xref: /freebsd/sys/dev/cxgbe/tom/t4_cpl_io.c (revision 77013d11e6483b970af25e13c9b892075742f7e5)
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->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->wrq);
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->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),
206 		    &toep->ofld_txq->wrq)) == NULL) {
207 			if (tc_idx >= 0)
208 				t4_release_cl_rl(sc, port_id, tc_idx);
209 			return (ENOMEM);
210 		}
211 
212 		flowc = wrtod(wr);
213 		memset(flowc, 0, wr->wr_len);
214 
215 		flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
216 		    V_FW_FLOWC_WR_NPARAMS(nparams));
217 		flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
218 		    V_FW_WR_FLOWID(toep->tid));
219 
220 		flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
221 		if (tc_idx == -1)
222 			flowc->mnemval[0].val = htobe32(0xff);
223 		else
224 			flowc->mnemval[0].val = htobe32(tc_idx);
225 
226 		txsd->tx_credits = flowclen16;
227 		txsd->plen = 0;
228 		toep->tx_credits -= txsd->tx_credits;
229 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
230 			toep->txsd_pidx = 0;
231 		toep->txsd_avail--;
232 		t4_wrq_tx(sc, wr);
233 	}
234 
235 	if (toep->params.tc_idx >= 0)
236 		t4_release_cl_rl(sc, port_id, toep->params.tc_idx);
237 	toep->params.tc_idx = tc_idx;
238 
239 	return (0);
240 }
241 #endif
242 
243 void
244 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
245 {
246 	struct wrqe *wr;
247 	struct cpl_abort_req *req;
248 	int tid = toep->tid;
249 	struct inpcb *inp = toep->inp;
250 	struct tcpcb *tp = intotcpcb(inp);	/* don't use if INP_DROPPED */
251 
252 	INP_WLOCK_ASSERT(inp);
253 
254 	CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
255 	    __func__, toep->tid,
256 	    inp->inp_flags & INP_DROPPED ? "inp dropped" :
257 	    tcpstates[tp->t_state],
258 	    toep->flags, inp->inp_flags,
259 	    toep->flags & TPF_ABORT_SHUTDOWN ?
260 	    " (abort already in progress)" : "");
261 
262 	if (toep->flags & TPF_ABORT_SHUTDOWN)
263 		return;	/* abort already in progress */
264 
265 	toep->flags |= TPF_ABORT_SHUTDOWN;
266 
267 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
268 	    ("%s: flowc_wr not sent for tid %d.", __func__, tid));
269 
270 	wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq);
271 	if (wr == NULL) {
272 		/* XXX */
273 		panic("%s: allocation failure.", __func__);
274 	}
275 	req = wrtod(wr);
276 
277 	INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
278 	if (inp->inp_flags & INP_DROPPED)
279 		req->rsvd0 = htobe32(snd_nxt);
280 	else
281 		req->rsvd0 = htobe32(tp->snd_nxt);
282 	req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
283 	req->cmd = CPL_ABORT_SEND_RST;
284 
285 	/*
286 	 * XXX: What's the correct way to tell that the inp hasn't been detached
287 	 * from its socket?  Should I even be flushing the snd buffer here?
288 	 */
289 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
290 		struct socket *so = inp->inp_socket;
291 
292 		if (so != NULL)	/* because I'm not sure.  See comment above */
293 			sbflush(&so->so_snd);
294 	}
295 
296 	t4_l2t_send(sc, wr, toep->l2te);
297 }
298 
299 /*
300  * Called when a connection is established to translate the TCP options
301  * reported by HW to FreeBSD's native format.
302  */
303 static void
304 assign_rxopt(struct tcpcb *tp, uint16_t opt)
305 {
306 	struct toepcb *toep = tp->t_toe;
307 	struct inpcb *inp = tp->t_inpcb;
308 	struct adapter *sc = td_adapter(toep->td);
309 
310 	INP_LOCK_ASSERT(inp);
311 
312 	toep->params.mtu_idx = G_TCPOPT_MSS(opt);
313 	tp->t_maxseg = sc->params.mtus[toep->params.mtu_idx];
314 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
315 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
316 	else
317 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
318 
319 	toep->params.emss = tp->t_maxseg;
320 	if (G_TCPOPT_TSTAMP(opt)) {
321 		toep->params.tstamp = 1;
322 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
323 		tp->t_flags |= TF_RCVD_TSTMP;	/* timestamps ok */
324 		tp->ts_recent = 0;		/* hmmm */
325 		tp->ts_recent_age = tcp_ts_getticks();
326 	} else
327 		toep->params.tstamp = 0;
328 
329 	if (G_TCPOPT_SACK(opt)) {
330 		toep->params.sack = 1;
331 		tp->t_flags |= TF_SACK_PERMIT;	/* should already be set */
332 	} else {
333 		toep->params.sack = 0;
334 		tp->t_flags &= ~TF_SACK_PERMIT;	/* sack disallowed by peer */
335 	}
336 
337 	if (G_TCPOPT_WSCALE_OK(opt))
338 		tp->t_flags |= TF_RCVD_SCALE;
339 
340 	/* Doing window scaling? */
341 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
342 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
343 		tp->rcv_scale = tp->request_r_scale;
344 		tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
345 	} else
346 		toep->params.wscale = 0;
347 
348 	CTR6(KTR_CXGBE,
349 	    "assign_rxopt: tid %d, mtu_idx %u, emss %u, ts %u, sack %u, wscale %u",
350 	    toep->tid, toep->params.mtu_idx, toep->params.emss,
351 	    toep->params.tstamp, toep->params.sack, toep->params.wscale);
352 }
353 
354 /*
355  * Completes some final bits of initialization for just established connections
356  * and changes their state to TCPS_ESTABLISHED.
357  *
358  * The ISNs are from the exchange of SYNs.
359  */
360 void
361 make_established(struct toepcb *toep, uint32_t iss, uint32_t irs, uint16_t opt)
362 {
363 	struct inpcb *inp = toep->inp;
364 	struct socket *so = inp->inp_socket;
365 	struct tcpcb *tp = intotcpcb(inp);
366 	uint16_t tcpopt = be16toh(opt);
367 
368 	INP_WLOCK_ASSERT(inp);
369 	KASSERT(tp->t_state == TCPS_SYN_SENT ||
370 	    tp->t_state == TCPS_SYN_RECEIVED,
371 	    ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
372 
373 	CTR6(KTR_CXGBE, "%s: tid %d, so %p, inp %p, tp %p, toep %p",
374 	    __func__, toep->tid, so, inp, tp, toep);
375 
376 	tcp_state_change(tp, TCPS_ESTABLISHED);
377 	tp->t_starttime = ticks;
378 	TCPSTAT_INC(tcps_connects);
379 
380 	tp->irs = irs;
381 	tcp_rcvseqinit(tp);
382 	tp->rcv_wnd = (u_int)toep->params.opt0_bufsize << 10;
383 	tp->rcv_adv += tp->rcv_wnd;
384 	tp->last_ack_sent = tp->rcv_nxt;
385 
386 	tp->iss = iss;
387 	tcp_sendseqinit(tp);
388 	tp->snd_una = iss + 1;
389 	tp->snd_nxt = iss + 1;
390 	tp->snd_max = iss + 1;
391 
392 	assign_rxopt(tp, tcpopt);
393 	send_flowc_wr(toep, tp);
394 
395 	soisconnected(so);
396 
397 	if (ulp_mode(toep) == ULP_MODE_TLS)
398 		tls_establish(toep);
399 }
400 
401 int
402 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
403 {
404 	struct wrqe *wr;
405 	struct cpl_rx_data_ack *req;
406 	uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
407 
408 	KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
409 
410 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
411 	if (wr == NULL)
412 		return (0);
413 	req = wrtod(wr);
414 
415 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
416 	req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
417 
418 	t4_wrq_tx(sc, wr);
419 	return (credits);
420 }
421 
422 void
423 send_rx_modulate(struct adapter *sc, struct toepcb *toep)
424 {
425 	struct wrqe *wr;
426 	struct cpl_rx_data_ack *req;
427 
428 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
429 	if (wr == NULL)
430 		return;
431 	req = wrtod(wr);
432 
433 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
434 	req->credit_dack = htobe32(F_RX_MODULATE_RX);
435 
436 	t4_wrq_tx(sc, wr);
437 }
438 
439 void
440 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp)
441 {
442 	struct adapter *sc = tod->tod_softc;
443 	struct inpcb *inp = tp->t_inpcb;
444 	struct socket *so = inp->inp_socket;
445 	struct sockbuf *sb = &so->so_rcv;
446 	struct toepcb *toep = tp->t_toe;
447 	int rx_credits;
448 
449 	INP_WLOCK_ASSERT(inp);
450 	SOCKBUF_LOCK_ASSERT(sb);
451 
452 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
453 	if (rx_credits > 0 &&
454 	    (tp->rcv_wnd <= 32 * 1024 || rx_credits >= 64 * 1024 ||
455 	    (rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
456 	    sbused(sb) + tp->rcv_wnd < sb->sb_lowat)) {
457 		rx_credits = send_rx_credits(sc, toep, rx_credits);
458 		tp->rcv_wnd += rx_credits;
459 		tp->rcv_adv += rx_credits;
460 	} else if (toep->flags & TPF_FORCE_CREDITS)
461 		send_rx_modulate(sc, toep);
462 }
463 
464 void
465 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
466 {
467 	struct inpcb *inp = tp->t_inpcb;
468 	struct socket *so = inp->inp_socket;
469 	struct sockbuf *sb = &so->so_rcv;
470 
471 	SOCKBUF_LOCK(sb);
472 	t4_rcvd_locked(tod, tp);
473 	SOCKBUF_UNLOCK(sb);
474 }
475 
476 /*
477  * Close a connection by sending a CPL_CLOSE_CON_REQ message.
