xref: /freebsd/sys/dev/cxgbe/tom/t4_cpl_io.c (revision af23369a6deaaeb612ab266eb88b8bb8d560c322)
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 <dev/iscsi/iscsi_proto.h>
71 
72 #include "common/common.h"
73 #include "common/t4_msg.h"
74 #include "common/t4_regs.h"
75 #include "common/t4_tcb.h"
76 #include "tom/t4_tom_l2t.h"
77 #include "tom/t4_tom.h"
78 
79 static void	t4_aiotx_cancel(struct kaiocb *job);
80 static void	t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep);
81 
82 void
83 send_flowc_wr(struct toepcb *toep, struct tcpcb *tp)
84 {
85 	struct wrqe *wr;
86 	struct fw_flowc_wr *flowc;
87 	unsigned int nparams, flowclen, paramidx;
88 	struct vi_info *vi = toep->vi;
89 	struct port_info *pi = vi->pi;
90 	struct adapter *sc = pi->adapter;
91 	unsigned int pfvf = sc->pf << S_FW_VIID_PFN;
92 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
93 
94 	KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT),
95 	    ("%s: flowc for tid %u sent already", __func__, toep->tid));
96 
97 	if (tp != NULL)
98 		nparams = 8;
99 	else
100 		nparams = 6;
101 	if (toep->params.tc_idx != -1) {
102 		MPASS(toep->params.tc_idx >= 0 &&
103 		    toep->params.tc_idx < sc->params.nsched_cls);
104 		nparams++;
105 	}
106 
107 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
108 
109 	wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq);
110 	if (wr == NULL) {
111 		/* XXX */
112 		panic("%s: allocation failure.", __func__);
113 	}
114 	flowc = wrtod(wr);
115 	memset(flowc, 0, wr->wr_len);
116 
117 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
118 	    V_FW_FLOWC_WR_NPARAMS(nparams));
119 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
120 	    V_FW_WR_FLOWID(toep->tid));
121 
122 #define FLOWC_PARAM(__m, __v) \
123 	do { \
124 		flowc->mnemval[paramidx].mnemonic = FW_FLOWC_MNEM_##__m; \
125 		flowc->mnemval[paramidx].val = htobe32(__v); \
126 		paramidx++; \
127 	} while (0)
128 
129 	paramidx = 0;
130 
131 	FLOWC_PARAM(PFNVFN, pfvf);
132 	FLOWC_PARAM(CH, pi->tx_chan);
133 	FLOWC_PARAM(PORT, pi->tx_chan);
134 	FLOWC_PARAM(IQID, toep->ofld_rxq->iq.abs_id);
135 	FLOWC_PARAM(SNDBUF, toep->params.sndbuf);
136 	if (tp) {
137 		FLOWC_PARAM(MSS, toep->params.emss);
138 		FLOWC_PARAM(SNDNXT, tp->snd_nxt);
139 		FLOWC_PARAM(RCVNXT, tp->rcv_nxt);
140 	} else
141 		FLOWC_PARAM(MSS, 512);
142 	CTR6(KTR_CXGBE,
143 	    "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x",
144 	    __func__, toep->tid, toep->params.emss, toep->params.sndbuf,
145 	    tp ? tp->snd_nxt : 0, tp ? tp->rcv_nxt : 0);
146 
147 	if (toep->params.tc_idx != -1)
148 		FLOWC_PARAM(SCHEDCLASS, toep->params.tc_idx);
149 #undef FLOWC_PARAM
150 
151 	KASSERT(paramidx == nparams, ("nparams mismatch"));
152 
153 	txsd->tx_credits = howmany(flowclen, 16);
154 	txsd->plen = 0;
155 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
156 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
157 	toep->tx_credits -= txsd->tx_credits;
158 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
159 		toep->txsd_pidx = 0;
160 	toep->txsd_avail--;
161 
162 	toep->flags |= TPF_FLOWC_WR_SENT;
163         t4_wrq_tx(sc, wr);
164 }
165 
166 #ifdef RATELIMIT
167 /*
168  * Input is Bytes/second (so_max_pacing_rate), chip counts in Kilobits/second.
169  */
170 static int
171 update_tx_rate_limit(struct adapter *sc, struct toepcb *toep, u_int Bps)
172 {
173 	int tc_idx, rc;
174 	const u_int kbps = (u_int) (uint64_t)Bps * 8ULL / 1000;
175 	const int port_id = toep->vi->pi->port_id;
176 
177 	CTR3(KTR_CXGBE, "%s: tid %u, rate %uKbps", __func__, toep->tid, kbps);
178 
179 	if (kbps == 0) {
180 		/* unbind */
181 		tc_idx = -1;
182 	} else {
183 		rc = t4_reserve_cl_rl_kbps(sc, port_id, kbps, &tc_idx);
184 		if (rc != 0)
185 			return (rc);
186 		MPASS(tc_idx >= 0 && tc_idx < sc->params.nsched_cls);
187 	}
188 
189 	if (toep->params.tc_idx != tc_idx) {
190 		struct wrqe *wr;
191 		struct fw_flowc_wr *flowc;
192 		int nparams = 1, flowclen, flowclen16;
193 		struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
194 
195 		flowclen = sizeof(*flowc) + nparams * sizeof(struct
196 		    fw_flowc_mnemval);
197 		flowclen16 = howmany(flowclen, 16);
198 		if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0 ||
199 		    (wr = alloc_wrqe(roundup2(flowclen, 16),
200 		    &toep->ofld_txq->wrq)) == NULL) {
201 			if (tc_idx >= 0)
202 				t4_release_cl_rl(sc, port_id, tc_idx);
203 			return (ENOMEM);
204 		}
205 
206 		flowc = wrtod(wr);
207 		memset(flowc, 0, wr->wr_len);
208 
209 		flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
210 		    V_FW_FLOWC_WR_NPARAMS(nparams));
211 		flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
212 		    V_FW_WR_FLOWID(toep->tid));
213 
214 		flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
215 		if (tc_idx == -1)
216 			flowc->mnemval[0].val = htobe32(0xff);
217 		else
218 			flowc->mnemval[0].val = htobe32(tc_idx);
219 
220 		txsd->tx_credits = flowclen16;
221 		txsd->plen = 0;
222 		toep->tx_credits -= txsd->tx_credits;
223 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
224 			toep->txsd_pidx = 0;
225 		toep->txsd_avail--;
226 		t4_wrq_tx(sc, wr);
227 	}
228 
229 	if (toep->params.tc_idx >= 0)
230 		t4_release_cl_rl(sc, port_id, toep->params.tc_idx);
231 	toep->params.tc_idx = tc_idx;
232 
233 	return (0);
234 }
235 #endif
236 
237 void
238 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
239 {
240 	struct wrqe *wr;
241 	struct cpl_abort_req *req;
242 	int tid = toep->tid;
243 	struct inpcb *inp = toep->inp;
244 	struct tcpcb *tp = intotcpcb(inp);	/* don't use if INP_DROPPED */
245 
246 	INP_WLOCK_ASSERT(inp);
247 
248 	CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
249 	    __func__, toep->tid,
250 	    inp->inp_flags & INP_DROPPED ? "inp dropped" :
251 	    tcpstates[tp->t_state],
252 	    toep->flags, inp->inp_flags,
253 	    toep->flags & TPF_ABORT_SHUTDOWN ?
