd[0];
1753 			toep->txsd_cidx = 0;
1754 		}
1755 	}
1756 
1757 	if (toep->tx_credits == toep->tx_total) {
1758 		toep->tx_nocompl = 0;
1759 		toep->plen_nocompl = 0;
1760 	}
1761 
1762 	if (toep->flags & TPF_TX_SUSPENDED &&
1763 	    toep->tx_credits >= toep->tx_total / 4) {
1764 #ifdef VERBOSE_TRACES
1765 		CTR2(KTR_CXGBE, "%s: tid %d calling t4_push_frames", __func__,
1766 		    tid);
1767 #endif
1768 		toep->flags &= ~TPF_TX_SUSPENDED;
1769 		CURVNET_SET(toep->vnet);
1770 		t4_push_data(sc, toep, plen);
1771 		CURVNET_RESTORE();
1772 	} else if (plen > 0) {
1773 		struct sockbuf *sb = &so->so_snd;
1774 		int sbu;
1775 
1776 		SOCKBUF_LOCK(sb);
1777 		sbu = sbused(sb);
1778 		if (ulp_mode(toep) == ULP_MODE_ISCSI) {
1779 
1780 			if (__predict_false(sbu > 0)) {
1781 				/*
1782 				 * The data trasmitted before the tid's ULP mode
1783 				 * changed to ISCSI is still in so_snd.
1784 				 * Incoming credits should account for so_snd
1785 				 * first.
1786 				 */
1787 				sbdrop_locked(sb, min(sbu, plen));
1788 				plen -= min(sbu, plen);
1789 			}
1790 			sowwakeup_locked(so);	/* unlocks so_snd */
1791 			rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1792 		} else {
1793 #ifdef VERBOSE_TRACES
1794 			CTR3(KTR_CXGBE, "%s: tid %d dropped %d bytes", __func__,
1795 			    tid, plen);
1796 #endif
1797 			sbdrop_locked(sb, plen);
1798 			if (tls_tx_key(toep) &&
1799 			    toep->tls.mode == TLS_MODE_TLSOM) {
1800 				struct tls_ofld_info *tls_ofld = &toep->tls;
1801 
1802 				MPASS(tls_ofld->sb_off >= plen);
1803 				tls_ofld->sb_off -= plen;
1804 			}
1805 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
1806 				t4_aiotx_queue_toep(so, toep);
1807 			sowwakeup_locked(so);	/* unlocks so_snd */
1808 		}
1809 		SOCKBUF_UNLOCK_ASSERT(sb);
1810 	}
1811 
1812 	INP_WUNLOCK(inp);
1813 
1814 	return (0);
1815 }
1816 
1817 void
1818 t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, struct toepcb *toep,
1819     uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie)
1820 {
1821 	struct wrqe *wr;
1822 	struct cpl_set_tcb_field *req;
1823 	struct ofld_tx_sdesc *txsd;
1824 
1825 	MPASS((cookie & ~M_COOKIE) == 0);
1826 	if (reply) {
1827 		MPASS(cookie != CPL_COOKIE_RESERVED);
1828 	}
1829 
1830 	wr = alloc_wrqe(sizeof(*req), wrq);
1831 	if (wr == NULL) {
1832 		/* XXX */
1833 		panic("%s: allocation failure.", __func__);
1834 	}
1835 	req = wrtod(wr);
1836 
1837 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1838 	req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1839 	if (reply == 0)
1840 		req->reply_ctrl |= htobe16(F_NO_REPLY);
1841 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie));
1842 	req->mask = htobe64(mask);
1843 	req->val = htobe64(val);
1844 	if ((wrq->eq.flags & EQ_TYPEMASK) == EQ_OFLD) {
1845 		txsd = &toep->txsd[toep->txsd_pidx];
1846 		txsd->tx_credits = howmany(sizeof(*req), 16);
1847 		txsd->plen = 0;
1848 		KASSERT(toep->tx_credits >= txsd->tx_credits &&
1849 		    toep->txsd_avail > 0,
1850 		    ("%s: not enough credits (%d)", __func__,
1851 		    toep->tx_credits));
1852 		toep->tx_credits -= txsd->tx_credits;
1853 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1854 			toep->txsd_pidx = 0;
1855 		toep->txsd_avail--;
1856 	}
1857 
1858 	t4_wrq_tx(sc, wr);
1859 }
1860 
1861 void
1862 t4_init_cpl_io_handlers(void)
1863 {
1864 
1865 	t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close);