478  */
479 int
480 t4_close_conn(struct adapter *sc, struct toepcb *toep)
481 {
482 	struct wrqe *wr;
483 	struct cpl_close_con_req *req;
484 	unsigned int tid = toep->tid;
485 
486 	CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
487 	    toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
488 
489 	if (toep->flags & TPF_FIN_SENT)
490 		return (0);
491 
492 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
493 	    ("%s: flowc_wr not sent for tid %u.", __func__, tid));
494 
495 	wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq);
496 	if (wr == NULL) {
497 		/* XXX */
498 		panic("%s: allocation failure.", __func__);
499 	}
500 	req = wrtod(wr);
501 
502         req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
503 	    V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
504 	req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
505 	    V_FW_WR_FLOWID(tid));
506         req->wr.wr_lo = cpu_to_be64(0);
507         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
508 	req->rsvd = 0;
509 
510 	toep->flags |= TPF_FIN_SENT;
511 	toep->flags &= ~TPF_SEND_FIN;
512 	t4_l2t_send(sc, wr, toep->l2te);
513 
514 	return (0);
515 }
516 
517 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
518 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
519 
520 /* Maximum amount of immediate data we could stuff in a WR */
521 static inline int
522 max_imm_payload(int tx_credits)
523 {
524 	const int n = 1;	/* Use no more than one desc for imm. data WR */
525 
526 	KASSERT(tx_credits >= 0 &&
527 		tx_credits <= MAX_OFLD_TX_CREDITS,
528 		("%s: %d credits", __func__, tx_credits));
529 
530 	if (tx_credits < MIN_OFLD_TX_CREDITS)
531 		return (0);
532 
533 	if (tx_credits >= (n * EQ_ESIZE) / 16)
534 		return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr));
535 	else
536 		return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr));
537 }
538 
539 /* Maximum number of SGL entries we could stuff in a WR */
540 static inline int
541 max_dsgl_nsegs(int tx_credits)
542 {
543 	int nseg = 1;	/* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
544 	int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS;
545 
546 	KASSERT(tx_credits >= 0 &&
547 		tx_credits <= MAX_OFLD_TX_CREDITS,
548 		("%s: %d credits", __func__, tx_credits));
549 
550 	if (tx_credits < MIN_OFLD_TX_CREDITS)
551 		return (0);
552 
553 	nseg += 2 * (sge_pair_credits * 16 / 24);
554 	if ((sge_pair_credits * 16) % 24 == 16)
555 		nseg++;
556 
557 	return (nseg);
558 }
559 
560 static inline void
561 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen,
562     unsigned int plen, uint8_t credits, int shove, int ulp_submode)
563 {
564 	struct fw_ofld_tx_data_wr *txwr = dst;
565 
566 	txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) |
567 	    V_FW_WR_IMMDLEN(immdlen));
568 	txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
569 	    V_FW_WR_LEN16(credits));
570 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ulp_mode(toep)) |
571 	    V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
572 	txwr->plen = htobe32(plen);
573 
574 	if (toep->params.tx_align > 0) {
575 		if (plen < 2 * toep->params.emss)
576 			txwr->lsodisable_to_flags |=
577 			    htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
578 		else
579 			txwr->lsodisable_to_flags |=
580 			    htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
581 				(toep->params.nagle == 0 ? 0 :
582 				F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
583 	}
584 }
585 
586 /*
587  * Generate a DSGL from a starting mbuf.  The total number of segments and the
588  * maximum segments in any one mbuf are provided.
589  */
590 static void
591 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
592 {
593 	struct mbuf *m;
594 	struct ulptx_sgl *usgl = dst;
595 	int i, j, rc;
596 	struct sglist sg;
597 	struct sglist_seg segs[n];
598 
599 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
600 
601 	sglist_init(&sg, n, segs);
602 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
603 	    V_ULPTX_NSGE(nsegs));
604 
605 	i = -1;
606 	for (m = start; m != stop; m = m->m_next) {
607 		if (m->m_flags & M_EXTPG)
608 			rc = sglist_append_mbuf_epg(&sg, m,
609 			    mtod(m, vm_offset_t), m->m_len);
610 		else
611 			rc = sglist_append(&sg, mtod(m, void *), m->m_len);
612 		if (__predict_false(rc != 0))
613 			panic("%s: sglist_append %d", __func__, rc);
614 
615 		for (j = 0; j < sg.sg_nseg; i++, j++) {
616 			if (i < 0) {
617 				usgl->len0 = htobe32(segs[j].ss_len);
618 				usgl->addr0 = htobe64(segs[j].ss_paddr);
619 			} else {
620 				usgl->sge[i / 2].len[i & 1] =
621 				    htobe32(segs[j].ss_len);
622 				usgl->sge[i / 2].addr[i & 1] =
623 				    htobe64(segs[j].ss_paddr);
624 			}
625 #ifdef INVARIANTS
626 			nsegs--;
627 #endif
628 		}
629 		sglist_reset(&sg);
630 	}
631 	if (i & 1)
632 		usgl->sge[i / 2].len[1] = htobe32(0);
633 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
634 	    __func__, nsegs, start, stop));
635 }
636 
637 /*
638  * Max number of SGL entries an offload tx work request can have.  This is 41
639  * (1 + 40) for a full 512B work request.
640  * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
641  */
642 #define OFLD_SGL_LEN (41)
643 
644 /*
645  * Send data and/or a FIN to the peer.
646  *
647  * The socket's so_snd buffer consists of a stream of data starting with sb_mb
648  * and linked together with m_next.  sb_sndptr, if set, is the last mbuf that
649  * was transmitted.
650  *
651  * drop indicates the number of bytes that should be dropped from the head of
652  * the send buffer.  It is an optimization that lets do_fw4_ack avoid creating
653  * contention on the send buffer lock (before this change it used to do
654  * sowwakeup and then t4_push_frames right after that when recovering from tx
655  * stalls).  When drop is set this function MUST drop the bytes and wake up any
656  * writers.
657  */
658 void
659 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
660 {
661 	struct mbuf *sndptr, *m, *sb_sndptr;
662 	struct fw_ofld_tx_data_wr *txwr;
663 	struct wrqe *wr;
664 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
665 	struct inpcb *inp = toep->inp;
666 	struct tcpcb *tp = intotcpcb(inp);
667 	struct socket *so = inp->inp_socket;
668 	struct sockbuf *sb = &so->so_snd;
669 	int tx_credits, shove, compl, sowwakeup;
670 	struct ofld_tx_sdesc *txsd;
671 	bool nomap_mbuf_seen;
672 
673 	INP_WLOCK_ASSERT(inp);
674 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
675 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
676 
677 	KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
678 	    ulp_mode(toep) == ULP_MODE_TCPDDP ||
679 	    ulp_mode(toep) == ULP_MODE_TLS ||
680 	    ulp_mode(toep) == ULP_MODE_RDMA,
681 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
682 
683 #ifdef VERBOSE_TRACES
684 	CTR5(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
685 	    __func__, toep->tid, toep->flags, tp->t_flags, drop);
686 #endif
687 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
688 		return;
689 
690 #ifdef RATELIMIT
691 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
692 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
693 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
694 	}
695 #endif
696 
697 	/*
698 	 * This function doesn't resume by itself.  Someone else must clear the
699 	 * flag and call this function.
