xref: /freebsd/sys/dev/cxgbe/tom/t4_cpl_io.c (revision 3c5ba95ad12285ad37c182a4bfc1b240ec6d18a7)
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_ratelimit.h"
36 
37 #ifdef TCP_OFFLOAD
38 #include <sys/param.h>
39 #include <sys/aio.h>
40 #include <sys/file.h>
41 #include <sys/kernel.h>
42 #include <sys/ktr.h>
43 #include <sys/module.h>
44 #include <sys/proc.h>
45 #include <sys/protosw.h>
46 #include <sys/domain.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/sglist.h>
50 #include <sys/taskqueue.h>
51 #include <netinet/in.h>
52 #include <netinet/in_pcb.h>
53 #include <netinet/ip.h>
54 #include <netinet/ip6.h>
55 #define TCPSTATES
56 #include <netinet/tcp_fsm.h>
57 #include <netinet/tcp_seq.h>
58 #include <netinet/tcp_var.h>
59 #include <netinet/toecore.h>
60 
61 #include <security/mac/mac_framework.h>
62 
63 #include <vm/vm.h>
64 #include <vm/vm_extern.h>
65 #include <vm/pmap.h>
66 #include <vm/vm_map.h>
67 #include <vm/vm_page.h>
68 
69 #include "common/common.h"
70 #include "common/t4_msg.h"
71 #include "common/t4_regs.h"
72 #include "common/t4_tcb.h"
73 #include "tom/t4_tom_l2t.h"
74 #include "tom/t4_tom.h"
75 
76 static void	t4_aiotx_cancel(struct kaiocb *job);
77 static void	t4_aiotx_queue_toep(struct toepcb *toep);
78 
79 static size_t
80 aiotx_mbuf_pgoff(struct mbuf *m)
81 {
82 	struct aiotx_buffer *ab;
83 
84 	MPASS(IS_AIOTX_MBUF(m));
85 	ab = m->m_ext.ext_arg1;
86 	return ((ab->ps.offset + (uintptr_t)m->m_ext.ext_arg2) % PAGE_SIZE);
87 }
88 
89 static vm_page_t *
90 aiotx_mbuf_pages(struct mbuf *m)
91 {
92 	struct aiotx_buffer *ab;
93 	int npages;
94 
95 	MPASS(IS_AIOTX_MBUF(m));
96 	ab = m->m_ext.ext_arg1;
97 	npages = (ab->ps.offset + (uintptr_t)m->m_ext.ext_arg2) / PAGE_SIZE;
98 	return (ab->ps.pages + npages);
99 }
100 
101 void
102 send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp)
103 {
104 	struct wrqe *wr;
105 	struct fw_flowc_wr *flowc;
106 	unsigned int nparams, flowclen, paramidx;
107 	struct vi_info *vi = toep->vi;
108 	struct port_info *pi = vi->pi;
109 	struct adapter *sc = pi->adapter;
110 	unsigned int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN;
111 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
112 
113 	KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
114 	    ("%s: flowc for tid %u sent already", __func__, toep->tid));
115 
116 	if (ftxp != NULL)
117 		nparams = 8;
118 	else
119 		nparams = 6;
120 	if (toep->ulp_mode == ULP_MODE_TLS)
121 		nparams++;
122 	if (toep->tls.fcplenmax != 0)
123 		nparams++;
124 	if (toep->tc_idx != -1) {
125 		MPASS(toep->tc_idx >= 0 &&
126 		    toep->tc_idx < sc->chip_params->nsched_cls);
127 		nparams++;
128 	}
129 
130 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
131 
132 	wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
133 	if (wr == NULL) {
134 		/* XXX */
135 		panic("%s: allocation failure.", __func__);
136 	}
137 	flowc = wrtod(wr);
138 	memset(flowc, 0, wr->wr_len);
139 
140 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
141 	    V_FW_FLOWC_WR_NPARAMS(nparams));
142 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
143 	    V_FW_WR_FLOWID(toep->tid));
144 
145 #define FLOWC_PARAM(__m, __v) \
146 	do { \
147 		flowc->mnemval[paramidx].mnemonic = FW_FLOWC_MNEM_##__m; \
148 		flowc->mnemval[paramidx].val = htobe32(__v); \
149 		paramidx++; \
150 	} while (0)
151 
152 	paramidx = 0;
153 
154 	FLOWC_PARAM(PFNVFN, pfvf);
155 	FLOWC_PARAM(CH, pi->tx_chan);
156 	FLOWC_PARAM(PORT, pi->tx_chan);
157 	FLOWC_PARAM(IQID, toep->ofld_rxq->iq.abs_id);
158 	if (ftxp) {
159 		uint32_t sndbuf = min(ftxp->snd_space, sc->tt.sndbuf);
160 
161 		FLOWC_PARAM(SNDNXT, ftxp->snd_nxt);
162 		FLOWC_PARAM(RCVNXT, ftxp->rcv_nxt);
163 		FLOWC_PARAM(SNDBUF, sndbuf);
164 		FLOWC_PARAM(MSS, ftxp->mss);
165 
166 		CTR6(KTR_CXGBE,
167 		    "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x",
168 		    __func__, toep->tid, ftxp->mss, sndbuf, ftxp->snd_nxt,
169 		    ftxp->rcv_nxt);
170 	} else {
171 		FLOWC_PARAM(SNDBUF, 512);
172 		FLOWC_PARAM(MSS, 512);
173 
174 		CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid);
175 	}
176 	if (toep->ulp_mode == ULP_MODE_TLS)
177 		FLOWC_PARAM(ULP_MODE, toep->ulp_mode);
178 	if (toep->tls.fcplenmax != 0)
179 		FLOWC_PARAM(TXDATAPLEN_MAX, toep->tls.fcplenmax);
180 	if (toep->tc_idx != -1)
181 		FLOWC_PARAM(SCHEDCLASS, toep->tc_idx);
182 #undef FLOWC_PARAM
183 
184 	KASSERT(paramidx == nparams, ("nparams mismatch"));
185 
186 	txsd->tx_credits = howmany(flowclen, 16);
187 	txsd->plen = 0;
188 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
189 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
190 	toep->tx_credits -= txsd->tx_credits;
191 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
192 		toep->txsd_pidx = 0;
193 	toep->txsd_avail--;
194 
195 	toep->flags |= TPF_FLOWC_WR_SENT;
196         t4_wrq_tx(sc, wr);
197 }
198 
199 #ifdef RATELIMIT
200 /*
201  * Input is Bytes/second (so_max_pacing_rate), chip counts in Kilobits/second.
202  */
203 static int
204 update_tx_rate_limit(struct adapter *sc, struct toepcb *toep, u_int Bps)
205 {
206 	int tc_idx, rc;
207 	const u_int kbps = (u_int) (uint64_t)Bps * 8ULL / 1000;
208 	const int port_id = toep->vi->pi->port_id;
209 
210 	CTR3(KTR_CXGBE, "%s: tid %u, rate %uKbps", __func__, toep->tid, kbps);
211 
212 	if (kbps == 0) {
213 		/* unbind */
214 		tc_idx = -1;
215 	} else {
216 		rc = t4_reserve_cl_rl_kbps(sc, port_id, kbps, &tc_idx);
217 		if (rc != 0)
218 			return (rc);
219 		MPASS(tc_idx >= 0 && tc_idx < sc->chip_params->nsched_cls);
220 	}
221 
222 	if (toep->tc_idx != tc_idx) {
223 		struct wrqe *wr;
224 		struct fw_flowc_wr *flowc;
225 		int nparams = 1, flowclen, flowclen16;
226 		struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
227 
228 		flowclen = sizeof(*flowc) + nparams * sizeof(struct
229 		    fw_flowc_mnemval);
230 		flowclen16 = howmany(flowclen, 16);
231 		if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0 ||
232 		    (wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq)) == NULL) {
233 			if (tc_idx >= 0)
234 				t4_release_cl_rl(sc, port_id, tc_idx);
235 			return (ENOMEM);
236 		}
237 
238 		flowc = wrtod(wr);
239 		memset(flowc, 0, wr->wr_len);
240 
241 		flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
242 		    V_FW_FLOWC_WR_NPARAMS(nparams));
243 		flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
244 		    V_FW_WR_FLOWID(toep->tid));
245 
246 		flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
247 		if (tc_idx == -1)
248 			flowc->mnemval[0].val = htobe32(0xff);
249 		else
250 			flowc->mnemval[0].val = htobe32(tc_idx);
251 
252 		txsd->tx_credits = flowclen16;
253 		txsd->plen = 0;
254 		toep->tx_credits -= txsd->tx_credits;
255 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
256 			toep->txsd_pidx = 0;
257 		toep->txsd_avail--;
258 		t4_wrq_tx(sc, wr);
259 	}
260 
261 	if (toep->tc_idx >= 0)
262 		t4_release_cl_rl(sc, port_id, toep->tc_idx);
263 	toep->tc_idx = tc_idx;
264 
265 	return (0);
266 }
267 #endif
268 
269 void
270 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
271 {
272 	struct wrqe *wr;
273 	struct cpl_abort_req *req;
274 	int tid = toep->tid;
275 	struct inpcb *inp = toep->inp;
276 	struct tcpcb *tp = intotcpcb(inp);	/* don't use if INP_DROPPED */
277 
278 	INP_WLOCK_ASSERT(inp);
279 
280 	CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
281 	    __func__, toep->tid,
282 	    inp->inp_flags & INP_DROPPED ? "inp dropped" :
283 	    tcpstates[tp->t_state],
284 	    toep->flags, inp->inp_flags,
285 	    toep->flags & TPF_ABORT_SHUTDOWN ?
286 	    " (abort already in progress)" : "");
287 
288 	if (toep->flags & TPF_ABORT_SHUTDOWN)
289 		return;	/* abort already in progress */
290 
291 	toep->flags |= TPF_ABORT_SHUTDOWN;
292 
293 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
294 	    ("%s: flowc_wr not sent for tid %d.", __func__, tid));
295 
296 	wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
297 	if (wr == NULL) {
298 		/* XXX */
299 		panic("%s: allocation failure.", __func__);
300 	}
301 	req = wrtod(wr);
302 
303 	INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
304 	if (inp->inp_flags & INP_DROPPED)
305 		req->rsvd0 = htobe32(snd_nxt);
306 	else
307 		req->rsvd0 = htobe32(tp->snd_nxt);
308 	req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
309 	req->cmd = CPL_ABORT_SEND_RST;
310 
311 	/*
312 	 * XXX: What's the correct way to tell that the inp hasn't been detached
313 	 * from its socket?  Should I even be flushing the snd buffer here?