254 	    " (abort already in progress)" : "");
255 
256 	if (toep->flags & TPF_ABORT_SHUTDOWN)
257 		return;	/* abort already in progress */
258 
259 	toep->flags |= TPF_ABORT_SHUTDOWN;
260 
261 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
262 	    ("%s: flowc_wr not sent for tid %d.", __func__, tid));
263 
264 	wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq);
265 	if (wr == NULL) {
266 		/* XXX */
267 		panic("%s: allocation failure.", __func__);
268 	}
269 	req = wrtod(wr);
270 
271 	INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
272 	if (inp->inp_flags & INP_DROPPED)
273 		req->rsvd0 = htobe32(snd_nxt);
274 	else
275 		req->rsvd0 = htobe32(tp->snd_nxt);
276 	req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
277 	req->cmd = CPL_ABORT_SEND_RST;
278 
279 	/*
280 	 * XXX: What's the correct way to tell that the inp hasn't been detached
281 	 * from its socket?  Should I even be flushing the snd buffer here?
282 	 */
283 	if ((inp->inp_flags & INP_DROPPED) == 0) {
284 		struct socket *so = inp->inp_socket;
285 
286 		if (so != NULL)	/* because I'm not sure.  See comment above */
287 			sbflush(&so->so_snd);
288 	}
289 
290 	t4_l2t_send(sc, wr, toep->l2te);
291 }
292 
293 /*
294  * Called when a connection is established to translate the TCP options
295  * reported by HW to FreeBSD's native format.
296  */
297 static void
298 assign_rxopt(struct tcpcb *tp, uint16_t opt)
299 {
300 	struct toepcb *toep = tp->t_toe;
301 	struct inpcb *inp = tptoinpcb(tp);
302 	struct adapter *sc = td_adapter(toep->td);
303 
304 	INP_LOCK_ASSERT(inp);
305 
306 	toep->params.mtu_idx = G_TCPOPT_MSS(opt);
307 	tp->t_maxseg = sc->params.mtus[toep->params.mtu_idx];
308 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
309 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
310 	else
311 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
312 
313 	toep->params.emss = tp->t_maxseg;
314 	if (G_TCPOPT_TSTAMP(opt)) {
315 		toep->params.tstamp = 1;
316 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
317 		tp->t_flags |= TF_RCVD_TSTMP;	/* timestamps ok */
318 		tp->ts_recent = 0;		/* hmmm */
319 		tp->ts_recent_age = tcp_ts_getticks();
320 	} else
321 		toep->params.tstamp = 0;
322 
323 	if (G_TCPOPT_SACK(opt)) {
324 		toep->params.sack = 1;
325 		tp->t_flags |= TF_SACK_PERMIT;	/* should already be set */
326 	} else {
327 		toep->params.sack = 0;
328 		tp->t_flags &= ~TF_SACK_PERMIT;	/* sack disallowed by peer */
329 	}
330 
331 	if (G_TCPOPT_WSCALE_OK(opt))
332 		tp->t_flags |= TF_RCVD_SCALE;
333 
334 	/* Doing window scaling? */
335 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
336 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
337 		tp->rcv_scale = tp->request_r_scale;
338 		tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
339 	} else
340 		toep->params.wscale = 0;
341 
342 	CTR6(KTR_CXGBE,
343 	    "assign_rxopt: tid %d, mtu_idx %u, emss %u, ts %u, sack %u, wscale %u",
344 	    toep->tid, toep->params.mtu_idx, toep->params.emss,
345 	    toep->params.tstamp, toep->params.sack, toep->params.wscale);
346 }
347 
348 /*
349  * Completes some final bits of initialization for just established connections
350  * and changes their state to TCPS_ESTABLISHED.
351  *
352  * The ISNs are from the exchange of SYNs.
353  */
354 void
355 make_established(struct toepcb *toep, uint32_t iss, uint32_t irs, uint16_t opt)
356 {
357 	struct inpcb *inp = toep->inp;
358 	struct socket *so = inp->inp_socket;
359 	struct tcpcb *tp = intotcpcb(inp);
360 	uint16_t tcpopt = be16toh(opt);
361 
362 	INP_WLOCK_ASSERT(inp);
363 	KASSERT(tp->t_state == TCPS_SYN_SENT ||
364 	    tp->t_state == TCPS_SYN_RECEIVED,
365 	    ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
366 
367 	CTR6(KTR_CXGBE, "%s: tid %d, so %p, inp %p, tp %p, toep %p",
368 	    __func__, toep->tid, so, inp, tp, toep);
369 
370 	tcp_state_change(tp, TCPS_ESTABLISHED);
371 	tp->t_starttime = ticks;
372 	TCPSTAT_INC(tcps_connects);
373 
374 	tp->irs = irs;
375 	tcp_rcvseqinit(tp);
376 	tp->rcv_wnd = (u_int)toep->params.opt0_bufsize << 10;
377 	tp->rcv_adv += tp->rcv_wnd;
378 	tp->last_ack_sent = tp->rcv_nxt;
379 
380 	tp->iss = iss;
381 	tcp_sendseqinit(tp);
382 	tp->snd_una = iss + 1;
383 	tp->snd_nxt = iss + 1;
384 	tp->snd_max = iss + 1;
385 
386 	assign_rxopt(tp, tcpopt);
387 	send_flowc_wr(toep, tp);
388 
389 	soisconnected(so);
390 }
391 
392 int
393 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
394 {
395 	struct wrqe *wr;
396 	struct cpl_rx_data_ack *req;
397 	uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
398 
399 	KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
400 
401 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
402 	if (wr == NULL)
403 		return (0);
404 	req = wrtod(wr);
405 
406 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
407 	req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
408 
409 	t4_wrq_tx(sc, wr);
410 	return (credits);
411 }
412 
413 void
414 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp)
415 {
416 	struct adapter *sc = tod->tod_softc;
417 	struct inpcb *inp = tptoinpcb(tp);
418 	struct socket *so = inp->inp_socket;
419 	struct sockbuf *sb = &so->so_rcv;
420 	struct toepcb *toep = tp->t_toe;
421 	int rx_credits;
422 
423 	INP_WLOCK_ASSERT(inp);
424 	SOCKBUF_LOCK_ASSERT(sb);
425 
426 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
427 	if (rx_credits > 0 &&
428 	    (tp->rcv_wnd <= 32 * 1024 || rx_credits >= 64 * 1024 ||
429 	    (rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
430 	    sbused(sb) + tp->rcv_wnd < sb->sb_lowat)) {
431 		rx_credits = send_rx_credits(sc, toep, rx_credits);
432 		tp->rcv_wnd += rx_credits;
433 		tp->rcv_adv += rx_credits;
434 	}
435 }
436 
437 void
438 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
439 {
440 	struct inpcb *inp = tptoinpcb(tp);
441 	struct socket *so = inp->inp_socket;
442 	struct sockbuf *sb = &so->so_rcv;
443 
444 	SOCKBUF_LOCK(sb);
445 	t4_rcvd_locked(tod, tp);
446 	SOCKBUF_UNLOCK(sb);
447 }
448 
449 /*
450  * Close a connection by sending a CPL_CLOSE_CON_REQ message.