1866 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl);
1867 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req);
1868 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl,
1869 	    CPL_COOKIE_TOM);
1870 	t4_register_cpl_handler(CPL_RX_DATA, do_rx_data);
1871 	t4_register_shared_cpl_handler(CPL_FW4_ACK, do_fw4_ack, CPL_COOKIE_TOM);
1872 }
1873 
1874 void
1875 t4_uninit_cpl_io_handlers(void)
1876 {
1877 
1878 	t4_register_cpl_handler(CPL_PEER_CLOSE, NULL);
1879 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, NULL);
1880 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, NULL);
1881 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, NULL, CPL_COOKIE_TOM);
1882 	t4_register_cpl_handler(CPL_RX_DATA, NULL);
1883 	t4_register_shared_cpl_handler(CPL_FW4_ACK, NULL, CPL_COOKIE_TOM);
1884 }
1885 
1886 /*
1887  * Use the 'backend1' field in AIO jobs to hold an error that should
1888  * be reported when the job is completed, the 'backend3' field to
1889  * store the amount of data sent by the AIO job so far, and the
1890  * 'backend4' field to hold a reference count on the job.
1891  *
1892  * Each unmapped mbuf holds a reference on the job as does the queue
1893  * so long as the job is queued.
1894  */
1895 #define	aio_error	backend1
1896 #define	aio_sent	backend3
1897 #define	aio_refs	backend4
1898 
1899 #define	jobtotid(job)							\
1900 	(((struct toepcb *)(so_sototcpcb((job)->fd_file->f_data)->t_toe))->tid)
1901 
1902 static void
1903 aiotx_free_job(struct kaiocb *job)
1904 {
1905 	long status;
1906 	int error;
1907 
1908 	if (refcount_release(&job->aio_refs) == 0)
1909 		return;
1910 
1911 	error = (intptr_t)job->aio_error;
1912 	status = job->aio_sent;
1913 #ifdef VERBOSE_TRACES
1914 	CTR5(KTR_CXGBE, "%s: tid %d completed %p len %ld, error %d", __func__,
1915 	    jobtotid(job), job, status, error);
1916 #endif
1917 	if (error != 0 && status != 0)
1918 		error = 0;
1919 	if (error == ECANCELED)
1920 		aio_cancel(job);
1921 	else if (error)
1922 		aio_complete(job, -1, error);
1923 	else {
1924 		job->msgsnd = 1;
1925 		aio_complete(job, status, 0);
1926 	}
1927 }
1928 
1929 static void
1930 aiotx_free_pgs(struct mbuf *m)
1931 {
1932 	struct kaiocb *job;
1933 	vm_page_t pg;
1934 
1935 	M_ASSERTEXTPG(m);
1936 	job = m->m_ext.ext_arg1;
1937 #ifdef VERBOSE_TRACES
1938 	CTR3(KTR_CXGBE, "%s: completed %d bytes for tid %d", __func__,
1939 	    m->m_len, jobtotid(job));
1940 #endif
1941 
1942 	for (int i = 0; i < m->m_epg_npgs; i++) {
1943 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
1944 		vm_page_unwire(pg, PQ_ACTIVE);
1945 	}
1946 
1947 	aiotx_free_job(job);
1948 }
1949 
1950 /*
1951  * Allocate a chain of unmapped mbufs describing the next 'len' bytes
1952  * of an AIO job.
1953  */
1954 static struct mbuf *
1955 alloc_aiotx_mbuf(struct kaiocb *job, int len)
1956 {
1957 	struct vmspace *vm;
1958 	vm_page_t pgs[MBUF_PEXT_MAX_PGS];
1959 	struct mbuf *m, *top, *last;
1960 	vm_map_t map;
1961 	vm_offset_t start;
1962 	int i, mlen, npages, pgoff;
1963 
1964 	KASSERT(job->aio_sent + len <= job->uaiocb.aio_nbytes,
1965 	    ("%s(%p, %d): request to send beyond end of buffer", __func__,
1966 	    job, len));
1967 
1968 	/*
1969 	 * The AIO subsystem will cancel and drain all requests before
1970 	 * permitting a process to exit or exec, so p_vmspace should
1971 	 * be stable here.