314 	 */
315 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
316 		struct socket *so = inp->inp_socket;
317 
318 		if (so != NULL)	/* because I'm not sure.  See comment above */
319 			sbflush(&so->so_snd);
320 	}
321 
322 	t4_l2t_send(sc, wr, toep->l2te);
323 }
324 
325 /*
326  * Called when a connection is established to translate the TCP options
327  * reported by HW to FreeBSD's native format.
328  */
329 static void
330 assign_rxopt(struct tcpcb *tp, unsigned int opt)
331 {
332 	struct toepcb *toep = tp->t_toe;
333 	struct inpcb *inp = tp->t_inpcb;
334 	struct adapter *sc = td_adapter(toep->td);
335 	int n;
336 
337 	INP_LOCK_ASSERT(inp);
338 
339 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
340 		n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
341 	else
342 		n = sizeof(struct ip) + sizeof(struct tcphdr);
343 	tp->t_maxseg = sc->params.mtus[G_TCPOPT_MSS(opt)] - n;
344 
345 	if (G_TCPOPT_TSTAMP(opt)) {
346 		tp->t_flags |= TF_RCVD_TSTMP;	/* timestamps ok */
347 		tp->ts_recent = 0;		/* hmmm */
348 		tp->ts_recent_age = tcp_ts_getticks();
349 		tp->t_maxseg -= TCPOLEN_TSTAMP_APPA;
350 	}
351 
352 	CTR5(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u), mss %u", __func__,
353 	    toep->tid, G_TCPOPT_MSS(opt), sc->params.mtus[G_TCPOPT_MSS(opt)],
354 	    tp->t_maxseg);
355 
356 	if (G_TCPOPT_SACK(opt))
357 		tp->t_flags |= TF_SACK_PERMIT;	/* should already be set */
358 	else
359 		tp->t_flags &= ~TF_SACK_PERMIT;	/* sack disallowed by peer */
360 
361 	if (G_TCPOPT_WSCALE_OK(opt))
362 		tp->t_flags |= TF_RCVD_SCALE;
363 
364 	/* Doing window scaling? */
365 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
366 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
367 		tp->rcv_scale = tp->request_r_scale;
368 		tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
369 	}
370 }
371 
372 /*
373  * Completes some final bits of initialization for just established connections
374  * and changes their state to TCPS_ESTABLISHED.
375  *
376  * The ISNs are from after the exchange of SYNs.  i.e., the true ISN + 1.
377  */
378 void
379 make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn,
380     uint16_t opt)
381 {
382 	struct inpcb *inp = toep->inp;
383 	struct socket *so = inp->inp_socket;
384 	struct tcpcb *tp = intotcpcb(inp);
385 	long bufsize;
386 	uint32_t iss = be32toh(snd_isn) - 1;	/* true ISS */
387 	uint32_t irs = be32toh(rcv_isn) - 1;	/* true IRS */
388 	uint16_t tcpopt = be16toh(opt);
389 	struct flowc_tx_params ftxp;
390 
391 	INP_WLOCK_ASSERT(inp);
392 	KASSERT(tp->t_state == TCPS_SYN_SENT ||
393 	    tp->t_state == TCPS_SYN_RECEIVED,
394 	    ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
395 
396 	CTR6(KTR_CXGBE, "%s: tid %d, so %p, inp %p, tp %p, toep %p",
397 	    __func__, toep->tid, so, inp, tp, toep);
398 
399 	tp->t_state = TCPS_ESTABLISHED;
400 	tp->t_starttime = ticks;
401 	TCPSTAT_INC(tcps_connects);
402 
403 	tp->irs = irs;
404 	tcp_rcvseqinit(tp);
405 	tp->rcv_wnd = toep->rx_credits << 10;
406 	tp->rcv_adv += tp->rcv_wnd;
407 	tp->last_ack_sent = tp->rcv_nxt;
408 
409 	/*
410 	 * If we were unable to send all rx credits via opt0, save the remainder
411 	 * in rx_credits so that they can be handed over with the next credit
412 	 * update.
413 	 */
414 	SOCKBUF_LOCK(&so->so_rcv);
415 	bufsize = select_rcv_wnd(so);
416 	SOCKBUF_UNLOCK(&so->so_rcv);
417 	toep->rx_credits = bufsize - tp->rcv_wnd;
418 
419 	tp->iss = iss;
420 	tcp_sendseqinit(tp);
421 	tp->snd_una = iss + 1;
422 	tp->snd_nxt = iss + 1;
423 	tp->snd_max = iss + 1;
424 
425 	assign_rxopt(tp, tcpopt);
426 
427 	SOCKBUF_LOCK(&so->so_snd);
428 	if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf)
429 		bufsize = V_tcp_autosndbuf_max;
430 	else
431 		bufsize = sbspace(&so->so_snd);
432 	SOCKBUF_UNLOCK(&so->so_snd);
433 
434 	ftxp.snd_nxt = tp->snd_nxt;
435 	ftxp.rcv_nxt = tp->rcv_nxt;
436 	ftxp.snd_space = bufsize;
437 	ftxp.mss = tp->t_maxseg;
438 	send_flowc_wr(toep, &ftxp);
439 
440 	soisconnected(so);
441 }
442 
443 int
444 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
445 {
446 	struct wrqe *wr;
447 	struct cpl_rx_data_ack *req;
448 	uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
449 
450 	KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
451 
452 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
453 	if (wr == NULL)
454 		return (0);
455 	req = wrtod(wr);
456 
457 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
458 	req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
459 
460 	t4_wrq_tx(sc, wr);
461 	return (credits);
462 }
463 
464 void
465 send_rx_modulate(struct adapter *sc, struct toepcb *toep)
466 {
467 	struct wrqe *wr;
468 	struct cpl_rx_data_ack *req;
469 
470 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
471 	if (wr == NULL)
472 		return;
473 	req = wrtod(wr);
474 
475 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
476 	req->credit_dack = htobe32(F_RX_MODULATE_RX);
477 
478 	t4_wrq_tx(sc, wr);
479 }
480 
481 void
482 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp)
483 {
484 	struct adapter *sc = tod->tod_softc;
485 	struct inpcb *inp = tp->t_inpcb;
486 	struct socket *so = inp->inp_socket;
487 	struct sockbuf *sb = &so->so_rcv;
488 	struct toepcb *toep = tp->t_toe;
489 	int credits;
490 
491 	INP_WLOCK_ASSERT(inp);
492 
493 	SOCKBUF_LOCK_ASSERT(sb);
494 	KASSERT(toep->sb_cc >= sbused(sb),
495 	    ("%s: sb %p has more data (%d) than last time (%d).",
496 	    __func__, sb, sbused(sb), toep->sb_cc));
497 
498 	credits = toep->sb_cc - sbused(sb);
499 	toep->sb_cc = sbused(sb);
500 	if (toep->ulp_mode == ULP_MODE_TLS) {
501 		if (toep->tls.rcv_over >= credits) {
502 			toep->tls.rcv_over -= credits;
503 			credits = 0;
504 		} else {
505 			credits -= toep->tls.rcv_over;
506 			toep->tls.rcv_over = 0;
507 		}
508 	}
509 	toep->rx_credits += credits;
510 
511 	if (toep->rx_credits > 0 &&
512 	    (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 ||
513 	    (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
514 	    toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) {
515 
516 		credits = send_rx_credits(sc, toep, toep->rx_credits);
517 		toep->rx_credits -= credits;
518 		tp->rcv_wnd += credits;
519 		tp->rcv_adv += credits;
520 	} else if (toep->flags & TPF_FORCE_CREDITS)
521 		send_rx_modulate(sc, toep);
522 }
523 
524 void
525 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
526 {
527 	struct inpcb *inp = tp->t_inpcb;
528 	struct socket *so = inp->inp_socket;
529 	struct sockbuf *sb = &so->so_rcv;
530 
531 	SOCKBUF_LOCK(sb);
532 	t4_rcvd_locked(tod, tp);
533 	SOCKBUF_UNLOCK(sb);
534 }
535 
536 /*
537  * Close a connection by sending a CPL_CLOSE_CON_REQ message.
538  */
539 int
540 t4_close_conn(struct adapter *sc, struct toepcb *toep)
541 {
542 	struct wrqe *wr;
543 	struct cpl_close_con_req *req;
544 	unsigned int tid = toep->tid;
545 
546 	CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
547 	    toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
548 
549 	if (toep->flags & TPF_FIN_SENT)
550 		return (0);
551 
552 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
553 	    ("%s: flowc_wr not sent for tid %u.", __func__, tid));
554 
555 	wr = alloc_wrqe(sizeof(*req), toep->ofld_txq);
556 	if (wr == NULL) {
557 		/* XXX */
558 		panic("%s: allocation failure.", __func__);
559 	}
560 	req = wrtod(wr);
561 
562         req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
563 	    V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
564 	req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
565 	    V_FW_WR_FLOWID(tid));
566         req->wr.wr_lo = cpu_to_be64(0);
567         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
568 	req->rsvd = 0;
569 
570 	toep->flags |= TPF_FIN_SENT;
571 	toep->flags &= ~TPF_SEND_FIN;
572 	t4_l2t_send(sc, wr, toep->l2te);
573 
574 	return (0);
575 }
576 
577 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
578 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
579 
580 /* Maximum amount of immediate data we could stuff in a WR */
581 static inline int
582 max_imm_payload(int tx_credits)
583 {
584 	const int n = 2;	/* Use only up to 2 desc for imm. data WR */
585 
586 	KASSERT(tx_credits >= 0 &&
587 		tx_credits <= MAX_OFLD_TX_CREDITS,
588 		("%s: %d credits", __func__, tx_credits));
589 
590 	if (tx_credits < MIN_OFLD_TX_CREDITS)
591 		return (0);
592 
593 	if (tx_credits >= (n * EQ_ESIZE) / 16)
594 		return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
595 	else
596 		return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
597 }
598 
599 /* Maximum number of SGL entries we could stuff in a WR */
600 static inline int
601 max_dsgl_nsegs(int tx_credits)
602 {
603 	int nseg = 1;	/* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
604 	int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
605 
606 	KASSERT(tx_credits >= 0 &&
607 		tx_credits <= MAX_OFLD_TX_CREDITS,
608 		("%s: %d credits", __func__, tx_credits));
609 
610 	if (tx_credits < MIN_OFLD_TX_CREDITS)
611 		return (0);
612 
613 	nseg += 2 * (sge_pair_credits * 16 / 24);
614 	if ((sge_pair_credits * 16) % 24 == 16)
615 		nseg++;
616 
617 	return (nseg);
618 }
619 
620 static inline void
621 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
622     unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign)
623 {
624 	struct fw_ofld_tx_data_wr *txwr = dst;
625 
626 	txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
627 	    V_FW_WR_IMMDLEN(immdlen));
628 	txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
629 	    V_FW_WR_LEN16(credits));
630 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) |
631 	    V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
632 	txwr->plen = htobe32(plen);
633 
634 	if (txalign > 0) {
635 		struct tcpcb *tp = intotcpcb(toep->inp);
636 
637 		if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi))
638 			txwr->lsodisable_to_flags |=
639 			    htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
640 		else
641 			txwr->lsodisable_to_flags |=
642 			    htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
643 				(tp->t_flags & TF_NODELAY ? 0 :
644 				F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
645 	}
646 }
647 
648 /*
649  * Generate a DSGL from a starting mbuf.  The total number of segments and the
650  * maximum segments in any one mbuf are provided.