700 	 */
701 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
702 		KASSERT(drop == 0,
703 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
704 		return;
705 	}
706 
707 	txsd = &toep->txsd[toep->txsd_pidx];
708 	do {
709 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
710 		max_imm = max_imm_payload(tx_credits);
711 		max_nsegs = max_dsgl_nsegs(tx_credits);
712 
713 		SOCKBUF_LOCK(sb);
714 		sowwakeup = drop;
715 		if (drop) {
716 			sbdrop_locked(sb, drop);
717 			drop = 0;
718 		}
719 		sb_sndptr = sb->sb_sndptr;
720 		sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
721 		plen = 0;
722 		nsegs = 0;
723 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
724 		nomap_mbuf_seen = false;
725 		for (m = sndptr; m != NULL; m = m->m_next) {
726 			int n;
727 
728 			if ((m->m_flags & M_NOTAVAIL) != 0)
729 				break;
730 			if (m->m_flags & M_EXTPG) {
731 #ifdef KERN_TLS
732 				if (m->m_epg_tls != NULL) {
733 					toep->flags |= TPF_KTLS;
734 					if (plen == 0) {
735 						SOCKBUF_UNLOCK(sb);
736 						t4_push_ktls(sc, toep, 0);
737 						return;
738 					}
739 					break;
740 				}
741 #endif
742 				n = sglist_count_mbuf_epg(m,
743 				    mtod(m, vm_offset_t), m->m_len);
744 			} else
745 				n = sglist_count(mtod(m, void *), m->m_len);
746 
747 			nsegs += n;
748 			plen += m->m_len;
749 
750 			/* This mbuf sent us _over_ the nsegs limit, back out */
751 			if (plen > max_imm && nsegs > max_nsegs) {
752 				nsegs -= n;
753 				plen -= m->m_len;
754 				if (plen == 0) {
755 					/* Too few credits */
756 					toep->flags |= TPF_TX_SUSPENDED;
757 					if (sowwakeup) {
758 						if (!TAILQ_EMPTY(
759 						    &toep->aiotx_jobq))
760 							t4_aiotx_queue_toep(so,
761 							    toep);
762 						sowwakeup_locked(so);
763 					} else
764 						SOCKBUF_UNLOCK(sb);
765 					SOCKBUF_UNLOCK_ASSERT(sb);
766 					return;
767 				}
768 				break;
769 			}
770 
771 			if (m->m_flags & M_EXTPG)
772 				nomap_mbuf_seen = true;
773 			if (max_nsegs_1mbuf < n)
774 				max_nsegs_1mbuf = n;
775 			sb_sndptr = m;	/* new sb->sb_sndptr if all goes well */
776 
777 			/* This mbuf put us right at the max_nsegs limit */
778 			if (plen > max_imm && nsegs == max_nsegs) {
779 				m = m->m_next;
780 				break;
781 			}
782 		}
783 
784 		if (sbused(sb) > sb->sb_hiwat * 5 / 8 &&
785 		    toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
786 			compl = 1;
787 		else
788 			compl = 0;
789 
790 		if (sb->sb_flags & SB_AUTOSIZE &&
791 		    V_tcp_do_autosndbuf &&
792 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
793 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
794 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
795 			    V_tcp_autosndbuf_max);
796 
797 			if (!sbreserve_locked(sb, newsize, so, NULL))
798 				sb->sb_flags &= ~SB_AUTOSIZE;
799 			else
800 				sowwakeup = 1;	/* room available */
801 		}
802 		if (sowwakeup) {
803 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
804 				t4_aiotx_queue_toep(so, toep);
805 			sowwakeup_locked(so);
806 		} else
807 			SOCKBUF_UNLOCK(sb);
808 		SOCKBUF_UNLOCK_ASSERT(sb);
809 
810 		/* nothing to send */
811 		if (plen == 0) {
812 			KASSERT(m == NULL || (m->m_flags & M_NOTAVAIL) != 0,
813 			    ("%s: nothing to send, but m != NULL is ready",
814 			    __func__));
815 			break;
816 		}
817 
818 		if (__predict_false(toep->flags & TPF_FIN_SENT))
819 			panic("%s: excess tx.", __func__);
820 
821 		shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
822 		if (plen <= max_imm && !nomap_mbuf_seen) {
823 
824 			/* Immediate data tx */
825 
826 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
827 					&toep->ofld_txq->wrq);
828 			if (wr == NULL) {
829 				/* XXX: how will we recover from this? */
830 				toep->flags |= TPF_TX_SUSPENDED;
831 				return;
832 			}
833 			txwr = wrtod(wr);
834 			credits = howmany(wr->wr_len, 16);
835 			write_tx_wr(txwr, toep, plen, plen, credits, shove, 0);
836 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
837 			nsegs = 0;
838 		} else {
839 			int wr_len;
840 
841 			/* DSGL tx */
842 
843 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
844 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
845 			wr = alloc_wrqe(roundup2(wr_len, 16),
846 			    &toep->ofld_txq->wrq);
847 			if (wr == NULL) {
848 				/* XXX: how will we recover from this? */
849 				toep->flags |= TPF_TX_SUSPENDED;
850 				return;
851 			}
852 			txwr = wrtod(wr);
853 			credits = howmany(wr_len, 16);
854 			write_tx_wr(txwr, toep, 0, plen, credits, shove, 0);
855 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
856 			    max_nsegs_1mbuf);
857 			if (wr_len & 0xf) {
858 				uint64_t *pad = (uint64_t *)
859 				    ((uintptr_t)txwr + wr_len);
860 				*pad = 0;
861 			}
862 		}
863 
864 		KASSERT(toep->tx_credits >= credits,
865 			("%s: not enough credits", __func__));
866 
867 		toep->tx_credits -= credits;
868 		toep->tx_nocompl += credits;
869 		toep->plen_nocompl += plen;
870 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
871 		    toep->tx_nocompl >= toep->tx_total / 4)
872 			compl = 1;
873 
874 		if (compl || ulp_mode(toep) == ULP_MODE_RDMA) {
875 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
876 			toep->tx_nocompl = 0;
877 			toep->plen_nocompl = 0;
878 		}
879 
880 		tp->snd_nxt += plen;
881 		tp->snd_max += plen;
882 
883 		SOCKBUF_LOCK(sb);
884 		KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
885 		sb->sb_sndptr = sb_sndptr;
886 		SOCKBUF_UNLOCK(sb);
887 
888 		toep->flags |= TPF_TX_DATA_SENT;
889 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
890 			toep->flags |= TPF_TX_SUSPENDED;
891 
892 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
893 		txsd->plen = plen;
894 		txsd->tx_credits = credits;
895 		txsd++;
896 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
897 			toep->txsd_pidx = 0;
898 			txsd = &toep->txsd[0];
899 		}
900 		toep->txsd_avail--;
901 
902 		t4_l2t_send(sc, wr, toep->l2te);
903 	} while (m != NULL && (m->m_flags & M_NOTAVAIL) == 0);
904 
905 	/* Send a FIN if requested, but only if there's no more data to send */
906 	if (m == NULL && toep->flags & TPF_SEND_FIN)
907 		t4_close_conn(sc, toep);
908 }
909 
910 static inline void
911 rqdrop_locked(struct mbufq *q, int plen)
912 {
913 	struct mbuf *m;
914 
915 	while (plen > 0) {
916 		m = mbufq_dequeue(q);
917 
918 		/* Too many credits. */
919 		MPASS(m != NULL);
920 		M_ASSERTPKTHDR(m);
921 
922 		/* Partial credits. */
923 		MPASS(plen >= m->m_pkthdr.len);
924 
925 		plen -= m->m_pkthdr.len;
926 		m_freem(m);
927 	}
928 }
929 
930 static struct wrqe *
931 write_iscsi_mbuf_wr(struct toepcb *toep, struct mbuf *sndptr)
932 {
933 	struct mbuf *m;
934 	struct fw_ofld_tx_data_wr *txwr;
935 	struct wrqe *wr;
936 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
937 	u_int adjusted_plen, ulp_submode;
938 	struct inpcb *inp = toep->inp;
939 	struct tcpcb *tp = intotcpcb(inp);
940 	int tx_credits, shove;
941 	static const u_int ulp_extra_len[] = {0, 4, 4, 8};
942 
943 	M_ASSERTPKTHDR(sndptr);
944 
945 	tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
946 	if (mbuf_raw_wr(sndptr)) {
947 		plen = sndptr->m_pkthdr.len;
948 		KASSERT(plen <= SGE_MAX_WR_LEN,
949 		    ("raw WR len %u is greater than max WR len", plen));
950 		if (plen > tx_credits * 16)
951 			return (NULL);
952 
953 		wr = alloc_wrqe(roundup2(plen, 16), &toep->ofld_txq->wrq);
954 		if (__predict_false(wr == NULL))
955 			return (NULL);
956 
957 		m_copydata(sndptr, 0, plen, wrtod(wr));
958 		return (wr);
959 	}
960 
961 	max_imm = max_imm_payload(tx_credits);
962 	max_nsegs = max_dsgl_nsegs(tx_credits);
963 
964 	plen = 0;
965 	nsegs = 0;
966 	max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
967 	for (m = sndptr; m != NULL; m = m->m_next) {
968 		int n = sglist_count(mtod(m, void *), m->m_len);
969 
970 		nsegs += n;
971 		plen += m->m_len;
972 
973 		/*
974 		 * This mbuf would send us _over_ the nsegs limit.
975 		 * Suspend tx because the PDU can't be sent out.
976 		 */
977 		if (plen > max_imm && nsegs > max_nsegs)
978 			return (NULL);
979 
980 		if (max_nsegs_1mbuf < n)
981 			max_nsegs_1mbuf = n;
982 	}
983 
984 	if (__predict_false(toep->flags & TPF_FIN_SENT))
985 		panic("%s: excess tx.", __func__);
986 
987 	/*
988 	 * We have a PDU to send.  All of it goes out in one WR so 'm'
989 	 * is NULL.  A PDU's length is always a multiple of 4.
990 	 */
991 	MPASS(m == NULL);
992 	MPASS((plen & 3) == 0);
993 	MPASS(sndptr->m_pkthdr.len == plen);
994 
995 	shove = !(tp->t_flags & TF_MORETOCOME);
996 	ulp_submode = mbuf_ulp_submode(sndptr);
997 	MPASS(ulp_submode < nitems(ulp_extra_len));
998 
999 	/*
1000 	 * plen doesn't include header and data digests, which are
1001 	 * generated and inserted in the right places by the TOE, but
1002 	 * they do occupy TCP sequence space and need to be accounted
1003 	 * for.