451  */
452 int
453 t4_close_conn(struct adapter *sc, struct toepcb *toep)
454 {
455 	struct wrqe *wr;
456 	struct cpl_close_con_req *req;
457 	unsigned int tid = toep->tid;
458 
459 	CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
460 	    toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
461 
462 	if (toep->flags & TPF_FIN_SENT)
463 		return (0);
464 
465 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
466 	    ("%s: flowc_wr not sent for tid %u.", __func__, tid));
467 
468 	wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq);
469 	if (wr == NULL) {
470 		/* XXX */
471 		panic("%s: allocation failure.", __func__);
472 	}
473 	req = wrtod(wr);
474 
475         req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
476 	    V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
477 	req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
478 	    V_FW_WR_FLOWID(tid));
479         req->wr.wr_lo = cpu_to_be64(0);
480         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
481 	req->rsvd = 0;
482 
483 	toep->flags |= TPF_FIN_SENT;
484 	toep->flags &= ~TPF_SEND_FIN;
485 	t4_l2t_send(sc, wr, toep->l2te);
486 
487 	return (0);
488 }
489 
490 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
491 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
492 #define MIN_ISO_TX_CREDITS  (howmany(sizeof(struct cpl_tx_data_iso), 16))
493 #define MIN_TX_CREDITS(iso)						\
494 	(MIN_OFLD_TX_CREDITS + ((iso) ? MIN_ISO_TX_CREDITS : 0))
495 
496 /* Maximum amount of immediate data we could stuff in a WR */
497 static inline int
498 max_imm_payload(int tx_credits, int iso)
499 {
500 	const int iso_cpl_size = iso ? sizeof(struct cpl_tx_data_iso) : 0;
501 	const int n = 1;	/* Use no more than one desc for imm. data WR */
502 
503 	KASSERT(tx_credits >= 0 &&
504 		tx_credits <= MAX_OFLD_TX_CREDITS,
505 		("%s: %d credits", __func__, tx_credits));
506 
507 	if (tx_credits < MIN_TX_CREDITS(iso))
508 		return (0);
509 
510 	if (tx_credits >= (n * EQ_ESIZE) / 16)
511 		return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr) -
512 		    iso_cpl_size);
513 	else
514 		return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr) -
515 		    iso_cpl_size);
516 }
517 
518 /* Maximum number of SGL entries we could stuff in a WR */
519 static inline int
520 max_dsgl_nsegs(int tx_credits, int iso)
521 {
522 	int nseg = 1;	/* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
523 	int sge_pair_credits = tx_credits - MIN_TX_CREDITS(iso);
524 
525 	KASSERT(tx_credits >= 0 &&
526 		tx_credits <= MAX_OFLD_TX_CREDITS,
527 		("%s: %d credits", __func__, tx_credits));
528 
529 	if (tx_credits < MIN_TX_CREDITS(iso))
530 		return (0);
531 
532 	nseg += 2 * (sge_pair_credits * 16 / 24);
533 	if ((sge_pair_credits * 16) % 24 == 16)
534 		nseg++;
535 
536 	return (nseg);
537 }
538 
539 static inline void
540 write_tx_wr(void *dst, struct toepcb *toep, int fw_wr_opcode,
541     unsigned int immdlen, unsigned int plen, uint8_t credits, int shove,
542     int ulp_submode)
543 {
544 	struct fw_ofld_tx_data_wr *txwr = dst;
545 
546 	txwr->op_to_immdlen = htobe32(V_WR_OP(fw_wr_opcode) |
547 	    V_FW_WR_IMMDLEN(immdlen));
548 	txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
549 	    V_FW_WR_LEN16(credits));
550 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ulp_mode(toep)) |
551 	    V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
552 	txwr->plen = htobe32(plen);
553 
554 	if (toep->params.tx_align > 0) {
555 		if (plen < 2 * toep->params.emss)
556 			txwr->lsodisable_to_flags |=
557 			    htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
558 		else
559 			txwr->lsodisable_to_flags |=
560 			    htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
561 				(toep->params.nagle == 0 ? 0 :
562 				F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
563 	}
564 }
565 
566 /*
567  * Generate a DSGL from a starting mbuf.  The total number of segments and the
568  * maximum segments in any one mbuf are provided.
569  */
570 static void
571 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
572 {
573 	struct mbuf *m;
574 	struct ulptx_sgl *usgl = dst;
575 	int i, j, rc;
576 	struct sglist sg;
577 	struct sglist_seg segs[n];
578 
579 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
580 
581 	sglist_init(&sg, n, segs);
582 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
583 	    V_ULPTX_NSGE(nsegs));
584 
585 	i = -1;
586 	for (m = start; m != stop; m = m->m_next) {
587 		if (m->m_flags & M_EXTPG)
588 			rc = sglist_append_mbuf_epg(&sg, m,
589 			    mtod(m, vm_offset_t), m->m_len);
590 		else
591 			rc = sglist_append(&sg, mtod(m, void *), m->m_len);
592 		if (__predict_false(rc != 0))
593 			panic("%s: sglist_append %d", __func__, rc);
594 
595 		for (j = 0; j < sg.sg_nseg; i++, j++) {
596 			if (i < 0) {
597 				usgl->len0 = htobe32(segs[j].ss_len);
598 				usgl->addr0 = htobe64(segs[j].ss_paddr);
599 			} else {
600 				usgl->sge[i / 2].len[i & 1] =
601 				    htobe32(segs[j].ss_len);
602 				usgl->sge[i / 2].addr[i & 1] =
603 				    htobe64(segs[j].ss_paddr);
604 			}
605 #ifdef INVARIANTS
606 			nsegs--;
607 #endif
608 		}
609 		sglist_reset(&sg);
610 	}
611 	if (i & 1)
612 		usgl->sge[i / 2].len[1] = htobe32(0);
613 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
614 	    __func__, nsegs, start, stop));
615 }
616 
617 /*
618  * Max number of SGL entries an offload tx work request can have.  This is 41
619  * (1 + 40) for a full 512B work request.
620  * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
621  */
622 #define OFLD_SGL_LEN (41)
623 
624 /*
625  * Send data and/or a FIN to the peer.