651  */
652 static void
653 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
654 {
655 	struct mbuf *m;
656 	struct ulptx_sgl *usgl = dst;
657 	int i, j, rc;
658 	struct sglist sg;
659 	struct sglist_seg segs[n];
660 
661 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
662 
663 	sglist_init(&sg, n, segs);
664 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
665 	    V_ULPTX_NSGE(nsegs));
666 
667 	i = -1;
668 	for (m = start; m != stop; m = m->m_next) {
669 		if (IS_AIOTX_MBUF(m))
670 			rc = sglist_append_vmpages(&sg, aiotx_mbuf_pages(m),
671 			    aiotx_mbuf_pgoff(m), m->m_len);
672 		else
673 			rc = sglist_append(&sg, mtod(m, void *), m->m_len);
674 		if (__predict_false(rc != 0))
675 			panic("%s: sglist_append %d", __func__, rc);
676 
677 		for (j = 0; j < sg.sg_nseg; i++, j++) {
678 			if (i < 0) {
679 				usgl->len0 = htobe32(segs[j].ss_len);
680 				usgl->addr0 = htobe64(segs[j].ss_paddr);
681 			} else {
682 				usgl->sge[i / 2].len[i & 1] =
683 				    htobe32(segs[j].ss_len);
684 				usgl->sge[i / 2].addr[i & 1] =
685 				    htobe64(segs[j].ss_paddr);
686 			}
687 #ifdef INVARIANTS
688 			nsegs--;
689 #endif
690 		}
691 		sglist_reset(&sg);
692 	}
693 	if (i & 1)
694 		usgl->sge[i / 2].len[1] = htobe32(0);
695 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
696 	    __func__, nsegs, start, stop));
697 }
698 
699 /*
700  * Max number of SGL entries an offload tx work request can have.  This is 41
701  * (1 + 40) for a full 512B work request.
702  * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
703  */
704 #define OFLD_SGL_LEN (41)
705 
706 /*
707  * Send data and/or a FIN to the peer.
708  *
709  * The socket's so_snd buffer consists of a stream of data starting with sb_mb
710  * and linked together with m_next.  sb_sndptr, if set, is the last mbuf that
711  * was transmitted.
712  *
713  * drop indicates the number of bytes that should be dropped from the head of
714  * the send buffer.  It is an optimization that lets do_fw4_ack avoid creating
715  * contention on the send buffer lock (before this change it used to do
716  * sowwakeup and then t4_push_frames right after that when recovering from tx
717  * stalls).  When drop is set this function MUST drop the bytes and wake up any
718  * writers.
719  */
720 void
721 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
722 {
723 	struct mbuf *sndptr, *m, *sb_sndptr;
724 	struct fw_ofld_tx_data_wr *txwr;
725 	struct wrqe *wr;
726 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
727 	struct inpcb *inp = toep->inp;
728 	struct tcpcb *tp = intotcpcb(inp);
729 	struct socket *so = inp->inp_socket;
730 	struct sockbuf *sb = &so->so_snd;
731 	int tx_credits, shove, compl, sowwakeup;
732 	struct ofld_tx_sdesc *txsd;
733 	bool aiotx_mbuf_seen;
734 
735 	INP_WLOCK_ASSERT(inp);
736 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
737 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
738 
739 	KASSERT(toep->ulp_mode == ULP_MODE_NONE ||
740 	    toep->ulp_mode == ULP_MODE_TCPDDP ||
741 	    toep->ulp_mode == ULP_MODE_TLS ||
742 	    toep->ulp_mode == ULP_MODE_RDMA,
743 	    ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
744 
745 #ifdef VERBOSE_TRACES
746 	CTR4(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
747 	    __func__, toep->tid, toep->flags, tp->t_flags);
748 #endif
749 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
750 		return;
751 
752 #ifdef RATELIMIT
753 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
754 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
755 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
756 	}
757 #endif
758 
759 	/*
760 	 * This function doesn't resume by itself.  Someone else must clear the
761 	 * flag and call this function.
762 	 */
763 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
764 		KASSERT(drop == 0,
765 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
766 		return;
767 	}
768 
769 	txsd = &toep->txsd[toep->txsd_pidx];
770 	do {
771 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
772 		max_imm = max_imm_payload(tx_credits);
773 		max_nsegs = max_dsgl_nsegs(tx_credits);
774 
775 		SOCKBUF_LOCK(sb);
776 		sowwakeup = drop;
777 		if (drop) {
778 			sbdrop_locked(sb, drop);
779 			drop = 0;
780 		}
781 		sb_sndptr = sb->sb_sndptr;
782 		sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
783 		plen = 0;
784 		nsegs = 0;
785 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
786 		aiotx_mbuf_seen = false;
787 		for (m = sndptr; m != NULL; m = m->m_next) {
788 			int n;
789 
790 			if (IS_AIOTX_MBUF(m))
791 				n = sglist_count_vmpages(aiotx_mbuf_pages(m),
792 				    aiotx_mbuf_pgoff(m), m->m_len);
793 			else
794 				n = sglist_count(mtod(m, void *), m->m_len);
795 
796 			nsegs += n;
797 			plen += m->m_len;
798 
799 			/* This mbuf sent us _over_ the nsegs limit, back out */
800 			if (plen > max_imm && nsegs > max_nsegs) {
801 				nsegs -= n;
802 				plen -= m->m_len;
803 				if (plen == 0) {
804 					/* Too few credits */
805 					toep->flags |= TPF_TX_SUSPENDED;
806 					if (sowwakeup) {
807 						if (!TAILQ_EMPTY(
808 						    &toep->aiotx_jobq))
809 							t4_aiotx_queue_toep(
810 							    toep);
811 						sowwakeup_locked(so);
812 					} else
813 						SOCKBUF_UNLOCK(sb);
814 					SOCKBUF_UNLOCK_ASSERT(sb);
815 					return;
816 				}
817 				break;
818 			}
819 
820 			if (IS_AIOTX_MBUF(m))
821 				aiotx_mbuf_seen = true;
822 			if (max_nsegs_1mbuf < n)
823 				max_nsegs_1mbuf = n;
824 			sb_sndptr = m;	/* new sb->sb_sndptr if all goes well */
825 
826 			/* This mbuf put us right at the max_nsegs limit */
827 			if (plen > max_imm && nsegs == max_nsegs) {
828 				m = m->m_next;
829 				break;
830 			}
831 		}
832 
833 		if (sbused(sb) > sb->sb_hiwat * 5 / 8 &&
834 		    toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
835 			compl = 1;
836 		else
837 			compl = 0;
838 
839 		if (sb->sb_flags & SB_AUTOSIZE &&
840 		    V_tcp_do_autosndbuf &&
841 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
842 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
843 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
844 			    V_tcp_autosndbuf_max);
845 
846 			if (!sbreserve_locked(sb, newsize, so, NULL))
847 				sb->sb_flags &= ~SB_AUTOSIZE;
848 			else
849 				sowwakeup = 1;	/* room available */
850 		}
851 		if (sowwakeup) {
852 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
853 				t4_aiotx_queue_toep(toep);
854 			sowwakeup_locked(so);
855 		} else
856 			SOCKBUF_UNLOCK(sb);
857 		SOCKBUF_UNLOCK_ASSERT(sb);
858 
859 		/* nothing to send */
860 		if (plen == 0) {
861 			KASSERT(m == NULL,
862 			    ("%s: nothing to send, but m != NULL", __func__));
863 			break;
864 		}
865 
866 		if (__predict_false(toep->flags & TPF_FIN_SENT))
867 			panic("%s: excess tx.", __func__);
868 
869 		shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
870 		if (plen <= max_imm && !aiotx_mbuf_seen) {
871 
872 			/* Immediate data tx */
873 
874 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
875 					toep->ofld_txq);
876 			if (wr == NULL) {
877 				/* XXX: how will we recover from this? */
878 				toep->flags |= TPF_TX_SUSPENDED;
879 				return;
880 			}
881 			txwr = wrtod(wr);
882 			credits = howmany(wr->wr_len, 16);
883 			write_tx_wr(txwr, toep, plen, plen, credits, shove, 0,
884 			    sc->tt.tx_align);
885 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
886 			nsegs = 0;
887 		} else {
888 			int wr_len;
889 
890 			/* DSGL tx */
891 
892 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
893 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
894 			wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
895 			if (wr == NULL) {
896 				/* XXX: how will we recover from this? */
897 				toep->flags |= TPF_TX_SUSPENDED;
898 				return;
899 			}
900 			txwr = wrtod(wr);
901 			credits = howmany(wr_len, 16);
902 			write_tx_wr(txwr, toep, 0, plen, credits, shove, 0,
903 			    sc->tt.tx_align);
904 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
905 			    max_nsegs_1mbuf);
906 			if (wr_len & 0xf) {
907 				uint64_t *pad = (uint64_t *)
908 				    ((uintptr_t)txwr + wr_len);
909 				*pad = 0;
910 			}
911 		}
912 
913 		KASSERT(toep->tx_credits >= credits,
914 			("%s: not enough credits", __func__));
915 
916 		toep->tx_credits -= credits;
917 		toep->tx_nocompl += credits;
918 		toep->plen_nocompl += plen;
919 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
920 		    toep->tx_nocompl >= toep->tx_total / 4)
921 			compl = 1;
922 
923 		if (compl || toep->ulp_mode == ULP_MODE_RDMA) {
924 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
925 			toep->tx_nocompl = 0;
926 			toep->plen_nocompl = 0;
927 		}
928 
929 		tp->snd_nxt += plen;
930 		tp->snd_max += plen;
931 
932 		SOCKBUF_LOCK(sb);
933 		KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
934 		sb->sb_sndptr = sb_sndptr;