1972 	 */
1973 	vm = job->userproc->p_vmspace;
1974 	map = &vm->vm_map;
1975 	start = (uintptr_t)job->uaiocb.aio_buf + job->aio_sent;
1976 	pgoff = start & PAGE_MASK;
1977 
1978 	top = NULL;
1979 	last = NULL;
1980 	while (len > 0) {
1981 		mlen = imin(len, MBUF_PEXT_MAX_PGS * PAGE_SIZE - pgoff);
1982 		KASSERT(mlen == len || ((start + mlen) & PAGE_MASK) == 0,
1983 		    ("%s: next start (%#jx + %#x) is not page aligned",
1984 		    __func__, (uintmax_t)start, mlen));
1985 
1986 		npages = vm_fault_quick_hold_pages(map, start, mlen,
1987 		    VM_PROT_WRITE, pgs, nitems(pgs));
1988 		if (npages < 0)
1989 			break;
1990 
1991 		m = mb_alloc_ext_pgs(M_WAITOK, aiotx_free_pgs);
1992 		if (m == NULL) {
1993 			vm_page_unhold_pages(pgs, npages);
1994 			break;
1995 		}
1996 
1997 		m->m_epg_1st_off = pgoff;
1998 		m->m_epg_npgs = npages;
1999 		if (npages == 1) {
2000 			KASSERT(mlen + pgoff <= PAGE_SIZE,
2001 			    ("%s: single page is too large (off %d len %d)",
2002 			    __func__, pgoff, mlen));
2003 			m->m_epg_last_len = mlen;
2004 		} else {
2005 			m->m_epg_last_len = mlen - (PAGE_SIZE - pgoff) -
2006 			    (npages - 2) * PAGE_SIZE;
2007 		}
2008 		for (i = 0; i < npages; i++)
2009 			m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pgs[i]);
2010 
2011 		m->m_len = mlen;
2012 		m->m_ext.ext_size = npages * PAGE_SIZE;
2013 		m->m_ext.ext_arg1 = job;
2014 		refcount_acquire(&job->aio_refs);
2015 
2016 #ifdef VERBOSE_TRACES
2017 		CTR5(KTR_CXGBE, "%s: tid %d, new mbuf %p for job %p, npages %d",
2018 		    __func__, jobtotid(job), m, job, npages);
2019 #endif
2020 
2021 		if (top == NULL)
2022 			top = m;
2023 		else
2024 			last->m_next = m;
2025 		last = m;
2026 
2027 		len -= mlen;
2028 		start += mlen;
2029 		pgoff = 0;
2030 	}
2031 
2032 	return (top);
2033 }
2034 
2035 static void
2036 t4_aiotx_process_job(struct toepcb *toep, struct socket *so, struct kaiocb *job)
2037 {
2038 	struct sockbuf *sb;
2039 	struct file *fp;
2040 	struct inpcb *inp;
2041 	struct tcpcb *tp;
2042 	struct mbuf *m;
2043 	int error, len;
2044 	bool moretocome, sendmore;
2045 
2046 	sb = &so->so_snd;
2047 	SOCKBUF_UNLOCK(sb);
2048 	fp = job->fd_file;
2049 	m = NULL;
2050 
2051 #ifdef MAC
2052 	error = mac_socket_check_send(fp->f_cred, so);
2053 	if (error != 0)
2054 		goto out;
2055 #endif
2056 
2057 	/* Inline sosend_generic(). */
2058 
2059 	error = sblock(sb, SBL_WAIT);
2060 	MPASS(error == 0);
2061 
2062 sendanother:
2063 	SOCKBUF_LOCK(sb);
2064 	if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
2065 		SOCKBUF_UNLOCK(sb);
2066 		sbunlock(sb);
2067 		if ((so->so_options & SO_NOSIGPIPE) == 0) {
2068 			PROC_LOCK(job->userproc);
2069 			kern_psignal(job->userproc, SIGPIPE);
2070 			PROC_UNLOCK(job->userproc);
2071 		}
2072 		error = EPIPE;
2073 		goto out;
2074 	}
2075 	if (so->so_error) {
2076 		error = so->so_error;
2077 		so->so_error = 0;
2078 		SOCKBUF_UNLOCK(sb);
2079 		sbunlock(sb);
2080 		goto out;
2081 	}
2082 	if ((so->so_state & SS_ISCONNECTED) == 0) {
2083 		SOCKBUF_UNLOCK(sb);
2084 		sbunlock(sb);
2085 		error = ENOTCONN;
2086 		goto out;
2087 	}
2088 	if (sbspace(sb) < sb->sb_lowat) {
2089 		MPASS(job->aio_sent == 0 || !(so->so_state & SS_NBIO));
2090 
2091 		/*
2092 		 * Don't block if there is too little room in the socket
2093 		 * buffer.  Instead, requeue the request.