626  *
627  * The socket's so_snd buffer consists of a stream of data starting with sb_mb
628  * and linked together with m_next.  sb_sndptr, if set, is the last mbuf that
629  * was transmitted.
630  *
631  * drop indicates the number of bytes that should be dropped from the head of
632  * the send buffer.  It is an optimization that lets do_fw4_ack avoid creating
633  * contention on the send buffer lock (before this change it used to do
634  * sowwakeup and then t4_push_frames right after that when recovering from tx
635  * stalls).  When drop is set this function MUST drop the bytes and wake up any
636  * writers.
637  */
638 void
639 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
640 {
641 	struct mbuf *sndptr, *m, *sb_sndptr;
642 	struct fw_ofld_tx_data_wr *txwr;
643 	struct wrqe *wr;
644 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
645 	struct inpcb *inp = toep->inp;
646 	struct tcpcb *tp = intotcpcb(inp);
647 	struct socket *so = inp->inp_socket;
648 	struct sockbuf *sb = &so->so_snd;
649 	int tx_credits, shove, compl, sowwakeup;
650 	struct ofld_tx_sdesc *txsd;
651 	bool nomap_mbuf_seen;
652 
653 	INP_WLOCK_ASSERT(inp);
654 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
655 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
656 
657 	KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
658 	    ulp_mode(toep) == ULP_MODE_TCPDDP ||
659 	    ulp_mode(toep) == ULP_MODE_TLS ||
660 	    ulp_mode(toep) == ULP_MODE_RDMA,
661 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
662 
663 #ifdef VERBOSE_TRACES
664 	CTR5(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
665 	    __func__, toep->tid, toep->flags, tp->t_flags, drop);
666 #endif
667 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
668 		return;
669 
670 #ifdef RATELIMIT
671 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
672 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
673 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
674 	}
675 #endif
676 
677 	/*
678 	 * This function doesn't resume by itself.  Someone else must clear the
679 	 * flag and call this function.
680 	 */
681 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
682 		KASSERT(drop == 0,
683 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
684 		return;
685 	}
686 
687 	txsd = &toep->txsd[toep->txsd_pidx];
688 	do {
689 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
690 		max_imm = max_imm_payload(tx_credits, 0);
691 		max_nsegs = max_dsgl_nsegs(tx_credits, 0);
692 
693 		SOCKBUF_LOCK(sb);
694 		sowwakeup = drop;
695 		if (drop) {
696 			sbdrop_locked(sb, drop);
697 			drop = 0;
698 		}
699 		sb_sndptr = sb->sb_sndptr;
700 		sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
701 		plen = 0;
702 		nsegs = 0;
703 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
704 		nomap_mbuf_seen = false;
705 		for (m = sndptr; m != NULL; m = m->m_next) {
706 			int n;
707 
708 			if ((m->m_flags & M_NOTAVAIL) != 0)
709 				break;
710 			if (m->m_flags & M_EXTPG) {
711 #ifdef KERN_TLS
712 				if (m->m_epg_tls != NULL) {
713 					toep->flags |= TPF_KTLS;
714 					if (plen == 0) {
715 						SOCKBUF_UNLOCK(sb);
716 						t4_push_ktls(sc, toep, 0);
717 						return;
718 					}
719 					break;
720 				}
721 #endif
722 				n = sglist_count_mbuf_epg(m,
723 				    mtod(m, vm_offset_t), m->m_len);
724 			} else
725 				n = sglist_count(mtod(m, void *), m->m_len);
726 
727 			nsegs += n;
728 			plen += m->m_len;
729 
730 			/* This mbuf sent us _over_ the nsegs limit, back out */
731 			if (plen > max_imm && nsegs > max_nsegs) {
732 				nsegs -= n;
733 				plen -= m->m_len;
734 				if (plen == 0) {
735 					/* Too few credits */
736 					toep->flags |= TPF_TX_SUSPENDED;
737 					if (sowwakeup) {
738 						if (!TAILQ_EMPTY(
739 						    &toep->aiotx_jobq))
740 							t4_aiotx_queue_toep(so,
741 							    toep);
742 						sowwakeup_locked(so);
743 					} else
744 						SOCKBUF_UNLOCK(sb);
745 					SOCKBUF_UNLOCK_ASSERT(sb);
746 					return;
747 				}
748 				break;
749 			}
750 
751 			if (m->m_flags & M_EXTPG)
752 				nomap_mbuf_seen = true;
753 			if (max_nsegs_1mbuf < n)
754 				max_nsegs_1mbuf = n;
755 			sb_sndptr = m;	/* new sb->sb_sndptr if all goes well */
756 
757 			/* This mbuf put us right at the max_nsegs limit */
758 			if (plen > max_imm && nsegs == max_nsegs) {
759 				m = m->m_next;
760 				break;
761 			}
762 		}
763 
764 		if (sbused(sb) > sb->sb_hiwat * 5 / 8 &&
765 		    toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
766 			compl = 1;
767 		else
768 			compl = 0;
769 
770 		if (sb->sb_flags & SB_AUTOSIZE &&
771 		    V_tcp_do_autosndbuf &&
772 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
773 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
774 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
775 			    V_tcp_autosndbuf_max);
776 
777 			if (!sbreserve_locked(so, SO_SND, newsize, NULL))
778 				sb->sb_flags &= ~SB_AUTOSIZE;
779 			else
780 				sowwakeup = 1;	/* room available */
781 		}
782 		if (sowwakeup) {
783 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
784 				t4_aiotx_queue_toep(so, toep);
785 			sowwakeup_locked(so);
786 		} else
787 			SOCKBUF_UNLOCK(sb);
788 		SOCKBUF_UNLOCK_ASSERT(sb);
789 
790 		/* nothing to send */
791 		if (plen == 0) {
792 			KASSERT(m == NULL || (m->m_flags & M_NOTAVAIL) != 0,
793 			    ("%s: nothing to send, but m != NULL is ready",
794 			    __func__));
795 			break;
796 		}
797 
798 		if (__predict_false(toep->flags & TPF_FIN_SENT))
799 			panic("%s: excess tx.", __func__);
800 
801 		shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
802 		if (plen <= max_imm && !nomap_mbuf_seen) {
803 
804 			/* Immediate data tx */
805 
806 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
807 					&toep->ofld_txq->wrq);
808 			if (wr == NULL) {
809 				/* XXX: how will we recover from this? */
810 				toep->flags |= TPF_TX_SUSPENDED;
811 				return;
812 			}
813 			txwr = wrtod(wr);
814 			credits = howmany(wr->wr_len, 16);
815 			write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, plen, plen,