935 		SOCKBUF_UNLOCK(sb);
936 
937 		toep->flags |= TPF_TX_DATA_SENT;
938 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
939 			toep->flags |= TPF_TX_SUSPENDED;
940 
941 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
942 		txsd->plen = plen;
943 		txsd->tx_credits = credits;
944 		txsd++;
945 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
946 			toep->txsd_pidx = 0;
947 			txsd = &toep->txsd[0];
948 		}
949 		toep->txsd_avail--;
950 
951 		t4_l2t_send(sc, wr, toep->l2te);
952 	} while (m != NULL);
953 
954 	/* Send a FIN if requested, but only if there's no more data to send */
955 	if (m == NULL && toep->flags & TPF_SEND_FIN)
956 		t4_close_conn(sc, toep);
957 }
958 
959 static inline void
960 rqdrop_locked(struct mbufq *q, int plen)
961 {
962 	struct mbuf *m;
963 
964 	while (plen > 0) {
965 		m = mbufq_dequeue(q);
966 
967 		/* Too many credits. */
968 		MPASS(m != NULL);
969 		M_ASSERTPKTHDR(m);
970 
971 		/* Partial credits. */
972 		MPASS(plen >= m->m_pkthdr.len);
973 
974 		plen -= m->m_pkthdr.len;
975 		m_freem(m);
976 	}
977 }
978 
979 void
980 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
981 {
982 	struct mbuf *sndptr, *m;
983 	struct fw_ofld_tx_data_wr *txwr;
984 	struct wrqe *wr;
985 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
986 	u_int adjusted_plen, ulp_submode;
987 	struct inpcb *inp = toep->inp;
988 	struct tcpcb *tp = intotcpcb(inp);
989 	int tx_credits, shove;
990 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
991 	struct mbufq *pduq = &toep->ulp_pduq;
992 	static const u_int ulp_extra_len[] = {0, 4, 4, 8};
993 
994 	INP_WLOCK_ASSERT(inp);
995 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
996 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
997 	KASSERT(toep->ulp_mode == ULP_MODE_ISCSI,
998 	    ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep));
999 
1000 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1001 		return;
1002 
1003 	/*
1004 	 * This function doesn't resume by itself.  Someone else must clear the
1005 	 * flag and call this function.
1006 	 */
1007 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1008 		KASSERT(drop == 0,
1009 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1010 		return;
1011 	}
1012 
1013 	if (drop)
1014 		rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
1015 
1016 	while ((sndptr = mbufq_first(pduq)) != NULL) {
1017 		M_ASSERTPKTHDR(sndptr);
1018 
1019 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
1020 		max_imm = max_imm_payload(tx_credits);
1021 		max_nsegs = max_dsgl_nsegs(tx_credits);
1022 
1023 		plen = 0;
1024 		nsegs = 0;
1025 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
1026 		for (m = sndptr; m != NULL; m = m->m_next) {
1027 			int n = sglist_count(mtod(m, void *), m->m_len);
1028 
1029 			nsegs += n;
1030 			plen += m->m_len;
1031 
1032 			/*
1033 			 * This mbuf would send us _over_ the nsegs limit.
1034 			 * Suspend tx because the PDU can't be sent out.
1035 			 */
1036 			if (plen > max_imm && nsegs > max_nsegs) {
1037 				toep->flags |= TPF_TX_SUSPENDED;
1038 				return;
1039 			}
1040 
1041 			if (max_nsegs_1mbuf < n)
1042 				max_nsegs_1mbuf = n;
1043 		}
1044 
1045 		if (__predict_false(toep->flags & TPF_FIN_SENT))
1046 			panic("%s: excess tx.", __func__);
1047 
1048 		/*
1049 		 * We have a PDU to send.  All of it goes out in one WR so 'm'
1050 		 * is NULL.  A PDU's length is always a multiple of 4.
1051 		 */
1052 		MPASS(m == NULL);
1053 		MPASS((plen & 3) == 0);
1054 		MPASS(sndptr->m_pkthdr.len == plen);
1055 
1056 		shove = !(tp->t_flags & TF_MORETOCOME);
1057 		ulp_submode = mbuf_ulp_submode(sndptr);
1058 		MPASS(ulp_submode < nitems(ulp_extra_len));
1059 
1060 		/*
1061 		 * plen doesn't include header and data digests, which are
1062 		 * generated and inserted in the right places by the TOE, but
1063 		 * they do occupy TCP sequence space and need to be accounted
1064 		 * for.
1065 		 */
1066 		adjusted_plen = plen + ulp_extra_len[ulp_submode];
1067 		if (plen <= max_imm) {
1068 
1069 			/* Immediate data tx */
1070 
1071 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
1072 					toep->ofld_txq);
1073 			if (wr == NULL) {
1074 				/* XXX: how will we recover from this? */
1075 				toep->flags |= TPF_TX_SUSPENDED;
1076 				return;
1077 			}
1078 			txwr = wrtod(wr);
1079 			credits = howmany(wr->wr_len, 16);
1080 			write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
1081 			    shove, ulp_submode, sc->tt.tx_align);
1082 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
1083 			nsegs = 0;
1084 		} else {
1085 			int wr_len;
1086 
1087 			/* DSGL tx */
1088 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
1089 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1090 			wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
1091 			if (wr == NULL) {
1092 				/* XXX: how will we recover from this? */
1093 				toep->flags |= TPF_TX_SUSPENDED;
1094 				return;
1095 			}
1096 			txwr = wrtod(wr);
1097 			credits = howmany(wr_len, 16);
1098 			write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
1099 			    shove, ulp_submode, sc->tt.tx_align);
1100 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
1101 			    max_nsegs_1mbuf);
1102 			if (wr_len & 0xf) {
1103 				uint64_t *pad = (uint64_t *)
1104 				    ((uintptr_t)txwr + wr_len);
1105 				*pad = 0;
1106 			}
1107 		}
1108 
1109 		KASSERT(toep->tx_credits >= credits,
1110 			("%s: not enough credits", __func__));
1111 
1112 		m = mbufq_dequeue(pduq);
1113 		MPASS(m == sndptr);
1114 		mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
1115 
1116 		toep->tx_credits -= credits;
1117 		toep->tx_nocompl += credits;
1118 		toep->plen_nocompl += plen;
1119 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
1120 		    toep->tx_nocompl >= toep->tx_total / 4) {
1121 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
1122 			toep->tx_nocompl = 0;
1123 			toep->plen_nocompl = 0;
1124 		}
1125 
1126 		tp->snd_nxt += adjusted_plen;
1127 		tp->snd_max += adjusted_plen;
1128 
1129 		toep->flags |= TPF_TX_DATA_SENT;
1130 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
1131 			toep->flags |= TPF_TX_SUSPENDED;
1132 
1133 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1134 		txsd->plen = plen;
1135 		txsd->tx_credits = credits;
1136 		txsd++;
1137 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1138 			toep->txsd_pidx = 0;
1139 			txsd = &toep->txsd[0];
1140 		}
1141 		toep->txsd_avail--;
1142 
1143 		t4_l2t_send(sc, wr, toep->l2te);
1144 	}
1145 
1146 	/* Send a FIN if requested, but only if there are no more PDUs to send */
1147 	if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
1148 		t4_close_conn(sc, toep);
1149 }
1150 
1151 int
1152 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
1153 {
1154 	struct adapter *sc = tod->tod_softc;
1155 #ifdef INVARIANTS
1156 	struct inpcb *inp = tp->t_inpcb;
1157 #endif
1158 	struct toepcb *toep = tp->t_toe;
1159 
1160 	INP_WLOCK_ASSERT(inp);
1161 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1162 	    ("%s: inp %p dropped.", __func__, inp));
1163 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1164 
1165 	if (toep->ulp_mode == ULP_MODE_ISCSI)
1166 		t4_push_pdus(sc, toep, 0);
1167 	else if (tls_tx_key(toep))
1168 		t4_push_tls_records(sc, toep, 0);
1169 	else
1170 		t4_push_frames(sc, toep, 0);
1171 
1172 	return (0);
1173 }
1174 
1175 int
1176 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
1177 {
1178 	struct adapter *sc = tod->tod_softc;
1179 #ifdef INVARIANTS
1180 	struct inpcb *inp = tp->t_inpcb;
1181 #endif
1182 	struct toepcb *toep = tp->t_toe;
1183 
1184 	INP_WLOCK_ASSERT(inp);
1185 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1186 	    ("%s: inp %p dropped.", __func__, inp));
1187 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1188 
1189 	toep->flags |= TPF_SEND_FIN;
1190 	if (tp->t_state >= TCPS_ESTABLISHED) {
1191 		if (toep->ulp_mode == ULP_MODE_ISCSI)
1192 			t4_push_pdus(sc, toep, 0);
1193 		else if (tls_tx_key(toep))
1194 			t4_push_tls_records(sc, toep, 0);
1195 		else
1196 			t4_push_frames(sc, toep, 0);
1197 	}
1198 
1199 	return (0);
1200 }
1201 
1202 int
1203 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1204 {
1205 	struct adapter *sc = tod->tod_softc;
1206 #if defined(INVARIANTS)
1207 	struct inpcb *inp = tp->t_inpcb;
1208 #endif
1209 	struct toepcb *toep = tp->t_toe;
1210 
1211 	INP_WLOCK_ASSERT(inp);
1212 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1213 	    ("%s: inp %p dropped.", __func__, inp));
1214 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1215 
1216 	/* hmmmm */
1217 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1218 	    ("%s: flowc for tid %u [%s] not sent already",
1219 	    __func__, toep->tid, tcpstates[tp->t_state]));
1220 
1221 	send_reset(sc, toep, 0);
1222 	return (0);
1223 }
1224 
1225 /*
1226  * Peer has sent us a FIN.