1004 	 */
1005 	adjusted_plen = plen + ulp_extra_len[ulp_submode];
1006 	if (plen <= max_imm) {
1007 
1008 		/* Immediate data tx */
1009 
1010 		wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
1011 				&toep->ofld_txq->wrq);
1012 		if (wr == NULL) {
1013 			/* XXX: how will we recover from this? */
1014 			return (NULL);
1015 		}
1016 		txwr = wrtod(wr);
1017 		credits = howmany(wr->wr_len, 16);
1018 		write_tx_wr(txwr, toep, plen, adjusted_plen, credits,
1019 		    shove, ulp_submode);
1020 		m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
1021 		nsegs = 0;
1022 	} else {
1023 		int wr_len;
1024 
1025 		/* DSGL tx */
1026 		wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
1027 		    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1028 		wr = alloc_wrqe(roundup2(wr_len, 16),
1029 		    &toep->ofld_txq->wrq);
1030 		if (wr == NULL) {
1031 			/* XXX: how will we recover from this? */
1032 			return (NULL);
1033 		}
1034 		txwr = wrtod(wr);
1035 		credits = howmany(wr_len, 16);
1036 		write_tx_wr(txwr, toep, 0, adjusted_plen, credits,
1037 		    shove, ulp_submode);
1038 		write_tx_sgl(txwr + 1, sndptr, m, nsegs, max_nsegs_1mbuf);
1039 		if (wr_len & 0xf) {
1040 			uint64_t *pad = (uint64_t *)((uintptr_t)txwr + wr_len);
1041 			*pad = 0;
1042 		}
1043 	}
1044 
1045 	tp->snd_nxt += adjusted_plen;
1046 	tp->snd_max += adjusted_plen;
1047 
1048 	counter_u64_add(toep->ofld_txq->tx_iscsi_pdus, 1);
1049 	counter_u64_add(toep->ofld_txq->tx_iscsi_octets, plen);
1050 
1051 	return (wr);
1052 }
1053 
1054 void
1055 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
1056 {
1057 	struct mbuf *sndptr, *m;
1058 	struct fw_wr_hdr *wrhdr;
1059 	struct wrqe *wr;
1060 	u_int plen, credits;
1061 	struct inpcb *inp = toep->inp;
1062 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
1063 	struct mbufq *pduq = &toep->ulp_pduq;
1064 
1065 	INP_WLOCK_ASSERT(inp);
1066 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1067 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1068 	KASSERT(ulp_mode(toep) == ULP_MODE_ISCSI,
1069 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1070 
1071 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1072 		return;
1073 
1074 	/*
1075 	 * This function doesn't resume by itself.  Someone else must clear the
1076 	 * flag and call this function.
1077 	 */
1078 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1079 		KASSERT(drop == 0,
1080 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1081 		return;
1082 	}
1083 
1084 	if (drop) {
1085 		struct socket *so = inp->inp_socket;
1086 		struct sockbuf *sb = &so->so_snd;
1087 		int sbu;
1088 
1089 		/*
1090 		 * An unlocked read is ok here as the data should only
1091 		 * transition from a non-zero value to either another
1092 		 * non-zero value or zero.  Once it is zero it should
1093 		 * stay zero.
1094 		 */
1095 		if (__predict_false(sbused(sb)) > 0) {
1096 			SOCKBUF_LOCK(sb);
1097 			sbu = sbused(sb);
1098 			if (sbu > 0) {
1099 				/*
1100 				 * The data transmitted before the
1101 				 * tid's ULP mode changed to ISCSI is
1102 				 * still in so_snd.  Incoming credits
1103 				 * should account for so_snd first.
1104 				 */
1105 				sbdrop_locked(sb, min(sbu, drop));
1106 				drop -= min(sbu, drop);
1107 			}
1108 			sowwakeup_locked(so);	/* unlocks so_snd */
1109 		}
1110 		rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
1111 	}
1112 
1113 	while ((sndptr = mbufq_first(pduq)) != NULL) {
1114 		wr = write_iscsi_mbuf_wr(toep, sndptr);
1115 		if (wr == NULL) {
1116 			toep->flags |= TPF_TX_SUSPENDED;
1117 			return;
1118 		}
1119 
1120 		plen = sndptr->m_pkthdr.len;
1121 		credits = howmany(wr->wr_len, 16);
1122 		KASSERT(toep->tx_credits >= credits,
1123 			("%s: not enough credits", __func__));
1124 
1125 		m = mbufq_dequeue(pduq);
1126 		MPASS(m == sndptr);
1127 		mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
1128 
1129 		toep->tx_credits -= credits;
1130 		toep->tx_nocompl += credits;
1131 		toep->plen_nocompl += plen;
1132 
1133 		/*
1134 		 * Ensure there are enough credits for a full-sized WR
1135 		 * as page pod WRs can be full-sized.
1136 		 */
1137 		if (toep->tx_credits <= SGE_MAX_WR_LEN * 5 / 4 &&
1138 		    toep->tx_nocompl >= toep->tx_total / 4) {
1139 			wrhdr = wrtod(wr);
1140 			wrhdr->hi |= htobe32(F_FW_WR_COMPL);
1141 			toep->tx_nocompl = 0;
1142 			toep->plen_nocompl = 0;
1143 		}
1144 
1145 		toep->flags |= TPF_TX_DATA_SENT;
1146 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
1147 			toep->flags |= TPF_TX_SUSPENDED;
1148 
1149 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1150 		txsd->plen = plen;
1151 		txsd->tx_credits = credits;
1152 		txsd++;
1153 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1154 			toep->txsd_pidx = 0;
1155 			txsd = &toep->txsd[0];
1156 		}
1157 		toep->txsd_avail--;
1158 
1159 		t4_l2t_send(sc, wr, toep->l2te);
1160 	}
1161 
1162 	/* Send a FIN if requested, but only if there are no more PDUs to send */
1163 	if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
1164 		t4_close_conn(sc, toep);
1165 }
1166 
1167 static inline void
1168 t4_push_data(struct adapter *sc, struct toepcb *toep, int drop)
1169 {
1170 
1171 	if (ulp_mode(toep) == ULP_MODE_ISCSI)
1172 		t4_push_pdus(sc, toep, drop);
1173 	else if (toep->flags & TPF_KTLS)
1174 		t4_push_ktls(sc, toep, drop);
1175 	else
1176 		t4_push_frames(sc, toep, drop);
1177 }
1178 
1179 int
1180 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
1181 {
1182 	struct adapter *sc = tod->tod_softc;
1183 #ifdef INVARIANTS
1184 	struct inpcb *inp = tp->t_inpcb;
1185 #endif
1186 	struct toepcb *toep = tp->t_toe;
1187 
1188 	INP_WLOCK_ASSERT(inp);
1189 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1190 	    ("%s: inp %p dropped.", __func__, inp));
1191 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1192 
1193 	t4_push_data(sc, toep, 0);
1194 
1195 	return (0);
1196 }
1197 
1198 int
1199 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
1200 {
1201 	struct adapter *sc = tod->tod_softc;
1202 #ifdef INVARIANTS
1203 	struct inpcb *inp = tp->t_inpcb;
1204 #endif
1205 	struct toepcb *toep = tp->t_toe;
1206 
1207 	INP_WLOCK_ASSERT(inp);
1208 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1209 	    ("%s: inp %p dropped.", __func__, inp));
1210 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1211 
1212 	toep->flags |= TPF_SEND_FIN;
1213 	if (tp->t_state >= TCPS_ESTABLISHED)
1214 		t4_push_data(sc, toep, 0);
1215 
1216 	return (0);
1217 }
1218 
1219 int
1220 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1221 {
1222 	struct adapter *sc = tod->tod_softc;
1223 #if defined(INVARIANTS)
1224 	struct inpcb *inp = tp->t_inpcb;
1225 #endif
1226 	struct toepcb *toep = tp->t_toe;
1227 
1228 	INP_WLOCK_ASSERT(inp);
1229 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1230 	    ("%s: inp %p dropped.", __func__, inp));
1231 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1232 
1233 	/* hmmmm */
1234 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1235 	    ("%s: flowc for tid %u [%s] not sent already",
1236 	    __func__, toep->tid, tcpstates[tp->t_state]));
1237 
1238 	send_reset(sc, toep, 0);
1239 	return (0);
1240 }
1241 
1242 /*
1243  * Peer has sent us a FIN.
1244  */
1245 static int
1246 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1247 {
1248 	struct adapter *sc = iq->adapter;
1249 	const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1250 	unsigned int tid = GET_TID(cpl);
1251 	struct toepcb *toep = lookup_tid(sc, tid);
1252 	struct inpcb *inp = toep->inp;
1253 	struct tcpcb *tp = NULL;
1254 	struct socket *so;
1255 	struct epoch_tracker et;
1256 #ifdef INVARIANTS
1257 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1258 #endif
1259 
1260 	KASSERT(opcode == CPL_PEER_CLOSE,
1261 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1262 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1263 
1264 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1265 		/*
1266 		 * do_pass_establish must have run before do_peer_close and if
1267 		 * this is still a synqe instead of a toepcb then the connection
1268 		 * must be getting aborted.
1269 		 */
1270 		MPASS(toep->flags & TPF_ABORT_SHUTDOWN);
1271 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1272 		    toep, toep->flags);
1273 		return (0);
1274 	}
1275 
1276 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1277 
1278 	CURVNET_SET(toep->vnet);
1279 	NET_EPOCH_ENTER(et);
1280 	INP_WLOCK(inp);
1281 	tp = intotcpcb(inp);
1282 
1283 	CTR6(KTR_CXGBE,
1284 	    "%s: tid %u (%s), toep_flags 0x%x, ddp_flags 0x%x, inp %p",
1285 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1286 	    toep->ddp.flags, inp);
1287 
1288 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1289 		goto done;
1290 
1291 	if (ulp_mode(toep) == ULP_MODE_RDMA ||
1292 	    (ulp_mode(toep) == ULP_MODE_ISCSI && chip_id(sc) >= CHELSIO_T6)) {
1293 		/*
1294 		 * There might be data received via DDP before the FIN
1295 		 * not reported to the driver.  Just assume the
1296 		 * sequence number in the CPL is correct as the
1297 		 * sequence number of the FIN.