816 			    credits, shove, 0);
817 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
818 			nsegs = 0;
819 		} else {
820 			int wr_len;
821 
822 			/* DSGL tx */
823 
824 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
825 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
826 			wr = alloc_wrqe(roundup2(wr_len, 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_len, 16);
835 			write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, 0, plen,
836 			    credits, shove, 0);
837 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
838 			    max_nsegs_1mbuf);
839 			if (wr_len & 0xf) {
840 				uint64_t *pad = (uint64_t *)
841 				    ((uintptr_t)txwr + wr_len);
842 				*pad = 0;
843 			}
844 		}
845 
846 		KASSERT(toep->tx_credits >= credits,
847 			("%s: not enough credits", __func__));
848 
849 		toep->tx_credits -= credits;
850 		toep->tx_nocompl += credits;
851 		toep->plen_nocompl += plen;
852 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
853 		    toep->tx_nocompl >= toep->tx_total / 4)
854 			compl = 1;
855 
856 		if (compl || ulp_mode(toep) == ULP_MODE_RDMA) {
857 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
858 			toep->tx_nocompl = 0;
859 			toep->plen_nocompl = 0;
860 		}
861 
862 		tp->snd_nxt += plen;
863 		tp->snd_max += plen;
864 
865 		SOCKBUF_LOCK(sb);
866 		KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
867 		sb->sb_sndptr = sb_sndptr;
868 		SOCKBUF_UNLOCK(sb);
869 
870 		toep->flags |= TPF_TX_DATA_SENT;
871 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
872 			toep->flags |= TPF_TX_SUSPENDED;
873 
874 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
875 		txsd->plen = plen;
876 		txsd->tx_credits = credits;
877 		txsd++;
878 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
879 			toep->txsd_pidx = 0;
880 			txsd = &toep->txsd[0];
881 		}
882 		toep->txsd_avail--;
883 
884 		t4_l2t_send(sc, wr, toep->l2te);
885 	} while (m != NULL && (m->m_flags & M_NOTAVAIL) == 0);
886 
887 	/* Send a FIN if requested, but only if there's no more data to send */
888 	if (m == NULL && toep->flags & TPF_SEND_FIN)
889 		t4_close_conn(sc, toep);
890 }
891 
892 static inline void
893 rqdrop_locked(struct mbufq *q, int plen)
894 {
895 	struct mbuf *m;
896 
897 	while (plen > 0) {
898 		m = mbufq_dequeue(q);
899 
900 		/* Too many credits. */
901 		MPASS(m != NULL);
902 		M_ASSERTPKTHDR(m);
903 
904 		/* Partial credits. */
905 		MPASS(plen >= m->m_pkthdr.len);
906 
907 		plen -= m->m_pkthdr.len;
908 		m_freem(m);
909 	}
910 }
911 
912 /*
913  * Not a bit in the TCB, but is a bit in the ulp_submode field of the
914  * CPL_TX_DATA flags field in FW_ISCSI_TX_DATA_WR.
915  */
916 #define	ULP_ISO		G_TX_ULP_SUBMODE(F_FW_ISCSI_TX_DATA_WR_ULPSUBMODE_ISO)
917 
918 static void
919 write_tx_data_iso(void *dst, u_int ulp_submode, uint8_t flags, uint16_t mss,
920     int len, int npdu)
921 {
922 	struct cpl_tx_data_iso *cpl;
923 	unsigned int burst_size;
924 	unsigned int last;
925 
926 	/*
927 	 * The firmware will set the 'F' bit on the last PDU when
928 	 * either condition is true:
929 	 *
930 	 * - this large PDU is marked as the "last" slice
931 	 *
932 	 * - the amount of data payload bytes equals the burst_size
933 	 *
934 	 * The strategy used here is to always set the burst_size
935 	 * artificially high (len includes the size of the template
936 	 * BHS) and only set the "last" flag if the original PDU had
937 	 * 'F' set.
938 	 */
939 	burst_size = len;
940 	last = !!(flags & CXGBE_ISO_F);
941 
942 	cpl = (struct cpl_tx_data_iso *)dst;
943 	cpl->op_to_scsi = htonl(V_CPL_TX_DATA_ISO_OP(CPL_TX_DATA_ISO) |
944 	    V_CPL_TX_DATA_ISO_FIRST(1) | V_CPL_TX_DATA_ISO_LAST(last) |
945 	    V_CPL_TX_DATA_ISO_CPLHDRLEN(0) |
946 	    V_CPL_TX_DATA_ISO_HDRCRC(!!(ulp_submode & ULP_CRC_HEADER)) |
947 	    V_CPL_TX_DATA_ISO_PLDCRC(!!(ulp_submode & ULP_CRC_DATA)) |
948 	    V_CPL_TX_DATA_ISO_IMMEDIATE(0) |
949 	    V_CPL_TX_DATA_ISO_SCSI(CXGBE_ISO_TYPE(flags)));
950 
951 	cpl->ahs_len = 0;
952 	cpl->mpdu = htons(DIV_ROUND_UP(mss, 4));
953 	cpl->burst_size = htonl(DIV_ROUND_UP(burst_size, 4));
954 	cpl->len = htonl(len);
955 	cpl->reserved2_seglen_offset = htonl(0);
956 	cpl->datasn_offset = htonl(0);
957 	cpl->buffer_offset = htonl(0);
958 	cpl->reserved3 = 0;
959 }
960 
961 static struct wrqe *
962 write_iscsi_mbuf_wr(struct toepcb *toep, struct mbuf *sndptr)
963 {
964 	struct mbuf *m;
965 	struct fw_ofld_tx_data_wr *txwr;
966 	struct cpl_tx_data_iso *cpl_iso;
967 	void *p;
968 	struct wrqe *wr;
969 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
970 	u_int adjusted_plen, imm_data, ulp_submode;
971 	struct inpcb *inp = toep->inp;
972 	struct tcpcb *tp = intotcpcb(inp);
973 	int tx_credits, shove, npdu, wr_len;
974 	uint16_t iso_mss;
975 	static const u_int ulp_extra_len[] = {0, 4, 4, 8};
976 	bool iso, nomap_mbuf_seen;
977 
978 	M_ASSERTPKTHDR(sndptr);
979 
980 	tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
981 	if (mbuf_raw_wr(sndptr)) {
982 		plen = sndptr->m_pkthdr.len;
983 		KASSERT(plen <= SGE_MAX_WR_LEN,
984 		    ("raw WR len %u is greater than max WR len", plen));
985 		if (plen > tx_credits * 16)
986 			return (NULL);
987 
988 		wr = alloc_wrqe(roundup2(plen, 16), &toep->ofld_txq->wrq);
989 		if (__predict_false(wr == NULL))
990 			return (NULL);
991 
992 		m_copydata(sndptr, 0, plen, wrtod(wr));
993 		return (wr);
994 	}
995 
996 	iso = mbuf_iscsi_iso(sndptr);
997 	max_imm = max_imm_payload(tx_credits, iso);
998 	max_nsegs = max_dsgl_nsegs(tx_credits, iso);
999 	iso_mss = mbuf_iscsi_iso_mss(sndptr);
1000 
1001 	plen = 0;
1002 	nsegs = 0;
1003 	max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
1004 	nomap_mbuf_seen = false;
1005 	for (m = sndptr; m != NULL; m = m->m_next) {
1006 		int n;
1007 
1008 		if (m->m_flags & M_EXTPG)
1009 			n = sglist_count_mbuf_epg(m, mtod(m, vm_offset_t),
1010 			    m->m_len);
1011 		else
1012 			n = sglist_count(mtod(m, void *), m->m_len);
1013 
1014 		nsegs += n;
1015 		plen += m->m_len;
1016 
1017 		/*
1018 		 * This mbuf would send us _over_ the nsegs limit.