1227  */
1228 static int
1229 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1230 {
1231 	struct adapter *sc = iq->adapter;
1232 	const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1233 	unsigned int tid = GET_TID(cpl);
1234 	struct toepcb *toep = lookup_tid(sc, tid);
1235 	struct inpcb *inp = toep->inp;
1236 	struct tcpcb *tp = NULL;
1237 	struct socket *so;
1238 	struct epoch_tracker et;
1239 #ifdef INVARIANTS
1240 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1241 #endif
1242 
1243 	KASSERT(opcode == CPL_PEER_CLOSE,
1244 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1245 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1246 
1247 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1248 #ifdef INVARIANTS
1249 		struct synq_entry *synqe = (void *)toep;
1250 
1251 		INP_WLOCK(synqe->lctx->inp);
1252 		if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1253 			KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1254 			    ("%s: listen socket closed but tid %u not aborted.",
1255 			    __func__, tid));
1256 		} else {
1257 			/*
1258 			 * do_pass_accept_req is still running and will
1259 			 * eventually take care of this tid.
1260 			 */
1261 		}
1262 		INP_WUNLOCK(synqe->lctx->inp);
1263 #endif
1264 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1265 		    toep, toep->flags);
1266 		return (0);
1267 	}
1268 
1269 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1270 
1271 	CURVNET_SET(toep->vnet);
1272 	INP_INFO_RLOCK_ET(&V_tcbinfo, et);
1273 	INP_WLOCK(inp);
1274 	tp = intotcpcb(inp);
1275 
1276 	CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__,
1277 	    tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp);
1278 
1279 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1280 		goto done;
1281 
1282 	tp->rcv_nxt++;	/* FIN */
1283 
1284 	so = inp->inp_socket;
1285 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1286 		DDP_LOCK(toep);
1287 		if (__predict_false(toep->ddp.flags &
1288 		    (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)))
1289 			handle_ddp_close(toep, tp, cpl->rcv_nxt);
1290 		DDP_UNLOCK(toep);
1291 	}
1292 	socantrcvmore(so);
1293 
1294 	if (toep->ulp_mode != ULP_MODE_RDMA) {
1295 		KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt),
1296 	    		("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1297 	    		be32toh(cpl->rcv_nxt)));
1298 	}
1299 
1300 	switch (tp->t_state) {
1301 	case TCPS_SYN_RECEIVED:
1302 		tp->t_starttime = ticks;
1303 		/* FALLTHROUGH */
1304 
1305 	case TCPS_ESTABLISHED:
1306 		tp->t_state = TCPS_CLOSE_WAIT;
1307 		break;
1308 
1309 	case TCPS_FIN_WAIT_1:
1310 		tp->t_state = TCPS_CLOSING;
1311 		break;
1312 
1313 	case TCPS_FIN_WAIT_2:
1314 		tcp_twstart(tp);
1315 		INP_UNLOCK_ASSERT(inp);	 /* safe, we have a ref on the inp */
1316 		INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1317 		CURVNET_RESTORE();
1318 
1319 		INP_WLOCK(inp);
1320 		final_cpl_received(toep);
1321 		return (0);
1322 
1323 	default:
1324 		log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1325 		    __func__, tid, tp->t_state);
1326 	}
1327 done:
1328 	INP_WUNLOCK(inp);
1329 	INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1330 	CURVNET_RESTORE();
1331 	return (0);
1332 }
1333 
1334 /*
1335  * Peer has ACK'd our FIN.
1336  */
1337 static int
1338 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1339     struct mbuf *m)
1340 {
1341 	struct adapter *sc = iq->adapter;
1342 	const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1343 	unsigned int tid = GET_TID(cpl);
1344 	struct toepcb *toep = lookup_tid(sc, tid);
1345 	struct inpcb *inp = toep->inp;
1346 	struct tcpcb *tp = NULL;
1347 	struct socket *so = NULL;
1348 	struct epoch_tracker et;
1349 #ifdef INVARIANTS
1350 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1351 #endif
1352 
1353 	KASSERT(opcode == CPL_CLOSE_CON_RPL,
1354 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1355 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1356 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1357 
1358 	CURVNET_SET(toep->vnet);
1359 	INP_INFO_RLOCK_ET(&V_tcbinfo, et);
1360 	INP_WLOCK(inp);
1361 	tp = intotcpcb(inp);
1362 
1363 	CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1364 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1365 
1366 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1367 		goto done;
1368 
1369 	so = inp->inp_socket;
1370 	tp->snd_una = be32toh(cpl->snd_nxt) - 1;	/* exclude FIN */
1371 
1372 	switch (tp->t_state) {
1373 	case TCPS_CLOSING:	/* see TCPS_FIN_WAIT_2 in do_peer_close too */
1374 		tcp_twstart(tp);
1375 release:
1376 		INP_UNLOCK_ASSERT(inp);	/* safe, we have a ref on the  inp */
1377 		INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1378 		CURVNET_RESTORE();
1379 
1380 		INP_WLOCK(inp);
1381 		final_cpl_received(toep);	/* no more CPLs expected */
1382 
1383 		return (0);
1384 	case TCPS_LAST_ACK:
1385 		if (tcp_close(tp))
1386 			INP_WUNLOCK(inp);
1387 		goto release;
1388 
1389 	case TCPS_FIN_WAIT_1:
1390 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1391 			soisdisconnected(so);
1392 		tp->t_state = TCPS_FIN_WAIT_2;
1393 		break;
1394 
1395 	default:
1396 		log(LOG_ERR,
1397 		    "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1398 		    __func__, tid, tcpstates[tp->t_state]);
1399 	}
1400 done:
1401 	INP_WUNLOCK(inp);
1402 	INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1403 	CURVNET_RESTORE();
1404 	return (0);
1405 }
1406 
1407 void
1408 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid,
1409     int rst_status)
1410 {
1411 	struct wrqe *wr;
1412 	struct cpl_abort_rpl *cpl;
1413 
1414 	wr = alloc_wrqe(sizeof(*cpl), ofld_txq);
1415 	if (wr == NULL) {
1416 		/* XXX */
1417 		panic("%s: allocation failure.", __func__);
1418 	}
1419 	cpl = wrtod(wr);
1420 
1421 	INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1422 	cpl->cmd = rst_status;
1423 
1424 	t4_wrq_tx(sc, wr);
1425 }
1426 
1427 static int
1428 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1429 {
1430 	switch (abort_reason) {
1431 	case CPL_ERR_BAD_SYN:
1432 	case CPL_ERR_CONN_RESET:
1433 		return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1434 	case CPL_ERR_XMIT_TIMEDOUT:
1435 	case CPL_ERR_PERSIST_TIMEDOUT:
1436 	case CPL_ERR_FINWAIT2_TIMEDOUT:
1437 	case CPL_ERR_KEEPALIVE_TIMEDOUT:
1438 		return (ETIMEDOUT);
1439 	default:
1440 		return (EIO);
1441 	}
1442 }
1443 
1444 /*
1445  * TCP RST from the peer, timeout, or some other such critical error.
1446  */
1447 static int
1448 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1449 {
1450 	struct adapter *sc = iq->adapter;
1451 	const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1452 	unsigned int tid = GET_TID(cpl);
1453 	struct toepcb *toep = lookup_tid(sc, tid);
1454 	struct sge_wrq *ofld_txq = toep->ofld_txq;
1455 	struct inpcb *inp;
1456 	struct tcpcb *tp;
1457 	struct epoch_tracker et;
1458 #ifdef INVARIANTS
1459 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1460 #endif
1461 
1462 	KASSERT(opcode == CPL_ABORT_REQ_RSS,
1463 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1464 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1465 
1466 	if (toep->flags & TPF_SYNQE)
1467 		return (do_abort_req_synqe(iq, rss, m));
1468 
1469 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1470 
1471 	if (negative_advice(cpl->status)) {
1472 		CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1473 		    __func__, cpl->status, tid, toep->flags);
1474 		return (0);	/* Ignore negative advice */
1475 	}
1476 
1477 	inp = toep->inp;
1478 	CURVNET_SET(toep->vnet);
1479 	INP_INFO_RLOCK_ET(&V_tcbinfo, et);	/* for tcp_close */
1480 	INP_WLOCK(inp);
1481 
1482 	tp = intotcpcb(inp);
1483 
1484 	CTR6(KTR_CXGBE,
1485 	    "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1486 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1487 	    inp->inp_flags, cpl->status);
1488 
1489 	/*
1490 	 * If we'd initiated an abort earlier the reply to it is responsible for
1491 	 * cleaning up resources.  Otherwise we tear everything down right here
1492 	 * right now.  We owe the T4 a CPL_ABORT_RPL no matter what.