2094 		 */
2095 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2096 			SOCKBUF_UNLOCK(sb);
2097 			sbunlock(sb);
2098 			error = ECANCELED;
2099 			goto out;
2100 		}
2101 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2102 		SOCKBUF_UNLOCK(sb);
2103 		sbunlock(sb);
2104 		goto out;
2105 	}
2106 
2107 	/*
2108 	 * Write as much data as the socket permits, but no more than a
2109 	 * a single sndbuf at a time.
2110 	 */
2111 	len = sbspace(sb);
2112 	if (len > job->uaiocb.aio_nbytes - job->aio_sent) {
2113 		len = job->uaiocb.aio_nbytes - job->aio_sent;
2114 		moretocome = false;
2115 	} else
2116 		moretocome = true;
2117 	if (len > toep->params.sndbuf) {
2118 		len = toep->params.sndbuf;
2119 		sendmore = true;
2120 	} else
2121 		sendmore = false;
2122 
2123 	if (!TAILQ_EMPTY(&toep->aiotx_jobq))
2124 		moretocome = true;
2125 	SOCKBUF_UNLOCK(sb);
2126 	MPASS(len != 0);
2127 
2128 	m = alloc_aiotx_mbuf(job, len);
2129 	if (m == NULL) {
2130 		sbunlock(sb);
2131 		error = EFAULT;
2132 		goto out;
2133 	}
2134 
2135 	/* Inlined tcp_usr_send(). */
2136 
2137 	inp = toep->inp;
2138 	INP_WLOCK(inp);
2139 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
2140 		INP_WUNLOCK(inp);
2141 		sbunlock(sb);
2142 		error = ECONNRESET;
2143 		goto out;
2144 	}
2145 
2146 	job->aio_sent += m_length(m, NULL);
2147 
2148 	sbappendstream(sb, m, 0);
2149 	m = NULL;
2150 
2151 	if (!(inp->inp_flags & INP_DROPPED)) {
2152 		tp = intotcpcb(inp);
2153 		if (moretocome)
2154 			tp->t_flags |= TF_MORETOCOME;
2155 		error = tp->t_fb->tfb_tcp_output(tp);
2156 		if (moretocome)
2157 			tp->t_flags &= ~TF_MORETOCOME;
2158 	}
2159 
2160 	INP_WUNLOCK(inp);
2161 	if (sendmore)
2162 		goto sendanother;
2163 	sbunlock(sb);
2164 
2165 	if (error)
2166 		goto out;
2167 
2168 	/*
2169 	 * If this is a blocking socket and the request has not been
2170 	 * fully completed, requeue it until the socket is ready
2171 	 * again.
2172 	 */
2173 	if (job->aio_sent < job->uaiocb.aio_nbytes &&
2174 	    !(so->so_state & SS_NBIO)) {
2175 		SOCKBUF_LOCK(sb);
2176 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2177 			SOCKBUF_UNLOCK(sb);
2178 			error = ECANCELED;
2179 			goto out;
2180 		}
2181 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2182 		return;
2183 	}
2184 
2185 	/*
2186 	 * If the request will not be requeued, drop the queue's
2187 	 * reference to the job.  Any mbufs in flight should still
2188 	 * hold a reference, but this drops the reference that the
2189 	 * queue owns while it is waiting to queue mbufs to the
2190 	 * socket.