1019 		 * Suspend tx because the PDU can't be sent out.
1020 		 */
1021 		if ((nomap_mbuf_seen || plen > max_imm) && nsegs > max_nsegs)
1022 			return (NULL);
1023 
1024 		if (m->m_flags & M_EXTPG)
1025 			nomap_mbuf_seen = true;
1026 		if (max_nsegs_1mbuf < n)
1027 			max_nsegs_1mbuf = n;
1028 	}
1029 
1030 	if (__predict_false(toep->flags & TPF_FIN_SENT))
1031 		panic("%s: excess tx.", __func__);
1032 
1033 	/*
1034 	 * We have a PDU to send.  All of it goes out in one WR so 'm'
1035 	 * is NULL.  A PDU's length is always a multiple of 4.
1036 	 */
1037 	MPASS(m == NULL);
1038 	MPASS((plen & 3) == 0);
1039 	MPASS(sndptr->m_pkthdr.len == plen);
1040 
1041 	shove = !(tp->t_flags & TF_MORETOCOME);
1042 
1043 	/*
1044 	 * plen doesn't include header and data digests, which are
1045 	 * generated and inserted in the right places by the TOE, but
1046 	 * they do occupy TCP sequence space and need to be accounted
1047 	 * for.
1048 	 */
1049 	ulp_submode = mbuf_ulp_submode(sndptr);
1050 	MPASS(ulp_submode < nitems(ulp_extra_len));
1051 	npdu = iso ? howmany(plen - ISCSI_BHS_SIZE, iso_mss) : 1;
1052 	adjusted_plen = plen + ulp_extra_len[ulp_submode] * npdu;
1053 	if (iso)
1054 		adjusted_plen += ISCSI_BHS_SIZE * (npdu - 1);
1055 	wr_len = sizeof(*txwr);
1056 	if (iso)
1057 		wr_len += sizeof(struct cpl_tx_data_iso);
1058 	if (plen <= max_imm && !nomap_mbuf_seen) {
1059 		/* Immediate data tx */
1060 		imm_data = plen;
1061 		wr_len += plen;
1062 		nsegs = 0;
1063 	} else {
1064 		/* DSGL tx */
1065 		imm_data = 0;
1066 		wr_len += sizeof(struct ulptx_sgl) +
1067 		    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1068 	}
1069 
1070 	wr = alloc_wrqe(roundup2(wr_len, 16), &toep->ofld_txq->wrq);
1071 	if (wr == NULL) {
1072 		/* XXX: how will we recover from this? */
1073 		return (NULL);
1074 	}
1075 	txwr = wrtod(wr);
1076 	credits = howmany(wr->wr_len, 16);
1077 
1078 	if (iso) {
1079 		write_tx_wr(txwr, toep, FW_ISCSI_TX_DATA_WR,
1080 		    imm_data + sizeof(struct cpl_tx_data_iso),
1081 		    adjusted_plen, credits, shove, ulp_submode | ULP_ISO);
1082 		cpl_iso = (struct cpl_tx_data_iso *)(txwr + 1);
1083 		MPASS(plen == sndptr->m_pkthdr.len);
1084 		write_tx_data_iso(cpl_iso, ulp_submode,
1085 		    mbuf_iscsi_iso_flags(sndptr), iso_mss, plen, npdu);
1086 		p = cpl_iso + 1;
1087 	} else {
1088 		write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, imm_data,
1089 		    adjusted_plen, credits, shove, ulp_submode);
1090 		p = txwr + 1;
1091 	}
1092 
1093 	if (imm_data != 0) {
1094 		m_copydata(sndptr, 0, plen, p);
1095 	} else {
1096 		write_tx_sgl(p, sndptr, m, nsegs, max_nsegs_1mbuf);
1097 		if (wr_len & 0xf) {
1098 			uint64_t *pad = (uint64_t *)((uintptr_t)txwr + wr_len);
1099 			*pad = 0;
1100 		}
1101 	}
1102 
1103 	KASSERT(toep->tx_credits >= credits,
1104 	    ("%s: not enough credits: credits %u "
1105 		"toep->tx_credits %u tx_credits %u nsegs %u "
1106 		"max_nsegs %u iso %d", __func__, credits,
1107 		toep->tx_credits, tx_credits, nsegs, max_nsegs, iso));
1108 
1109 	tp->snd_nxt += adjusted_plen;
1110 	tp->snd_max += adjusted_plen;
1111 
1112 	counter_u64_add(toep->ofld_txq->tx_iscsi_pdus, npdu);
1113 	counter_u64_add(toep->ofld_txq->tx_iscsi_octets, plen);
1114 	if (iso)
1115 		counter_u64_add(toep->ofld_txq->tx_iscsi_iso_wrs, 1);
1116 
1117 	return (wr);
1118 }
1119 
1120 void
1121 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
1122 {
1123 	struct mbuf *sndptr, *m;
1124 	struct fw_wr_hdr *wrhdr;
1125 	struct wrqe *wr;
1126 	u_int plen, credits;
1127 	struct inpcb *inp = toep->inp;
1128 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
1129 	struct mbufq *pduq = &toep->ulp_pduq;
1130 
1131 	INP_WLOCK_ASSERT(inp);
1132 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1133 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1134 	KASSERT(ulp_mode(toep) == ULP_MODE_ISCSI,
1135 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1136 
1137 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1138 		return;
1139 
1140 	/*
1141 	 * This function doesn't resume by itself.  Someone else must clear the
1142 	 * flag and call this function.
1143 	 */
1144 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1145 		KASSERT(drop == 0,
1146 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1147 		return;
1148 	}
1149 
1150 	if (drop) {
1151 		struct socket *so = inp->inp_socket;
1152 		struct sockbuf *sb = &so->so_snd;
1153 		int sbu;
1154 
1155 		/*
1156 		 * An unlocked read is ok here as the data should only
1157 		 * transition from a non-zero value to either another
1158 		 * non-zero value or zero.  Once it is zero it should
1159 		 * stay zero.
1160 		 */
1161 		if (__predict_false(sbused(sb)) > 0) {
1162 			SOCKBUF_LOCK(sb);
1163 			sbu = sbused(sb);
1164 			if (sbu > 0) {
1165 				/*
1166 				 * The data transmitted before the
1167 				 * tid's ULP mode changed to ISCSI is
1168 				 * still in so_snd.  Incoming credits
1169 				 * should account for so_snd first.