1493 	 */
1494 	if (toep->flags & TPF_ABORT_SHUTDOWN) {
1495 		INP_WUNLOCK(inp);
1496 		goto done;
1497 	}
1498 	toep->flags |= TPF_ABORT_SHUTDOWN;
1499 
1500 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1501 		struct socket *so = inp->inp_socket;
1502 
1503 		if (so != NULL)
1504 			so_error_set(so, abort_status_to_errno(tp,
1505 			    cpl->status));
1506 		tp = tcp_close(tp);
1507 		if (tp == NULL)
1508 			INP_WLOCK(inp);	/* re-acquire */
1509 	}
1510 
1511 	final_cpl_received(toep);
1512 done:
1513 	INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1514 	CURVNET_RESTORE();
1515 	send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1516 	return (0);
1517 }
1518 
1519 /*
1520  * Reply to the CPL_ABORT_REQ (send_reset)
1521  */
1522 static int
1523 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1524 {
1525 	struct adapter *sc = iq->adapter;
1526 	const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1527 	unsigned int tid = GET_TID(cpl);
1528 	struct toepcb *toep = lookup_tid(sc, tid);
1529 	struct inpcb *inp = toep->inp;
1530 #ifdef INVARIANTS
1531 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1532 #endif
1533 
1534 	KASSERT(opcode == CPL_ABORT_RPL_RSS,
1535 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1536 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1537 
1538 	if (toep->flags & TPF_SYNQE)
1539 		return (do_abort_rpl_synqe(iq, rss, m));
1540 
1541 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1542 
1543 	CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1544 	    __func__, tid, toep, inp, cpl->status);
1545 
1546 	KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1547 	    ("%s: wasn't expecting abort reply", __func__));
1548 
1549 	INP_WLOCK(inp);
1550 	final_cpl_received(toep);
1551 
1552 	return (0);
1553 }
1554 
1555 static int
1556 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1557 {
1558 	struct adapter *sc = iq->adapter;
1559 	const struct cpl_rx_data *cpl = mtod(m, const void *);
1560 	unsigned int tid = GET_TID(cpl);
1561 	struct toepcb *toep = lookup_tid(sc, tid);
1562 	struct inpcb *inp = toep->inp;
1563 	struct tcpcb *tp;
1564 	struct socket *so;
1565 	struct sockbuf *sb;
1566 	struct epoch_tracker et;
1567 	int len;
1568 	uint32_t ddp_placed = 0;
1569 
1570 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1571 #ifdef INVARIANTS
1572 		struct synq_entry *synqe = (void *)toep;
1573 
1574 		INP_WLOCK(synqe->lctx->inp);
1575 		if (synqe->flags & TPF_SYNQE_HAS_L2TE) {
1576 			KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN,
1577 			    ("%s: listen socket closed but tid %u not aborted.",
1578 			    __func__, tid));
1579 		} else {
1580 			/*
1581 			 * do_pass_accept_req is still running and will
1582 			 * eventually take care of this tid.
1583 			 */
1584 		}
1585 		INP_WUNLOCK(synqe->lctx->inp);
1586 #endif
1587 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1588 		    toep, toep->flags);
1589 		m_freem(m);
1590 		return (0);
1591 	}
1592 
1593 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1594 
1595 	/* strip off CPL header */
1596 	m_adj(m, sizeof(*cpl));
1597 	len = m->m_pkthdr.len;
1598 
1599 	INP_WLOCK(inp);
1600 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1601 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1602 		    __func__, tid, len, inp->inp_flags);
1603 		INP_WUNLOCK(inp);
1604 		m_freem(m);
1605 		return (0);
1606 	}
1607 
1608 	tp = intotcpcb(inp);
1609 
1610 	if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1611 		ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1612 
1613 	tp->rcv_nxt += len;
1614 	if (tp->rcv_wnd < len) {
1615 		KASSERT(toep->ulp_mode == ULP_MODE_RDMA,
1616 				("%s: negative window size", __func__));
1617 	}
1618 
1619 	tp->rcv_wnd -= len;
1620 	tp->t_rcvtime = ticks;
1621 
1622 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
1623 		DDP_LOCK(toep);
1624 	so = inp_inpcbtosocket(inp);
1625 	sb = &so->so_rcv;
1626 	SOCKBUF_LOCK(sb);
1627 
1628 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1629 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1630 		    __func__, tid, len);
1631 		m_freem(m);
1632 		SOCKBUF_UNLOCK(sb);
1633 		if (toep->ulp_mode == ULP_MODE_TCPDDP)
1634 			DDP_UNLOCK(toep);
1635 		INP_WUNLOCK(inp);
1636 
1637 		CURVNET_SET(toep->vnet);
1638 		INP_INFO_RLOCK_ET(&V_tcbinfo, et);
1639 		INP_WLOCK(inp);
1640 		tp = tcp_drop(tp, ECONNRESET);
1641 		if (tp)
1642 			INP_WUNLOCK(inp);
1643 		INP_INFO_RUNLOCK_ET(&V_tcbinfo, et);
1644 		CURVNET_RESTORE();
1645 
1646 		return (0);
1647 	}
1648 
1649 	/* receive buffer autosize */
1650 	MPASS(toep->vnet == so->so_vnet);
1651 	CURVNET_SET(toep->vnet);
1652 	if (sb->sb_flags & SB_AUTOSIZE &&
1653 	    V_tcp_do_autorcvbuf &&
1654 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1655 	    len > (sbspace(sb) / 8 * 7)) {
1656 		unsigned int hiwat = sb->sb_hiwat;
1657 		unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc,
1658 		    V_tcp_autorcvbuf_max);
1659 
1660 		if (!sbreserve_locked(sb, newsize, so, NULL))
1661 			sb->sb_flags &= ~SB_AUTOSIZE;
1662 		else
1663 			toep->rx_credits += newsize - hiwat;
1664 	}
1665 
1666 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1667 		int changed = !(toep->ddp.flags & DDP_ON) ^ cpl->ddp_off;
1668 
1669 		if (toep->ddp.waiting_count != 0 || toep->ddp.active_count != 0)
1670 			CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)",
1671 			    __func__, tid, len);
1672 
1673 		if (changed) {
1674 			if (toep->ddp.flags & DDP_SC_REQ)
1675 				toep->ddp.flags ^= DDP_ON | DDP_SC_REQ;
1676 			else {
1677 				KASSERT(cpl->ddp_off == 1,
1678 				    ("%s: DDP switched on by itself.",
1679 				    __func__));
1680 
1681 				/* Fell out of DDP mode */
1682 				toep->ddp.flags &= ~DDP_ON;
1683 				CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
1684 				    __func__);
1685 
1686 				insert_ddp_data(toep, ddp_placed);
1687 			}
1688 		}
1689 
1690 		if (toep->ddp.flags & DDP_ON) {
1691 			/*
1692 			 * CPL_RX_DATA with DDP on can only be an indicate.
1693 			 * Start posting queued AIO requests via DDP.  The
1694 			 * payload that arrived in this indicate is appended
1695 			 * to the socket buffer as usual.
1696 			 */
1697 			handle_ddp_indicate(toep);
1698 		}
1699 	}
1700 
1701 	KASSERT(toep->sb_cc >= sbused(sb),
1702 	    ("%s: sb %p has more data (%d) than last time (%d).",
1703 	    __func__, sb, sbused(sb), toep->sb_cc));
1704 	toep->rx_credits += toep->sb_cc - sbused(sb);
1705 	sbappendstream_locked(sb, m, 0);
1706 	toep->sb_cc = sbused(sb);
1707 	if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) {
1708 		int credits;
1709 
1710 		credits = send_rx_credits(sc, toep, toep->rx_credits);
1711 		toep->rx_credits -= credits;
1712 		tp->rcv_wnd += credits;
1713 		tp->rcv_adv += credits;
1714 	}
1715 
1716 	if (toep->ulp_mode == ULP_MODE_TCPDDP && toep->ddp.waiting_count > 0 &&
1717 	    sbavail(sb) != 0) {
1718 		CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__,
1719 		    tid);
1720 		ddp_queue_toep(toep);
1721 	}
1722 	sorwakeup_locked(so);
1723 	SOCKBUF_UNLOCK_ASSERT(sb);
1724 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
1725 		DDP_UNLOCK(toep);
1726 
1727 	INP_WUNLOCK(inp);
1728 	CURVNET_RESTORE();
1729 	return (0);
1730 }
1731 
1732 static int
1733 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1734 {
1735 	struct adapter *sc = iq->adapter;
1736 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1737 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1738 	struct toepcb *toep = lookup_tid(sc, tid);
1739 	struct inpcb *inp;
1740 	struct tcpcb *tp;
1741 	struct socket *so;
1742 	uint8_t credits = cpl->credits;
1743 	struct ofld_tx_sdesc *txsd;
1744 	int plen;
1745 #ifdef INVARIANTS
1746 	unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1747 #endif
1748 
1749 	/*
1750 	 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1751 	 * now this comes back carrying the credits for the flowc.
1752 	 */
1753 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1754 		KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1755 		    ("%s: credits for a synq entry %p", __func__, toep));
1756 		return (0);
1757 	}
1758 
1759 	inp = toep->inp;
1760 
1761 	KASSERT(opcode == CPL_FW4_ACK,
1762 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1763 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1764 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1765 
1766 	INP_WLOCK(inp);
1767 
1768 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1769 		INP_WUNLOCK(inp);
1770 		return (0);
1771 	}
1772 
1773 	KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1774 	    ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1775 
1776 	tp = intotcpcb(inp);
1777 
1778 	if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1779 		tcp_seq snd_una = be32toh(cpl->snd_una);
1780 
1781 #ifdef INVARIANTS
1782 		if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1783 			log(LOG_ERR,
1784 			    "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1785 			    __func__, snd_una, toep->tid, tp->snd_una);
1786 		}
1787 #endif
1788 
1789 		if (tp->snd_una != snd_una) {
1790 			tp->snd_una = snd_una;
1791 			tp->ts_recent_age = tcp_ts_getticks();
1792 		}
1793 	}
1794 
1795 #ifdef VERBOSE_TRACES
1796 	CTR3(KTR_CXGBE, "%s: tid %d credits %u", __func__, tid, credits);
1797 #endif
1798 	so = inp->inp_socket;
1799 	txsd = &toep->txsd[toep->txsd_cidx];
1800 	plen = 0;
1801 	while (credits) {
1802 		KASSERT(credits >= txsd->tx_credits,
1803 		    ("%s: too many (or partial) credits", __func__));
1804 		credits -= txsd->tx_credits;
1805 		toep->tx_credits += txsd->tx_credits;
1806 		plen += txsd->plen;
1807 		if (txsd->iv_buffer) {
1808 			free(txsd->iv_buffer, M_CXGBE);
1809 			txsd->iv_buffer = NULL;
1810 		}
1811 		txsd++;
1812 		toep->txsd_avail++;
1813 		KASSERT(toep->txsd_avail <= toep->txsd_total,
1814 		    ("%s: txsd avail > total", __func__));
1815 		if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1816 			txsd = &toep->txsd[0];
1817 			toep->txsd_cidx = 0;
1818 		}
1819 	}
1820 
1821 	if (toep->tx_credits == toep->tx_total) {
1822 		toep->tx_nocompl = 0;
1823 		toep->plen_nocompl = 0;
1824 	}
1825 
1826 	if (toep->flags & TPF_TX_SUSPENDED &&
1827 	    toep->tx_credits >= toep->tx_total / 4) {
1828 #ifdef VERBOSE_TRACES
1829 		CTR2(KTR_CXGBE, "%s: tid %d calling t4_push_frames", __func__,
1830 		    tid);
1831 #endif
1832 		toep->flags &= ~TPF_TX_SUSPENDED;
1833 		CURVNET_SET(toep->vnet);
1834 		if (toep->ulp_mode == ULP_MODE_ISCSI)
1835 			t4_push_pdus(sc, toep, plen);
1836 		else if (tls_tx_key(toep))
1837 			t4_push_tls_records(sc, toep, plen);
1838 		else
1839 			t4_push_frames(sc, toep, plen);
1840 		CURVNET_RESTORE();
1841 	} else if (plen > 0) {
1842 		struct sockbuf *sb = &so->so_snd;
1843 		int sbu;
1844 
1845 		SOCKBUF_LOCK(sb);
1846 		sbu = sbused(sb);
1847 		if (toep->ulp_mode == ULP_MODE_ISCSI) {
1848 
1849 			if (__predict_false(sbu > 0)) {
1850 				/*
1851 				 * The data trasmitted before the tid's ULP mode
1852 				 * changed to ISCSI is still in so_snd.