1298 		 */
1299 	} else {
1300 		KASSERT(tp->rcv_nxt + 1 == be32toh(cpl->rcv_nxt),
1301 		    ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1302 		    be32toh(cpl->rcv_nxt)));
1303 	}
1304 
1305 	tp->rcv_nxt = be32toh(cpl->rcv_nxt);
1306 
1307 	so = inp->inp_socket;
1308 	socantrcvmore(so);
1309 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1310 		DDP_LOCK(toep);
1311 		if (__predict_false(toep->ddp.flags &
1312 		    (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)))
1313 			handle_ddp_close(toep, tp, cpl->rcv_nxt);
1314 		DDP_UNLOCK(toep);
1315 	}
1316 
1317 	switch (tp->t_state) {
1318 	case TCPS_SYN_RECEIVED:
1319 		tp->t_starttime = ticks;
1320 		/* FALLTHROUGH */
1321 
1322 	case TCPS_ESTABLISHED:
1323 		tcp_state_change(tp, TCPS_CLOSE_WAIT);
1324 		break;
1325 
1326 	case TCPS_FIN_WAIT_1:
1327 		tcp_state_change(tp, TCPS_CLOSING);
1328 		break;
1329 
1330 	case TCPS_FIN_WAIT_2:
1331 		restore_so_proto(so, inp->inp_vflag & INP_IPV6);
1332 		tcp_twstart(tp);
1333 		INP_UNLOCK_ASSERT(inp);	 /* safe, we have a ref on the inp */
1334 		NET_EPOCH_EXIT(et);
1335 		CURVNET_RESTORE();
1336 
1337 		INP_WLOCK(inp);
1338 		final_cpl_received(toep);
1339 		return (0);
1340 
1341 	default:
1342 		log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1343 		    __func__, tid, tp->t_state);
1344 	}
1345 done:
1346 	INP_WUNLOCK(inp);
1347 	NET_EPOCH_EXIT(et);
1348 	CURVNET_RESTORE();
1349 	return (0);
1350 }
1351 
1352 /*
1353  * Peer has ACK'd our FIN.
1354  */
1355 static int
1356 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1357     struct mbuf *m)
1358 {
1359 	struct adapter *sc = iq->adapter;
1360 	const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1361 	unsigned int tid = GET_TID(cpl);
1362 	struct toepcb *toep = lookup_tid(sc, tid);
1363 	struct inpcb *inp = toep->inp;
1364 	struct tcpcb *tp = NULL;
1365 	struct socket *so = NULL;
1366 	struct epoch_tracker et;
1367 #ifdef INVARIANTS
1368 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1369 #endif
1370 
1371 	KASSERT(opcode == CPL_CLOSE_CON_RPL,
1372 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1373 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1374 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1375 
1376 	CURVNET_SET(toep->vnet);
1377 	NET_EPOCH_ENTER(et);
1378 	INP_WLOCK(inp);
1379 	tp = intotcpcb(inp);
1380 
1381 	CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1382 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1383 
1384 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1385 		goto done;
1386 
1387 	so = inp->inp_socket;
1388 	tp->snd_una = be32toh(cpl->snd_nxt) - 1;	/* exclude FIN */
1389 
1390 	switch (tp->t_state) {
1391 	case TCPS_CLOSING:	/* see TCPS_FIN_WAIT_2 in do_peer_close too */
1392 		restore_so_proto(so, inp->inp_vflag & INP_IPV6);
1393 		tcp_twstart(tp);
1394 release:
1395 		INP_UNLOCK_ASSERT(inp);	/* safe, we have a ref on the  inp */
1396 		NET_EPOCH_EXIT(et);
1397 		CURVNET_RESTORE();
1398 
1399 		INP_WLOCK(inp);
1400 		final_cpl_received(toep);	/* no more CPLs expected */
1401 
1402 		return (0);
1403 	case TCPS_LAST_ACK:
1404 		if (tcp_close(tp))
1405 			INP_WUNLOCK(inp);
1406 		goto release;
1407 
1408 	case TCPS_FIN_WAIT_1:
1409 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1410 			soisdisconnected(so);
1411 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
1412 		break;
1413 
1414 	default:
1415 		log(LOG_ERR,
1416 		    "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1417 		    __func__, tid, tcpstates[tp->t_state]);
1418 	}
1419 done:
1420 	INP_WUNLOCK(inp);
1421 	NET_EPOCH_EXIT(et);
1422 	CURVNET_RESTORE();
1423 	return (0);
1424 }
1425 
1426 void
1427 send_abort_rpl(struct adapter *sc, struct sge_ofld_txq *ofld_txq, int tid,
1428     int rst_status)
1429 {
1430 	struct wrqe *wr;
1431 	struct cpl_abort_rpl *cpl;
1432 
1433 	wr = alloc_wrqe(sizeof(*cpl), &ofld_txq->wrq);
1434 	if (wr == NULL) {
1435 		/* XXX */
1436 		panic("%s: allocation failure.", __func__);
1437 	}
1438 	cpl = wrtod(wr);
1439 
1440 	INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1441 	cpl->cmd = rst_status;
1442 
1443 	t4_wrq_tx(sc, wr);
1444 }
1445 
1446 static int
1447 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1448 {
1449 	switch (abort_reason) {
1450 	case CPL_ERR_BAD_SYN:
1451 	case CPL_ERR_CONN_RESET:
1452 		return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1453 	case CPL_ERR_XMIT_TIMEDOUT:
1454 	case CPL_ERR_PERSIST_TIMEDOUT:
1455 	case CPL_ERR_FINWAIT2_TIMEDOUT:
1456 	case CPL_ERR_KEEPALIVE_TIMEDOUT:
1457 		return (ETIMEDOUT);
1458 	default:
1459 		return (EIO);
1460 	}
1461 }
1462 
1463 /*
1464  * TCP RST from the peer, timeout, or some other such critical error.
1465  */
1466 static int
1467 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1468 {
1469 	struct adapter *sc = iq->adapter;
1470 	const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1471 	unsigned int tid = GET_TID(cpl);
1472 	struct toepcb *toep = lookup_tid(sc, tid);
1473 	struct sge_ofld_txq *ofld_txq = toep->ofld_txq;
1474 	struct inpcb *inp;
1475 	struct tcpcb *tp;
1476 	struct epoch_tracker et;
1477 #ifdef INVARIANTS
1478 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1479 #endif
1480 
1481 	KASSERT(opcode == CPL_ABORT_REQ_RSS,
1482 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1483 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1484 
1485 	if (toep->flags & TPF_SYNQE)
1486 		return (do_abort_req_synqe(iq, rss, m));
1487 
1488 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1489 
1490 	if (negative_advice(cpl->status)) {
1491 		CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1492 		    __func__, cpl->status, tid, toep->flags);
1493 		return (0);	/* Ignore negative advice */
1494 	}
1495 
1496 	inp = toep->inp;
1497 	CURVNET_SET(toep->vnet);
1498 	NET_EPOCH_ENTER(et);	/* for tcp_close */
1499 	INP_WLOCK(inp);
1500 
1501 	tp = intotcpcb(inp);
1502 
1503 	CTR6(KTR_CXGBE,
1504 	    "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1505 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1506 	    inp->inp_flags, cpl->status);
1507 
1508 	/*
1509 	 * If we'd initiated an abort earlier the reply to it is responsible for
1510 	 * cleaning up resources.  Otherwise we tear everything down right here
1511 	 * right now.  We owe the T4 a CPL_ABORT_RPL no matter what.
1512 	 */
1513 	if (toep->flags & TPF_ABORT_SHUTDOWN) {
1514 		INP_WUNLOCK(inp);
1515 		goto done;
1516 	}
1517 	toep->flags |= TPF_ABORT_SHUTDOWN;
1518 
1519 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
1520 		struct socket *so = inp->inp_socket;
1521 
1522 		if (so != NULL)
1523 			so_error_set(so, abort_status_to_errno(tp,
1524 			    cpl->status));
1525 		tp = tcp_close(tp);
1526 		if (tp == NULL)
1527 			INP_WLOCK(inp);	/* re-acquire */
1528 	}
1529 
1530 	final_cpl_received(toep);
1531 done:
1532 	NET_EPOCH_EXIT(et);
1533 	CURVNET_RESTORE();
1534 	send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1535 	return (0);
1536 }
1537 
1538 /*
1539  * Reply to the CPL_ABORT_REQ (send_reset)
1540  */
1541 static int
1542 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1543 {
1544 	struct adapter *sc = iq->adapter;
1545 	const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1546 	unsigned int tid = GET_TID(cpl);
1547 	struct toepcb *toep = lookup_tid(sc, tid);
1548 	struct inpcb *inp = toep->inp;
1549 #ifdef INVARIANTS
1550 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1551 #endif
1552 
1553 	KASSERT(opcode == CPL_ABORT_RPL_RSS,
1554 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1555 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1556 
1557 	if (toep->flags & TPF_SYNQE)
1558 		return (do_abort_rpl_synqe(iq, rss, m));
1559 
1560 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1561 
1562 	CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1563 	    __func__, tid, toep, inp, cpl->status);
1564 
1565 	KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1566 	    ("%s: wasn't expecting abort reply", __func__));
1567 
1568 	INP_WLOCK(inp);
1569 	final_cpl_received(toep);
1570 
1571 	return (0);
1572 }
1573 
1574 static int
1575 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1576 {
1577 	struct adapter *sc = iq->adapter;
1578 	const struct cpl_rx_data *cpl = mtod(m, const void *);
1579 	unsigned int tid = GET_TID(cpl);
1580 	struct toepcb *toep = lookup_tid(sc, tid);
1581 	struct inpcb *inp = toep->inp;
1582 	struct tcpcb *tp;
1583 	struct socket *so;
1584 	struct sockbuf *sb;
1585 	struct epoch_tracker et;
1586 	int len, rx_credits;
1587 	uint32_t ddp_placed = 0;
1588 
1589 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1590 		/*
1591 		 * do_pass_establish must have run before do_rx_data and if this
1592 		 * is still a synqe instead of a toepcb then the connection must
1593 		 * be getting aborted.