1170 				 */
1171 				sbdrop_locked(sb, min(sbu, drop));
1172 				drop -= min(sbu, drop);
1173 			}
1174 			sowwakeup_locked(so);	/* unlocks so_snd */
1175 		}
1176 		rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
1177 	}
1178 
1179 	while ((sndptr = mbufq_first(pduq)) != NULL) {
1180 		wr = write_iscsi_mbuf_wr(toep, sndptr);
1181 		if (wr == NULL) {
1182 			toep->flags |= TPF_TX_SUSPENDED;
1183 			return;
1184 		}
1185 
1186 		plen = sndptr->m_pkthdr.len;
1187 		credits = howmany(wr->wr_len, 16);
1188 		KASSERT(toep->tx_credits >= credits,
1189 			("%s: not enough credits", __func__));
1190 
1191 		m = mbufq_dequeue(pduq);
1192 		MPASS(m == sndptr);
1193 		mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
1194 
1195 		toep->tx_credits -= credits;
1196 		toep->tx_nocompl += credits;
1197 		toep->plen_nocompl += plen;
1198 
1199 		/*
1200 		 * Ensure there are enough credits for a full-sized WR
1201 		 * as page pod WRs can be full-sized.
1202 		 */
1203 		if (toep->tx_credits <= SGE_MAX_WR_LEN * 5 / 4 &&
1204 		    toep->tx_nocompl >= toep->tx_total / 4) {
1205 			wrhdr = wrtod(wr);
1206 			wrhdr->hi |= htobe32(F_FW_WR_COMPL);
1207 			toep->tx_nocompl = 0;
1208 			toep->plen_nocompl = 0;
1209 		}
1210 
1211 		toep->flags |= TPF_TX_DATA_SENT;
1212 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
1213 			toep->flags |= TPF_TX_SUSPENDED;
1214 
1215 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1216 		txsd->plen = plen;
1217 		txsd->tx_credits = credits;
1218 		txsd++;
1219 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1220 			toep->txsd_pidx = 0;
1221 			txsd = &toep->txsd[0];
1222 		}
1223 		toep->txsd_avail--;
1224 
1225 		t4_l2t_send(sc, wr, toep->l2te);
1226 	}
1227 
1228 	/* Send a FIN if requested, but only if there are no more PDUs to send */
1229 	if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
1230 		t4_close_conn(sc, toep);
1231 }
1232 
1233 static inline void
1234 t4_push_data(struct adapter *sc, struct toepcb *toep, int drop)
1235 {
1236 
1237 	if (ulp_mode(toep) == ULP_MODE_ISCSI)
1238 		t4_push_pdus(sc, toep, drop);
1239 	else if (toep->flags & TPF_KTLS)
1240 		t4_push_ktls(sc, toep, drop);
1241 	else
1242 		t4_push_frames(sc, toep, drop);
1243 }
1244 
1245 int
1246 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
1247 {
1248 	struct adapter *sc = tod->tod_softc;
1249 #ifdef INVARIANTS
1250 	struct inpcb *inp = tptoinpcb(tp);
1251 #endif
1252 	struct toepcb *toep = tp->t_toe;
1253 
1254 	INP_WLOCK_ASSERT(inp);
1255 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1256 	    ("%s: inp %p dropped.", __func__, inp));
1257 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1258 
1259 	t4_push_data(sc, toep, 0);
1260 
1261 	return (0);
1262 }
1263 
1264 int
1265 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
1266 {
1267 	struct adapter *sc = tod->tod_softc;
1268 #ifdef INVARIANTS
1269 	struct inpcb *inp = tptoinpcb(tp);
1270 #endif
1271 	struct toepcb *toep = tp->t_toe;
1272 
1273 	INP_WLOCK_ASSERT(inp);
1274 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1275 	    ("%s: inp %p dropped.", __func__, inp));
1276 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1277 
1278 	toep->flags |= TPF_SEND_FIN;
1279 	if (tp->t_state >= TCPS_ESTABLISHED)
1280 		t4_push_data(sc, toep, 0);
1281 
1282 	return (0);
1283 }
1284 
1285 int
1286 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1287 {
1288 	struct adapter *sc = tod->tod_softc;
1289 #if defined(INVARIANTS)
1290 	struct inpcb *inp = tptoinpcb(tp);
1291 #endif
1292 	struct toepcb *toep = tp->t_toe;
1293 
1294 	INP_WLOCK_ASSERT(inp);
1295 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1296 	    ("%s: inp %p dropped.", __func__, inp));
1297 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1298 
1299 	/* hmmmm */
1300 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1301 	    ("%s: flowc for tid %u [%s] not sent already",
1302 	    __func__, toep->tid, tcpstates[tp->t_state]));
1303 
1304 	send_reset(sc, toep, 0);
1305 	return (0);
1306 }
1307 
1308 /*
1309  * Peer has sent us a FIN.
1310  */
1311 static int
1312 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1313 {
1314 	struct adapter *sc = iq->adapter;
1315 	const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1316 	unsigned int tid = GET_TID(cpl);
1317 	struct toepcb *toep = lookup_tid(sc, tid);
1318 	struct inpcb *inp = toep->inp;
1319 	struct tcpcb *tp = NULL;
1320 	struct socket *so;
1321 	struct epoch_tracker et;
1322 #ifdef INVARIANTS
1323 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1324 #endif
1325 
1326 	KASSERT(opcode == CPL_PEER_CLOSE,
1327 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1328 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1329 
1330 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1331 		/*
1332 		 * do_pass_establish must have run before do_peer_close and if
1333 		 * this is still a synqe instead of a toepcb then the connection
1334 		 * must be getting aborted.
1335 		 */
1336 		MPASS(toep->flags & TPF_ABORT_SHUTDOWN);
1337 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1338 		    toep, toep->flags);
1339 		return (0);
1340 	}
1341 
1342 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1343 
1344 	CURVNET_SET(toep->vnet);
1345 	NET_EPOCH_ENTER(et);
1346 	INP_WLOCK(inp);
1347 	tp = intotcpcb(inp);
1348 
1349 	CTR6(KTR_CXGBE,
1350 	    "%s: tid %u (%s), toep_flags 0x%x, ddp_flags 0x%x, inp %p",
1351 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1352 	    toep->ddp.flags, inp);
1353 
1354 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1355 		goto done;
1356 
1357 	so = inp->inp_socket;
1358 	socantrcvmore(so);
1359 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1360 		DDP_LOCK(toep);
1361 		if (__predict_false(toep->ddp.flags &
1362 		    (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)))
1363 			handle_ddp_close(toep, tp, cpl->rcv_nxt);
1364 		DDP_UNLOCK(toep);
1365 	}
1366 
1367 	if (ulp_mode(toep) == ULP_MODE_RDMA ||
1368 	    (ulp_mode(toep) == ULP_MODE_ISCSI && chip_id(sc) >= CHELSIO_T6)) {
1369 		/*
1370 		 * There might be data received via DDP before the FIN
1371 		 * not reported to the driver.  Just assume the
1372 		 * sequence number in the CPL is correct as the
1373 		 * sequence number of the FIN.