1853 				 * Incoming credits should account for so_snd
1854 				 * first.
1855 				 */
1856 				sbdrop_locked(sb, min(sbu, plen));
1857 				plen -= min(sbu, plen);
1858 			}
1859 			sowwakeup_locked(so);	/* unlocks so_snd */
1860 			rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1861 		} else {
1862 #ifdef VERBOSE_TRACES
1863 			CTR3(KTR_CXGBE, "%s: tid %d dropped %d bytes", __func__,
1864 			    tid, plen);
1865 #endif
1866 			sbdrop_locked(sb, plen);
1867 			if (tls_tx_key(toep)) {
1868 				struct tls_ofld_info *tls_ofld = &toep->tls;
1869 
1870 				MPASS(tls_ofld->sb_off >= plen);
1871 				tls_ofld->sb_off -= plen;
1872 			}
1873 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
1874 				t4_aiotx_queue_toep(toep);
1875 			sowwakeup_locked(so);	/* unlocks so_snd */
1876 		}
1877 		SOCKBUF_UNLOCK_ASSERT(sb);
1878 	}
1879 
1880 	INP_WUNLOCK(inp);
1881 
1882 	return (0);
1883 }
1884 
1885 void
1886 t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, struct toepcb *toep,
1887     uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie)
1888 {
1889 	struct wrqe *wr;
1890 	struct cpl_set_tcb_field *req;
1891 	struct ofld_tx_sdesc *txsd;
1892 
1893 	MPASS((cookie & ~M_COOKIE) == 0);
1894 	if (reply) {
1895 		MPASS(cookie != CPL_COOKIE_RESERVED);
1896 	}
1897 
1898 	wr = alloc_wrqe(sizeof(*req), wrq);
1899 	if (wr == NULL) {
1900 		/* XXX */
1901 		panic("%s: allocation failure.", __func__);
1902 	}
1903 	req = wrtod(wr);
1904 
1905 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1906 	req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1907 	if (reply == 0)
1908 		req->reply_ctrl |= htobe16(F_NO_REPLY);
1909 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie));
1910 	req->mask = htobe64(mask);
1911 	req->val = htobe64(val);
1912 	if ((wrq->eq.flags & EQ_TYPEMASK) == EQ_OFLD) {
1913 		txsd = &toep->txsd[toep->txsd_pidx];
1914 		txsd->tx_credits = howmany(sizeof(*req), 16);
1915 		txsd->plen = 0;
1916 		KASSERT(toep->tx_credits >= txsd->tx_credits &&
1917 		    toep->txsd_avail > 0,
1918 		    ("%s: not enough credits (%d)", __func__,
1919 		    toep->tx_credits));
1920 		toep->tx_credits -= txsd->tx_credits;
1921 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1922 			toep->txsd_pidx = 0;
1923 		toep->txsd_avail--;
1924 	}
1925 
1926 	t4_wrq_tx(sc, wr);
1927 }
1928 
1929 void
1930 t4_init_cpl_io_handlers(void)
1931 {
1932 
1933 	t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close);
1934 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl);
1935 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req);
1936 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl,
1937 	    CPL_COOKIE_TOM);
1938 	t4_register_cpl_handler(CPL_RX_DATA, do_rx_data);
1939 	t4_register_shared_cpl_handler(CPL_FW4_ACK, do_fw4_ack, CPL_COOKIE_TOM);
1940 }
1941 
1942 void
1943 t4_uninit_cpl_io_handlers(void)
1944 {
1945 
1946 	t4_register_cpl_handler(CPL_PEER_CLOSE, NULL);
1947 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, NULL);
1948 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, NULL);
1949 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, NULL, CPL_COOKIE_TOM);
1950 	t4_register_cpl_handler(CPL_RX_DATA, NULL);
1951 	t4_register_shared_cpl_handler(CPL_FW4_ACK, NULL, CPL_COOKIE_TOM);
1952 }
1953 
1954 /*
1955  * Use the 'backend3' field in AIO jobs to store the amount of data
1956  * sent by the AIO job so far and the 'backend4' field to hold an
1957  * error that should be reported when the job is completed.
1958  */
1959 #define	aio_sent	backend3
1960 #define	aio_error	backend4
1961 
1962 #define	jobtotid(job)							\
1963 	(((struct toepcb *)(so_sototcpcb((job)->fd_file->f_data)->t_toe))->tid)
1964 
1965 static void
1966 free_aiotx_buffer(struct aiotx_buffer *ab)
1967 {
1968 	struct kaiocb *job;
1969 	long status;
1970 	int error;
1971 
1972 	if (refcount_release(&ab->refcount) == 0)
1973 		return;
1974 
1975 	job = ab->job;
1976 	error = job->aio_error;
1977 	status = job->aio_sent;
1978 	vm_page_unhold_pages(ab->ps.pages, ab->ps.npages);
1979 	free(ab, M_CXGBE);
1980 #ifdef VERBOSE_TRACES
1981 	CTR5(KTR_CXGBE, "%s: tid %d completed %p len %ld, error %d", __func__,
1982 	    jobtotid(job), job, status, error);
1983 #endif
1984 	if (error == ECANCELED && status != 0)
1985 		error = 0;
1986 	if (error == ECANCELED)
1987 		aio_cancel(job);
1988 	else if (error)
1989 		aio_complete(job, -1, error);
1990 	else
1991 		aio_complete(job, status, 0);
1992 }
1993 
1994 static void
1995 t4_aiotx_mbuf_free(struct mbuf *m)
1996 {
1997 	struct aiotx_buffer *ab = m->m_ext.ext_arg1;
1998 
1999 #ifdef VERBOSE_TRACES
2000 	CTR3(KTR_CXGBE, "%s: completed %d bytes for tid %d", __func__,
2001 	    m->m_len, jobtotid(ab->job));
2002 #endif
2003 	free_aiotx_buffer(ab);
2004 }
2005 
2006 /*
2007  * Hold the buffer backing an AIO request and return an AIO transmit
2008  * buffer.
2009  */
2010 static int
2011 hold_aio(struct kaiocb *job)
2012 {
2013 	struct aiotx_buffer *ab;
2014 	struct vmspace *vm;
2015 	vm_map_t map;
2016 	vm_offset_t start, end, pgoff;
2017 	int n;
2018 
2019 	MPASS(job->backend1 == NULL);
2020 
2021 	/*
2022 	 * The AIO subsystem will cancel and drain all requests before
2023 	 * permitting a process to exit or exec, so p_vmspace should
2024 	 * be stable here.
2025 	 */
2026 	vm = job->userproc->p_vmspace;
2027 	map = &vm->vm_map;
2028 	start = (uintptr_t)job->uaiocb.aio_buf;
2029 	pgoff = start & PAGE_MASK;
2030 	end = round_page(start + job->uaiocb.aio_nbytes);
2031 	start = trunc_page(start);
2032 	n = atop(end - start);
2033 
2034 	ab = malloc(sizeof(*ab) + n * sizeof(vm_page_t), M_CXGBE, M_WAITOK |
2035 	    M_ZERO);
2036 	refcount_init(&ab->refcount, 1);
2037 	ab->ps.pages = (vm_page_t *)(ab + 1);
2038 	ab->ps.npages = vm_fault_quick_hold_pages(map, start, end - start,
2039 	    VM_PROT_WRITE, ab->ps.pages, n);
2040 	if (ab->ps.npages < 0) {
2041 		free(ab, M_CXGBE);
2042 		return (EFAULT);
2043 	}
2044 
2045 	KASSERT(ab->ps.npages == n,
2046 	    ("hold_aio: page count mismatch: %d vs %d", ab->ps.npages, n));
2047 
2048 	ab->ps.offset = pgoff;
2049 	ab->ps.len = job->uaiocb.aio_nbytes;
2050 	ab->job = job;
2051 	job->backend1 = ab;
2052 #ifdef VERBOSE_TRACES
2053 	CTR5(KTR_CXGBE, "%s: tid %d, new pageset %p for job %p, npages %d",
2054 	    __func__, jobtotid(job), &ab->ps, job, ab->ps.npages);
2055 #endif
2056 	return (0);
2057 }
2058 
2059 static void
2060 t4_aiotx_process_job(struct toepcb *toep, struct socket *so, struct kaiocb *job)
2061 {
2062 	struct adapter *sc;
2063 	struct sockbuf *sb;
2064 	struct file *fp;
2065 	struct aiotx_buffer *ab;
2066 	struct inpcb *inp;
2067 	struct tcpcb *tp;
2068 	struct mbuf *m;
2069 	int error;
2070 	bool moretocome, sendmore;
2071 
2072 	sc = td_adapter(toep->td);
2073 	sb = &so->so_snd;
2074 	SOCKBUF_UNLOCK(sb);
2075 	fp = job->fd_file;
2076 	ab = job->backend1;
2077 	m = NULL;
2078 
2079 #ifdef MAC
2080 	error = mac_socket_check_send(fp->f_cred, so);
2081 	if (error != 0)
2082 		goto out;
2083 #endif
2084 
2085 	if (ab == NULL) {
2086 		error = hold_aio(job);
2087 		if (error != 0)
2088 			goto out;
2089 		ab = job->backend1;
2090 	}
2091 
2092 	/* Inline sosend_generic(). */
2093 
2094 	job->msgsnd = 1;
2095 
2096 	error = sblock(sb, SBL_WAIT);
2097 	MPASS(error == 0);
2098 
2099 sendanother:
2100 	m = m_get(M_WAITOK, MT_DATA);
2101 
2102 	SOCKBUF_LOCK(sb);
2103 	if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
2104 		SOCKBUF_UNLOCK(sb);
2105 		sbunlock(sb);
2106 		if ((so->so_options & SO_NOSIGPIPE) == 0) {
2107 			PROC_LOCK(job->userproc);
2108 			kern_psignal(job->userproc, SIGPIPE);
2109 			PROC_UNLOCK(job->userproc);
2110 		}
2111 		error = EPIPE;
2112 		goto out;
2113 	}
2114 	if (so->so_error) {
2115 		error = so->so_error;
2116 		so->so_error = 0;
2117 		SOCKBUF_UNLOCK(sb);
2118 		sbunlock(sb);
2119 		goto out;
2120 	}
2121 	if ((so->so_state & SS_ISCONNECTED) == 0) {
2122 		SOCKBUF_UNLOCK(sb);
2123 		sbunlock(sb);
2124 		error = ENOTCONN;
2125 		goto out;
2126 	}
2127 	if (sbspace(sb) < sb->sb_lowat) {
2128 		MPASS(job->aio_sent == 0 || !(so->so_state & SS_NBIO));
2129 
2130 		/*
2131 		 * Don't block if there is too little room in the socket
2132 		 * buffer.  Instead, requeue the request.