1594 		 */
1595 		MPASS(toep->flags & TPF_ABORT_SHUTDOWN);
1596 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1597 		    toep, toep->flags);
1598 		m_freem(m);
1599 		return (0);
1600 	}
1601 
1602 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1603 
1604 	/* strip off CPL header */
1605 	m_adj(m, sizeof(*cpl));
1606 	len = m->m_pkthdr.len;
1607 
1608 	INP_WLOCK(inp);
1609 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1610 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1611 		    __func__, tid, len, inp->inp_flags);
1612 		INP_WUNLOCK(inp);
1613 		m_freem(m);
1614 		return (0);
1615 	}
1616 
1617 	tp = intotcpcb(inp);
1618 
1619 	if (__predict_false(ulp_mode(toep) == ULP_MODE_TLS &&
1620 	   toep->flags & TPF_TLS_RECEIVE)) {
1621 		/* Received "raw" data on a TLS socket. */
1622 		CTR3(KTR_CXGBE, "%s: tid %u, raw TLS data (%d bytes)",
1623 		    __func__, tid, len);
1624 		do_rx_data_tls(cpl, toep, m);
1625 		return (0);
1626 	}
1627 
1628 	if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1629 		ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1630 
1631 	tp->rcv_nxt += len;
1632 	if (tp->rcv_wnd < len) {
1633 		KASSERT(ulp_mode(toep) == ULP_MODE_RDMA,
1634 				("%s: negative window size", __func__));
1635 	}
1636 
1637 	tp->rcv_wnd -= len;
1638 	tp->t_rcvtime = ticks;
1639 
1640 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1641 		DDP_LOCK(toep);
1642 	so = inp_inpcbtosocket(inp);
1643 	sb = &so->so_rcv;
1644 	SOCKBUF_LOCK(sb);
1645 
1646 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1647 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1648 		    __func__, tid, len);
1649 		m_freem(m);
1650 		SOCKBUF_UNLOCK(sb);
1651 		if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1652 			DDP_UNLOCK(toep);
1653 		INP_WUNLOCK(inp);
1654 
1655 		CURVNET_SET(toep->vnet);
1656 		NET_EPOCH_ENTER(et);
1657 		INP_WLOCK(inp);
1658 		tp = tcp_drop(tp, ECONNRESET);
1659 		if (tp)
1660 			INP_WUNLOCK(inp);
1661 		NET_EPOCH_EXIT(et);
1662 		CURVNET_RESTORE();
1663 
1664 		return (0);
1665 	}
1666 
1667 	/* receive buffer autosize */
1668 	MPASS(toep->vnet == so->so_vnet);
1669 	CURVNET_SET(toep->vnet);
1670 	if (sb->sb_flags & SB_AUTOSIZE &&
1671 	    V_tcp_do_autorcvbuf &&
1672 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1673 	    len > (sbspace(sb) / 8 * 7)) {
1674 		unsigned int hiwat = sb->sb_hiwat;
1675 		unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
1676 		    V_tcp_autorcvbuf_max);
1677 
1678 		if (!sbreserve_locked(sb, newsize, so, NULL))
1679 			sb->sb_flags &= ~SB_AUTOSIZE;
1680 	}
1681 
1682 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1683 		int changed = !(toep->ddp.flags & DDP_ON) ^ cpl->ddp_off;
1684 
1685 		if (toep->ddp.waiting_count != 0 || toep->ddp.active_count != 0)
1686 			CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)",
1687 			    __func__, tid, len);
1688 
1689 		if (changed) {
1690 			if (toep->ddp.flags & DDP_SC_REQ)
1691 				toep->ddp.flags ^= DDP_ON | DDP_SC_REQ;
1692 			else {
1693 				KASSERT(cpl->ddp_off == 1,
1694 				    ("%s: DDP switched on by itself.",
1695 				    __func__));
1696 
1697 				/* Fell out of DDP mode */
1698 				toep->ddp.flags &= ~DDP_ON;
1699 				CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
1700 				    __func__);
1701 
1702 				insert_ddp_data(toep, ddp_placed);
1703 			}
1704 		}
1705 
1706 		if (toep->ddp.flags & DDP_ON) {
1707 			/*
1708 			 * CPL_RX_DATA with DDP on can only be an indicate.
1709 			 * Start posting queued AIO requests via DDP.  The
1710 			 * payload that arrived in this indicate is appended
1711 			 * to the socket buffer as usual.
1712 			 */
1713 			handle_ddp_indicate(toep);
1714 		}
1715 	}
1716 
1717 	sbappendstream_locked(sb, m, 0);
1718 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
1719 	if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) {
1720 		rx_credits = send_rx_credits(sc, toep, rx_credits);
1721 		tp->rcv_wnd += rx_credits;
1722 		tp->rcv_adv += rx_credits;
1723 	}
1724 
1725 	if (ulp_mode(toep) == ULP_MODE_TCPDDP && toep->ddp.waiting_count > 0 &&
1726 	    sbavail(sb) != 0) {
1727 		CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__,
1728 		    tid);
1729 		ddp_queue_toep(toep);
1730 	}
1731 	sorwakeup_locked(so);
1732 	SOCKBUF_UNLOCK_ASSERT(sb);
1733 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1734 		DDP_UNLOCK(toep);
1735 
1736 	INP_WUNLOCK(inp);
1737 	CURVNET_RESTORE();
1738 	return (0);
1739 }
1740 
1741 static int
1742 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1743 {
1744 	struct adapter *sc = iq->adapter;
1745 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1746 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1747 	struct toepcb *toep = lookup_tid(sc, tid);
1748 	struct inpcb *inp;
1749 	struct tcpcb *tp;
1750 	struct socket *so;
1751 	uint8_t credits = cpl->credits;
1752 	struct ofld_tx_sdesc *txsd;
1753 	int plen;
1754 #ifdef INVARIANTS
1755 	unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1756 #endif
1757 
1758 	/*
1759 	 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1760 	 * now this comes back carrying the credits for the flowc.
1761 	 */
1762 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1763 		KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1764 		    ("%s: credits for a synq entry %p", __func__, toep));
1765 		return (0);
1766 	}
1767 
1768 	inp = toep->inp;
1769 
1770 	KASSERT(opcode == CPL_FW4_ACK,
1771 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1772 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1773 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1774 
1775 	INP_WLOCK(inp);
1776 
1777 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1778 		INP_WUNLOCK(inp);
1779 		return (0);
1780 	}
1781 
1782 	KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0,
1783 	    ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1784 
1785 	tp = intotcpcb(inp);
1786 
1787 	if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1788 		tcp_seq snd_una = be32toh(cpl->snd_una);
1789 
1790 #ifdef INVARIANTS
1791 		if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1792 			log(LOG_ERR,
1793 			    "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1794 			    __func__, snd_una, toep->tid, tp->snd_una);
1795 		}
1796 #endif
1797 
1798 		if (tp->snd_una != snd_una) {
1799 			tp->snd_una = snd_una;
1800 			tp->ts_recent_age = tcp_ts_getticks();
1801 		}
1802 	}
1803 
1804 #ifdef VERBOSE_TRACES
1805 	CTR3(KTR_CXGBE, "%s: tid %d credits %u", __func__, tid, credits);
1806 #endif
1807 	so = inp->inp_socket;
1808 	txsd = &toep->txsd[toep->txsd_cidx];
1809 	plen = 0;
1810 	while (credits) {
1811 		KASSERT(credits >= txsd->tx_credits,
1812 		    ("%s: too many (or partial) credits", __func__));
1813 		credits -= txsd->tx_credits;
1814 		toep->tx_credits += txsd->tx_credits;
1815 		plen += txsd->plen;
1816 		txsd++;
1817 		toep->txsd_avail++;
1818 		KASSERT(toep->txsd_avail <= toep->txsd_total,
1819 		    ("%s: txsd avail > total", __func__));
1820 		if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1821 			txsd = &toep->txsd[0];
1822 			toep->txsd_cidx = 0;
1823 		}
1824 	}
1825 
1826 	if (toep->tx_credits == toep->tx_total) {
1827 		toep->tx_nocompl = 0;
1828 		toep->plen_nocompl = 0;
1829 	}
1830 
1831 	if (toep->flags & TPF_TX_SUSPENDED &&
1832 	    toep->tx_credits >= toep->tx_total / 4) {
1833 #ifdef VERBOSE_TRACES
1834 		CTR2(KTR_CXGBE, "%s: tid %d calling t4_push_frames", __func__,
1835 		    tid);
1836 #endif
1837 		toep->flags &= ~TPF_TX_SUSPENDED;
1838 		CURVNET_SET(toep->vnet);
1839 		t4_push_data(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 (ulp_mode(toep) == ULP_MODE_ISCSI) {
1848 			if (__predict_false(sbu > 0)) {
1849 				/*
1850 				 * The data transmitted before the
1851 				 * tid's ULP mode changed to ISCSI is
1852 				 * still in so_snd.  Incoming credits
1853 				 * should account for so_snd first.