2191 	 */
2192 	aiotx_free_job(job);
2193 
2194 out:
2195 	if (error) {
2196 		job->aio_error = (void *)(intptr_t)error;
2197 		aiotx_free_job(job);
2198 	}
2199 	if (m != NULL)
2200 		m_free(m);
2201 	SOCKBUF_LOCK(sb);
2202 }
2203 
2204 static void
2205 t4_aiotx_task(void *context, int pending)
2206 {
2207 	struct toepcb *toep = context;
2208 	struct socket *so;
2209 	struct kaiocb *job;
2210 
2211 	so = toep->aiotx_so;
2212 	CURVNET_SET(toep->vnet);
2213 	SOCKBUF_LOCK(&so->so_snd);
2214 	while (!TAILQ_EMPTY(&toep->aiotx_jobq) && sowriteable(so)) {
2215 		job = TAILQ_FIRST(&toep->aiotx_jobq);
2216 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2217 		if (!aio_clear_cancel_function(job))
2218 			continue;
2219 
2220 		t4_aiotx_process_job(toep, so, job);
2221 	}
2222 	toep->aiotx_so = NULL;
2223 	SOCKBUF_UNLOCK(&so->so_snd);
2224 	CURVNET_RESTORE();
2225 
2226 	free_toepcb(toep);
2227 	SOCK_LOCK(so);
2228 	sorele(so);
2229 }
2230 
2231 static void
2232 t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep)
2233 {
2234 
2235 	SOCKBUF_LOCK_ASSERT(&toep->inp->inp_socket->so_snd);
2236 #ifdef VERBOSE_TRACES
2237 	CTR3(KTR_CXGBE, "%s: queueing aiotx task for tid %d, active = %s",
2238 	    __func__, toep->tid, toep->aiotx_so != NULL ? "true" : "false");
2239 #endif
2240 	if (toep->aiotx_so != NULL)
2241 		return;
2242 	soref(so);
2243 	toep->aiotx_so = so;
2244 	hold_toepcb(toep);
2245 	soaio_enqueue(&toep->aiotx_task);
2246 }
2247 
2248 static void
2249 t4_aiotx_cancel(struct kaiocb *job)
2250 {
2251 	struct socket *so;
2252 	struct sockbuf *sb;
2253 	struct tcpcb *tp;
2254 	struct toepcb *toep;
2255 
2256 	so = job->fd_file->f_data;
2257 	tp = so_sototcpcb(so);
2258 	toep = tp->t_toe;
2259 	MPASS(job->uaiocb.aio_lio_opcode == LIO_WRITE);
2260 	sb = &so->so_snd;
2261 
2262 	SOCKBUF_LOCK(sb);
2263 	if (!aio_cancel_cleared(job))
2264 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2265 	SOCKBUF_UNLOCK(sb);
2266 
2267 	job->aio_error = (void *)(intptr_t)ECANCELED;
2268 	aiotx_free_job(job);
2269 }
2270 
2271 int
2272 t4_aio_queue_aiotx(struct socket *so, struct kaiocb *job)
2273 {
2274 	struct tcpcb *tp = so_sototcpcb(so);
2275 	struct toepcb *toep = tp->t_toe;
2276 	struct adapter *sc = td_adapter(toep->td);
2277 
2278 	/* This only handles writes. */
2279 	if (job->uaiocb.aio_lio_opcode != LIO_WRITE)
2280 		return (EOPNOTSUPP);
2281 
2282 	if (!sc->tt.tx_zcopy)
2283 		return (EOPNOTSUPP);
2284 
2285 	if (tls_tx_key(toep))
2286 		return (EOPNOTSUPP);
2287 
2288 	SOCKBUF_LOCK(&so->so_snd);
2289 #ifdef VERBOSE_TRACES
2290 	CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid);
2291 #endif
2292 	if (!aio_set_cancel_function(job, t4_aiotx_cancel))
2293 		panic("new job was cancelled");
2294 	refcount_init(&job->aio_refs, 1);
2295 	TAILQ_INSERT_TAIL(&toep->aiotx_jobq, job, list);
2296 	if (sowriteable(so))
2297 		t4_aiotx_queue_toep(so, toep);
2298 	SOCKBUF_UNLOCK(&so->so_snd);
2299 	return (0);
2300 }
2301 
2302 void
2303 aiotx_init_toep(struct toepcb *toep)
2304 {
2305 
2306 	TAILQ_INIT(&toep->aiotx_jobq);
2307 	TASK_INIT(&toep->aiotx_task, 0, t4_aiotx_task, toep);
2308 }
2309 #endif
2310