2133 		 */
2134 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2135 			SOCKBUF_UNLOCK(sb);
2136 			sbunlock(sb);
2137 			error = ECANCELED;
2138 			goto out;
2139 		}
2140 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2141 		SOCKBUF_UNLOCK(sb);
2142 		sbunlock(sb);
2143 		goto out;
2144 	}
2145 
2146 	/*
2147 	 * Write as much data as the socket permits, but no more than a
2148 	 * a single sndbuf at a time.
2149 	 */
2150 	m->m_len = sbspace(sb);
2151 	if (m->m_len > ab->ps.len - job->aio_sent) {
2152 		m->m_len = ab->ps.len - job->aio_sent;
2153 		moretocome = false;
2154 	} else
2155 		moretocome = true;
2156 	if (m->m_len > sc->tt.sndbuf) {
2157 		m->m_len = sc->tt.sndbuf;
2158 		sendmore = true;
2159 	} else
2160 		sendmore = false;
2161 
2162 	if (!TAILQ_EMPTY(&toep->aiotx_jobq))
2163 		moretocome = true;
2164 	SOCKBUF_UNLOCK(sb);
2165 	MPASS(m->m_len != 0);
2166 
2167 	/* Inlined tcp_usr_send(). */
2168 
2169 	inp = toep->inp;
2170 	INP_WLOCK(inp);
2171 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
2172 		INP_WUNLOCK(inp);
2173 		sbunlock(sb);
2174 		error = ECONNRESET;
2175 		goto out;
2176 	}
2177 
2178 	refcount_acquire(&ab->refcount);
2179 	m_extadd(m, NULL, ab->ps.len, t4_aiotx_mbuf_free, ab,
2180 	    (void *)(uintptr_t)job->aio_sent, 0, EXT_NET_DRV);
2181 	m->m_ext.ext_flags |= EXT_FLAG_AIOTX;
2182 	job->aio_sent += m->m_len;
2183 
2184 	sbappendstream(sb, m, 0);
2185 	m = NULL;
2186 
2187 	if (!(inp->inp_flags & INP_DROPPED)) {
2188 		tp = intotcpcb(inp);
2189 		if (moretocome)
2190 			tp->t_flags |= TF_MORETOCOME;
2191 		error = tp->t_fb->tfb_tcp_output(tp);
2192 		if (moretocome)
2193 			tp->t_flags &= ~TF_MORETOCOME;
2194 	}
2195 
2196 	INP_WUNLOCK(inp);
2197 	if (sendmore)
2198 		goto sendanother;
2199 	sbunlock(sb);
2200 
2201 	if (error)
2202 		goto out;
2203 
2204 	/*
2205 	 * If this is a non-blocking socket and the request has not
2206 	 * been fully completed, requeue it until the socket is ready
2207 	 * again.
2208 	 */
2209 	if (job->aio_sent < job->uaiocb.aio_nbytes &&
2210 	    !(so->so_state & SS_NBIO)) {
2211 		SOCKBUF_LOCK(sb);
2212 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2213 			SOCKBUF_UNLOCK(sb);
2214 			error = ECANCELED;
2215 			goto out;
2216 		}
2217 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2218 		return;
2219 	}
2220 
2221 	/*
2222 	 * If the request will not be requeued, drop a reference on
2223 	 * the aiotx buffer.  Any mbufs in flight should still
2224 	 * contain a reference, but this drops the reference that the
2225 	 * job owns while it is waiting to queue mbufs to the socket.
2226 	 */
2227 	free_aiotx_buffer(ab);
2228 
2229 out:
2230 	if (error) {
2231 		if (ab != NULL) {
2232 			job->aio_error = error;
2233 			free_aiotx_buffer(ab);
2234 		} else {
2235 			MPASS(job->aio_sent == 0);
2236 			aio_complete(job, -1, error);
2237 		}
2238 	}
2239 	if (m != NULL)
2240 		m_free(m);
2241 	SOCKBUF_LOCK(sb);
2242 }
2243 
2244 static void
2245 t4_aiotx_task(void *context, int pending)
2246 {
2247 	struct toepcb *toep = context;
2248 	struct inpcb *inp = toep->inp;
2249 	struct socket *so = inp->inp_socket;
2250 	struct kaiocb *job;
2251 
2252 	CURVNET_SET(toep->vnet);
2253 	SOCKBUF_LOCK(&so->so_snd);
2254 	while (!TAILQ_EMPTY(&toep->aiotx_jobq) && sowriteable(so)) {
2255 		job = TAILQ_FIRST(&toep->aiotx_jobq);
2256 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2257 		if (!aio_clear_cancel_function(job))
2258 			continue;
2259 
2260 		t4_aiotx_process_job(toep, so, job);
2261 	}
2262 	toep->aiotx_task_active = false;
2263 	SOCKBUF_UNLOCK(&so->so_snd);
2264 	CURVNET_RESTORE();
2265 
2266 	free_toepcb(toep);
2267 }
2268 
2269 static void
2270 t4_aiotx_queue_toep(struct toepcb *toep)
2271 {
2272 
2273 	SOCKBUF_LOCK_ASSERT(&toep->inp->inp_socket->so_snd);
2274 #ifdef VERBOSE_TRACES
2275 	CTR3(KTR_CXGBE, "%s: queueing aiotx task for tid %d, active = %s",
2276 	    __func__, toep->tid, toep->aiotx_task_active ? "true" : "false");
2277 #endif
2278 	if (toep->aiotx_task_active)
2279 		return;
2280 	toep->aiotx_task_active = true;
2281 	hold_toepcb(toep);
2282 	soaio_enqueue(&toep->aiotx_task);
2283 }
2284 
2285 static void
2286 t4_aiotx_cancel(struct kaiocb *job)
2287 {
2288 	struct aiotx_buffer *ab;
2289 	struct socket *so;
2290 	struct sockbuf *sb;
2291 	struct tcpcb *tp;
2292 	struct toepcb *toep;
2293 
2294 	so = job->fd_file->f_data;
2295 	tp = so_sototcpcb(so);
2296 	toep = tp->t_toe;
2297 	MPASS(job->uaiocb.aio_lio_opcode == LIO_WRITE);
2298 	sb = &so->so_snd;
2299 
2300 	SOCKBUF_LOCK(sb);
2301 	if (!aio_cancel_cleared(job))
2302 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2303 	SOCKBUF_UNLOCK(sb);
2304 
2305 	ab = job->backend1;
2306 	if (ab != NULL)
2307 		free_aiotx_buffer(ab);
2308 	else
2309 		aio_cancel(job);
2310 }
2311 
2312 int
2313 t4_aio_queue_aiotx(struct socket *so, struct kaiocb *job)
2314 {
2315 	struct tcpcb *tp = so_sototcpcb(so);
2316 	struct toepcb *toep = tp->t_toe;
2317 	struct adapter *sc = td_adapter(toep->td);
2318 
2319 	/* This only handles writes. */
2320 	if (job->uaiocb.aio_lio_opcode != LIO_WRITE)
2321 		return (EOPNOTSUPP);
2322 
2323 	if (!sc->tt.tx_zcopy)
2324 		return (EOPNOTSUPP);
2325 
2326 	if (tls_tx_key(toep))
2327 		return (EOPNOTSUPP);
2328 
2329 	SOCKBUF_LOCK(&so->so_snd);
2330 #ifdef VERBOSE_TRACES
2331 	CTR2(KTR_CXGBE, "%s: queueing %p", __func__, job);
2332 #endif
2333 	if (!aio_set_cancel_function(job, t4_aiotx_cancel))
2334 		panic("new job was cancelled");
2335 	TAILQ_INSERT_TAIL(&toep->aiotx_jobq, job, list);
2336 	if (sowriteable(so))
2337 		t4_aiotx_queue_toep(toep);
2338 	SOCKBUF_UNLOCK(&so->so_snd);
2339 	return (0);
2340 }
2341 
2342 void
2343 aiotx_init_toep(struct toepcb *toep)
2344 {
2345 
2346 	TAILQ_INIT(&toep->aiotx_jobq);
2347 	TASK_INIT(&toep->aiotx_task, 0, t4_aiotx_task, toep);
2348 }
2349 #endif
2350