1854 				 */
1855 				sbdrop_locked(sb, min(sbu, plen));
1856 				plen -= min(sbu, plen);
1857 			}
1858 			sowwakeup_locked(so);	/* unlocks so_snd */
1859 			rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1860 		} else {
1861 #ifdef VERBOSE_TRACES
1862 			CTR3(KTR_CXGBE, "%s: tid %d dropped %d bytes", __func__,
1863 			    tid, plen);
1864 #endif
1865 			sbdrop_locked(sb, plen);
1866 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
1867 				t4_aiotx_queue_toep(so, toep);
1868 			sowwakeup_locked(so);	/* unlocks so_snd */
1869 		}
1870 		SOCKBUF_UNLOCK_ASSERT(sb);
1871 	}
1872 
1873 	INP_WUNLOCK(inp);
1874 
1875 	return (0);
1876 }
1877 
1878 void
1879 t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, struct toepcb *toep,
1880     uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie)
1881 {
1882 	struct wrqe *wr;
1883 	struct cpl_set_tcb_field *req;
1884 	struct ofld_tx_sdesc *txsd;
1885 
1886 	MPASS((cookie & ~M_COOKIE) == 0);
1887 	if (reply) {
1888 		MPASS(cookie != CPL_COOKIE_RESERVED);
1889 	}
1890 
1891 	wr = alloc_wrqe(sizeof(*req), wrq);
1892 	if (wr == NULL) {
1893 		/* XXX */
1894 		panic("%s: allocation failure.", __func__);
1895 	}
1896 	req = wrtod(wr);
1897 
1898 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1899 	req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1900 	if (reply == 0)
1901 		req->reply_ctrl |= htobe16(F_NO_REPLY);
1902 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie));
1903 	req->mask = htobe64(mask);
1904 	req->val = htobe64(val);
1905 	if (wrq->eq.type == EQ_OFLD) {
1906 		txsd = &toep->txsd[toep->txsd_pidx];
1907 		txsd->tx_credits = howmany(sizeof(*req), 16);
1908 		txsd->plen = 0;
1909 		KASSERT(toep->tx_credits >= txsd->tx_credits &&
1910 		    toep->txsd_avail > 0,
1911 		    ("%s: not enough credits (%d)", __func__,
1912 		    toep->tx_credits));
1913 		toep->tx_credits -= txsd->tx_credits;
1914 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1915 			toep->txsd_pidx = 0;
1916 		toep->txsd_avail--;
1917 	}
1918 
1919 	t4_wrq_tx(sc, wr);
1920 }
1921 
1922 void
1923 t4_init_cpl_io_handlers(void)
1924 {
1925 
1926 	t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close);
1927 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl);
1928 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req);
1929 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl,
1930 	    CPL_COOKIE_TOM);
1931 	t4_register_cpl_handler(CPL_RX_DATA, do_rx_data);
1932 	t4_register_shared_cpl_handler(CPL_FW4_ACK, do_fw4_ack, CPL_COOKIE_TOM);
1933 }
1934 
1935 void
1936 t4_uninit_cpl_io_handlers(void)
1937 {
1938 
1939 	t4_register_cpl_handler(CPL_PEER_CLOSE, NULL);
1940 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, NULL);
1941 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, NULL);
1942 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, NULL, CPL_COOKIE_TOM);
1943 	t4_register_cpl_handler(CPL_RX_DATA, NULL);
1944 	t4_register_shared_cpl_handler(CPL_FW4_ACK, NULL, CPL_COOKIE_TOM);
1945 }
1946 
1947 /*
1948  * Use the 'backend1' field in AIO jobs to hold an error that should
1949  * be reported when the job is completed, the 'backend3' field to
1950  * store the amount of data sent by the AIO job so far, and the
1951  * 'backend4' field to hold a reference count on the job.
1952  *
1953  * Each unmapped mbuf holds a reference on the job as does the queue
1954  * so long as the job is queued.
1955  */
1956 #define	aio_error	backend1
1957 #define	aio_sent	backend3
1958 #define	aio_refs	backend4
1959 
1960 #define	jobtotid(job)							\
1961 	(((struct toepcb *)(so_sototcpcb((job)->fd_file->f_data)->t_toe))->tid)
1962 
1963 static void
1964 aiotx_free_job(struct kaiocb *job)
1965 {
1966 	long status;
1967 	int error;
1968 
1969 	if (refcount_release(&job->aio_refs) == 0)
1970 		return;
1971 
1972 	error = (intptr_t)job->aio_error;
1973 	status = job->aio_sent;
1974 #ifdef VERBOSE_TRACES
1975 	CTR5(KTR_CXGBE, "%s: tid %d completed %p len %ld, error %d", __func__,
1976 	    jobtotid(job), job, status, error);
1977 #endif
1978 	if (error != 0 && status != 0)
1979 		error = 0;
1980 	if (error == ECANCELED)
1981 		aio_cancel(job);
1982 	else if (error)
1983 		aio_complete(job, -1, error);
1984 	else {
1985 		job->msgsnd = 1;
1986 		aio_complete(job, status, 0);
1987 	}
1988 }
1989 
1990 static void
1991 aiotx_free_pgs(struct mbuf *m)
1992 {
1993 	struct kaiocb *job;
1994 	vm_page_t pg;
1995 
1996 	M_ASSERTEXTPG(m);
1997 	job = m->m_ext.ext_arg1;
1998 #ifdef VERBOSE_TRACES
1999 	CTR3(KTR_CXGBE, "%s: completed %d bytes for tid %d", __func__,
2000 	    m->m_len, jobtotid(job));
2001 #endif
2002 
2003 	for (int i = 0; i < m->m_epg_npgs; i++) {
2004 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
2005 		vm_page_unwire(pg, PQ_ACTIVE);
2006 	}
2007 
2008 	aiotx_free_job(job);
2009 }
2010 
2011 /*
2012  * Allocate a chain of unmapped mbufs describing the next 'len' bytes
2013  * of an AIO job.
2014  */
2015 static struct mbuf *
2016 alloc_aiotx_mbuf(struct kaiocb *job, int len)
2017 {
2018 	struct vmspace *vm;
2019 	vm_page_t pgs[MBUF_PEXT_MAX_PGS];
2020 	struct mbuf *m, *top, *last;
2021 	vm_map_t map;
2022 	vm_offset_t start;
2023 	int i, mlen, npages, pgoff;
2024 
2025 	KASSERT(job->aio_sent + len <= job->uaiocb.aio_nbytes,
2026 	    ("%s(%p, %d): request to send beyond end of buffer", __func__,
2027 	    job, len));
2028 
2029 	/*
2030 	 * The AIO subsystem will cancel and drain all requests before
2031 	 * permitting a process to exit or exec, so p_vmspace should
2032 	 * be stable here.
2033 	 */
2034 	vm = job->userproc->p_vmspace;
2035 	map = &vm->vm_map;
2036 	start = (uintptr_t)job->uaiocb.aio_buf + job->aio_sent;
2037 	pgoff = start & PAGE_MASK;
2038 
2039 	top = NULL;
2040 	last = NULL;
2041 	while (len > 0) {
2042 		mlen = imin(len, MBUF_PEXT_MAX_PGS * PAGE_SIZE - pgoff);
2043 		KASSERT(mlen == len || ((start + mlen) & PAGE_MASK) == 0,
2044 		    ("%s: next start (%#jx + %#x) is not page aligned",
2045 		    __func__, (uintmax_t)start, mlen));
2046 
2047 		npages = vm_fault_quick_hold_pages(map, start, mlen,
2048 		    VM_PROT_WRITE, pgs, nitems(pgs));
2049 		if (npages < 0)
2050 			break;
2051 
2052 		m = mb_alloc_ext_pgs(M_WAITOK, aiotx_free_pgs);
2053 		if (m == NULL) {
2054 			vm_page_unhold_pages(pgs, npages);
2055 			break;
2056 		}
2057 
2058 		m->m_epg_1st_off = pgoff;
2059 		m->m_epg_npgs = npages;
2060 		if (npages == 1) {
2061 			KASSERT(mlen + pgoff <= PAGE_SIZE,
2062 			    ("%s: single page is too large (off %d len %d)",
2063 			    __func__, pgoff, mlen));
2064 			m->m_epg_last_len = mlen;
2065 		} else {
2066 			m->m_epg_last_len = mlen - (PAGE_SIZE - pgoff) -
2067 			    (npages - 2) * PAGE_SIZE;
2068 		}
2069 		for (i = 0; i < npages; i++)
2070 			m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pgs[i]);
2071 
2072 		m->m_len = mlen;
2073 		m->m_ext.ext_size = npages * PAGE_SIZE;
2074 		m->m_ext.ext_arg1 = job;
2075 		refcount_acquire(&job->aio_refs);
2076 
2077 #ifdef VERBOSE_TRACES
2078 		CTR5(KTR_CXGBE, "%s: tid %d, new mbuf %p for job %p, npages %d",
2079 		    __func__, jobtotid(job), m, job, npages);
2080 #endif
2081 
2082 		if (top == NULL)
2083 			top = m;
2084 		else
2085 			last->m_next = m;
2086 		last = m;
2087 
2088 		len -= mlen;
2089 		start += mlen;
2090 		pgoff = 0;
2091 	}
2092 
2093 	return (top);
2094 }
2095 
2096 static void
2097 t4_aiotx_process_job(struct toepcb *toep, struct socket *so, struct kaiocb *job)
2098 {
2099 	struct sockbuf *sb;
2100 	struct file *fp;
2101 	struct inpcb *inp;
2102 	struct tcpcb *tp;
2103 	struct mbuf *m;
2104 	int error, len;
2105 	bool moretocome, sendmore;
2106 
2107 	sb = &so->so_snd;
2108 	SOCKBUF_UNLOCK(sb);
2109 	fp = job->fd_file;
2110 	m = NULL;
2111 
2112 #ifdef MAC
2113 	error = mac_socket_check_send(fp->f_cred, so);
2114 	if (error != 0)
2115 		goto out;
2116 #endif
2117 
2118 	/* Inline sosend_generic(). */
2119 
2120 	error = sblock(sb, SBL_WAIT);
2121 	MPASS(error == 0);
2122 
2123 sendanother:
2124 	SOCKBUF_LOCK(sb);
2125 	if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
2126 		SOCKBUF_UNLOCK(sb);
2127 		sbunlock(sb);
2128 		if ((so->so_options & SO_NOSIGPIPE) == 0) {
2129 			PROC_LOCK(job->userproc);
2130 			kern_psignal(job->userproc, SIGPIPE);
2131 			PROC_UNLOCK(job->userproc);
2132 		}
2133 		error = EPIPE;
2134 		goto out;
2135 	}
2136 	if (so->so_error) {
2137 		error = so->so_error;
2138 		so->so_error = 0;
2139 		SOCKBUF_UNLOCK(sb);
2140 		sbunlock(sb);
2141 		goto out;
2142 	}
2143 	if ((so->so_state & SS_ISCONNECTED) == 0) {
2144 		SOCKBUF_UNLOCK(sb);
2145 		sbunlock(sb);
2146 		error = ENOTCONN;
2147 		goto out;
2148 	}
2149 	if (sbspace(sb) < sb->sb_lowat) {
2150 		MPASS(job->aio_sent == 0 || !(so->so_state & SS_NBIO));
2151 
2152 		/*
2153 		 * Don't block if there is too little room in the socket
2154 		 * buffer.  Instead, requeue the request.
2155 		 */
2156 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2157 			SOCKBUF_UNLOCK(sb);
2158 			sbunlock(sb);
2159 			error = ECANCELED;
2160 			goto out;
2161 		}
2162 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2163 		SOCKBUF_UNLOCK(sb);
2164 		sbunlock(sb);
2165 		goto out;
2166 	}
2167 
2168 	/*
2169 	 * Write as much data as the socket permits, but no more than a
2170 	 * a single sndbuf at a time.
2171 	 */
2172 	len = sbspace(sb);
2173 	if (len > job->uaiocb.aio_nbytes - job->aio_sent) {
2174 		len = job->uaiocb.aio_nbytes - job->aio_sent;
2175 		moretocome = false;
2176 	} else
2177 		moretocome = true;
2178 	if (len > toep->params.sndbuf) {
2179 		len = toep->params.sndbuf;
2180 		sendmore = true;
2181 	} else
2182 		sendmore = false;
2183 
2184 	if (!TAILQ_EMPTY(&toep->aiotx_jobq))
2185 		moretocome = true;
2186 	SOCKBUF_UNLOCK(sb);
2187 	MPASS(len != 0);
2188 
2189 	m = alloc_aiotx_mbuf(job, len);
2190 	if (m == NULL) {
2191 		sbunlock(sb);
2192 		error = EFAULT;
2193 		goto out;
2194 	}
2195 
2196 	/* Inlined tcp_usr_send(). */
2197 
2198 	inp = toep->inp;
2199 	INP_WLOCK(inp);
2200 	if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
2201 		INP_WUNLOCK(inp);
2202 		sbunlock(sb);
2203 		error = ECONNRESET;
2204 		goto out;
2205 	}
2206 
2207 	job->aio_sent += m_length(m, NULL);
2208 
2209 	sbappendstream(sb, m, 0);
2210 	m = NULL;
2211 
2212 	if (!(inp->inp_flags & INP_DROPPED)) {
2213 		tp = intotcpcb(inp);
2214 		if (moretocome)
2215 			tp->t_flags |= TF_MORETOCOME;
2216 		error = tp->t_fb->tfb_tcp_output(tp);
2217 		if (moretocome)
2218 			tp->t_flags &= ~TF_MORETOCOME;
2219 	}
2220 
2221 	INP_WUNLOCK(inp);
2222 	if (sendmore)
2223 		goto sendanother;
2224 	sbunlock(sb);
2225 
2226 	if (error)
2227 		goto out;
2228 
2229 	/*
2230 	 * If this is a blocking socket and the request has not been
2231 	 * fully completed, requeue it until the socket is ready
2232 	 * again.
2233 	 */
2234 	if (job->aio_sent < job->uaiocb.aio_nbytes &&
2235 	    !(so->so_state & SS_NBIO)) {
2236 		SOCKBUF_LOCK(sb);
2237 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2238 			SOCKBUF_UNLOCK(sb);
2239 			error = ECANCELED;
2240 			goto out;
2241 		}
2242 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2243 		return;
2244 	}
2245 
2246 	/*
2247 	 * If the request will not be requeued, drop the queue's
2248 	 * reference to the job.  Any mbufs in flight should still
2249 	 * hold a reference, but this drops the reference that the
2250 	 * queue owns while it is waiting to queue mbufs to the
2251 	 * socket.
2252 	 */
2253 	aiotx_free_job(job);
2254 
2255 out:
2256 	if (error) {
2257 		job->aio_error = (void *)(intptr_t)error;
2258 		aiotx_free_job(job);
2259 	}
2260 	m_freem(m);
2261 	SOCKBUF_LOCK(sb);
2262 }
2263 
2264 static void
2265 t4_aiotx_task(void *context, int pending)
2266 {
2267 	struct toepcb *toep = context;
2268 	struct socket *so;
2269 	struct kaiocb *job;
2270 
2271 	so = toep->aiotx_so;
2272 	CURVNET_SET(toep->vnet);
2273 	SOCKBUF_LOCK(&so->so_snd);
2274 	while (!TAILQ_EMPTY(&toep->aiotx_jobq) && sowriteable(so)) {
2275 		job = TAILQ_FIRST(&toep->aiotx_jobq);
2276 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2277 		if (!aio_clear_cancel_function(job))
2278 			continue;
2279 
2280 		t4_aiotx_process_job(toep, so, job);
2281 	}
2282 	toep->aiotx_so = NULL;
2283 	SOCKBUF_UNLOCK(&so->so_snd);
2284 	CURVNET_RESTORE();
2285 
2286 	free_toepcb(toep);
2287 	SOCK_LOCK(so);
2288 	sorele(so);
2289 }
2290 
2291 static void
2292 t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep)
2293 {
2294 
2295 	SOCKBUF_LOCK_ASSERT(&toep->inp->inp_socket->so_snd);
2296 #ifdef VERBOSE_TRACES
2297 	CTR3(KTR_CXGBE, "%s: queueing aiotx task for tid %d, active = %s",
2298 	    __func__, toep->tid, toep->aiotx_so != NULL ? "true" : "false");
2299 #endif
2300 	if (toep->aiotx_so != NULL)
2301 		return;
2302 	soref(so);
2303 	toep->aiotx_so = so;
2304 	hold_toepcb(toep);
2305 	soaio_enqueue(&toep->aiotx_task);
2306 }
2307 
2308 static void
2309 t4_aiotx_cancel(struct kaiocb *job)
2310 {
2311 	struct socket *so;
2312 	struct sockbuf *sb;
2313 	struct tcpcb *tp;
2314 	struct toepcb *toep;
2315 
2316 	so = job->fd_file->f_data;
2317 	tp = so_sototcpcb(so);
2318 	toep = tp->t_toe;
2319 	MPASS(job->uaiocb.aio_lio_opcode == LIO_WRITE);
2320 	sb = &so->so_snd;
2321 
2322 	SOCKBUF_LOCK(sb);
2323 	if (!aio_cancel_cleared(job))
2324 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2325 	SOCKBUF_UNLOCK(sb);
2326 
2327 	job->aio_error = (void *)(intptr_t)ECANCELED;
2328 	aiotx_free_job(job);
2329 }
2330 
2331 int
2332 t4_aio_queue_aiotx(struct socket *so, struct kaiocb *job)
2333 {
2334 	struct tcpcb *tp = so_sototcpcb(so);
2335 	struct toepcb *toep = tp->t_toe;
2336 	struct adapter *sc = td_adapter(toep->td);
2337 
2338 	/* This only handles writes. */
2339 	if (job->uaiocb.aio_lio_opcode != LIO_WRITE)
2340 		return (EOPNOTSUPP);
2341 
2342 	if (!sc->tt.tx_zcopy)
2343 		return (EOPNOTSUPP);
2344 
2345 	if (tls_tx_key(toep))
2346 		return (EOPNOTSUPP);
2347 
2348 	SOCKBUF_LOCK(&so->so_snd);
2349 #ifdef VERBOSE_TRACES
2350 	CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid);
2351 #endif
2352 	if (!aio_set_cancel_function(job, t4_aiotx_cancel))
2353 		panic("new job was cancelled");
2354 	refcount_init(&job->aio_refs, 1);
2355 	TAILQ_INSERT_TAIL(&toep->aiotx_jobq, job, list);
2356 	if (sowriteable(so))
2357 		t4_aiotx_queue_toep(so, toep);
2358 	SOCKBUF_UNLOCK(&so->so_snd);
2359 	return (0);
2360 }
2361 
2362 void
2363 aiotx_init_toep(struct toepcb *toep)
2364 {
2365 
2366 	TAILQ_INIT(&toep->aiotx_jobq);
2367 	TASK_INIT(&toep->aiotx_task, 0, t4_aiotx_task, toep);
2368 }
2369 #endif
2370