1374 		 */
1375 	} else {
1376 		KASSERT(tp->rcv_nxt + 1 == be32toh(cpl->rcv_nxt),
1377 		    ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1378 		    be32toh(cpl->rcv_nxt)));
1379 	}
1380 
1381 	tp->rcv_nxt = be32toh(cpl->rcv_nxt);
1382 
1383 	switch (tp->t_state) {
1384 	case TCPS_SYN_RECEIVED:
1385 		tp->t_starttime = ticks;
1386 		/* FALLTHROUGH */
1387 
1388 	case TCPS_ESTABLISHED:
1389 		tcp_state_change(tp, TCPS_CLOSE_WAIT);
1390 		break;
1391 
1392 	case TCPS_FIN_WAIT_1:
1393 		tcp_state_change(tp, TCPS_CLOSING);
1394 		break;
1395 
1396 	case TCPS_FIN_WAIT_2:
1397 		restore_so_proto(so, inp->inp_vflag & INP_IPV6);
1398 		tcp_twstart(tp);
1399 		INP_UNLOCK_ASSERT(inp);	 /* safe, we have a ref on the inp */
1400 		NET_EPOCH_EXIT(et);
1401 		CURVNET_RESTORE();
1402 
1403 		INP_WLOCK(inp);
1404 		final_cpl_received(toep);
1405 		return (0);
1406 
1407 	default:
1408 		log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1409 		    __func__, tid, tp->t_state);
1410 	}
1411 done:
1412 	INP_WUNLOCK(inp);
1413 	NET_EPOCH_EXIT(et);
1414 	CURVNET_RESTORE();
1415 	return (0);
1416 }
1417 
1418 /*
1419  * Peer has ACK'd our FIN.
1420  */
1421 static int
1422 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1423     struct mbuf *m)
1424 {
1425 	struct adapter *sc = iq->adapter;
1426 	const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1427 	unsigned int tid = GET_TID(cpl);
1428 	struct toepcb *toep = lookup_tid(sc, tid);
1429 	struct inpcb *inp = toep->inp;
1430 	struct tcpcb *tp = NULL;
1431 	struct socket *so = NULL;
1432 	struct epoch_tracker et;
1433 #ifdef INVARIANTS
1434 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1435 #endif
1436 
1437 	KASSERT(opcode == CPL_CLOSE_CON_RPL,
1438 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1439 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1440 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1441 
1442 	CURVNET_SET(toep->vnet);
1443 	NET_EPOCH_ENTER(et);
1444 	INP_WLOCK(inp);
1445 	tp = intotcpcb(inp);
1446 
1447 	CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1448 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1449 
1450 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1451 		goto done;
1452 
1453 	so = inp->inp_socket;
1454 	tp->snd_una = be32toh(cpl->snd_nxt) - 1;	/* exclude FIN */
1455 
1456 	switch (tp->t_state) {
1457 	case TCPS_CLOSING:	/* see TCPS_FIN_WAIT_2 in do_peer_close too */
1458 		restore_so_proto(so, inp->inp_vflag & INP_IPV6);
1459 		tcp_twstart(tp);
1460 release:
1461 		INP_UNLOCK_ASSERT(inp);	/* safe, we have a ref on the  inp */
1462 		NET_EPOCH_EXIT(et);
1463 		CURVNET_RESTORE();
1464 
1465 		INP_WLOCK(inp);
1466 		final_cpl_received(toep);	/* no more CPLs expected */
1467 
1468 		return (0);
1469 	case TCPS_LAST_ACK:
1470 		if (tcp_close(tp))
1471 			INP_WUNLOCK(inp);
1472 		goto release;
1473 
1474 	case TCPS_FIN_WAIT_1:
1475 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1476 			soisdisconnected(so);
1477 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
1478 		break;
1479 
1480 	default:
1481 		log(LOG_ERR,
1482 		    "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1483 		    __func__, tid, tcpstates[tp->t_state]);
1484 	}
1485 done:
1486 	INP_WUNLOCK(inp);
1487 	NET_EPOCH_EXIT(et);
1488 	CURVNET_RESTORE();
1489 	return (0);
1490 }
1491 
1492 void
1493 send_abort_rpl(struct adapter *sc, struct sge_ofld_txq *ofld_txq, int tid,
1494     int rst_status)
1495 {
1496 	struct wrqe *wr;
1497 	struct cpl_abort_rpl *cpl;
1498 
1499 	wr = alloc_wrqe(sizeof(*cpl), &ofld_txq->wrq);
1500 	if (wr == NULL) {
1501 		/* XXX */
1502 		panic("%s: allocation failure.", __func__);
1503 	}
1504 	cpl = wrtod(wr);
1505 
1506 	INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1507 	cpl->cmd = rst_status;
1508 
1509 	t4_wrq_tx(sc, wr);
1510 }
1511 
1512 static int
1513 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1514 {
1515 	switch (abort_reason) {
1516 	case CPL_ERR_BAD_SYN:
1517 	case CPL_ERR_CONN_RESET:
1518 		return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1519 	case CPL_ERR_XMIT_TIMEDOUT:
1520 	case CPL_ERR_PERSIST_TIMEDOUT:
1521 	case CPL_ERR_FINWAIT2_TIMEDOUT:
1522 	case CPL_ERR_KEEPALIVE_TIMEDOUT:
1523 		return (ETIMEDOUT);
1524 	default:
1525 		return (EIO);
1526 	}
1527 }
1528 
1529 /*
1530  * TCP RST from the peer, timeout, or some other such critical error.
1531  */
1532 static int
1533 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1534 {
1535 	struct adapter *sc = iq->adapter;
1536 	const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1537 	unsigned int tid = GET_TID(cpl);
1538 	struct toepcb *toep = lookup_tid(sc, tid);
1539 	struct sge_ofld_txq *ofld_txq = toep->ofld_txq;
1540 	struct inpcb *inp;
1541 	struct tcpcb *tp;
1542 	struct epoch_tracker et;
1543 #ifdef INVARIANTS
1544 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1545 #endif
1546 
1547 	KASSERT(opcode == CPL_ABORT_REQ_RSS,
1548 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1549 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1550 
1551 	if (toep->flags & TPF_SYNQE)
1552 		return (do_abort_req_synqe(iq, rss, m));
1553 
1554 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1555 
1556 	if (negative_advice(cpl->status)) {
1557 		CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1558 		    __func__, cpl->status, tid, toep->flags);
1559 		return (0);	/* Ignore negative advice */
1560 	}
1561 
1562 	inp = toep->inp;
1563 	CURVNET_SET(toep->vnet);
1564 	NET_EPOCH_ENTER(et);	/* for tcp_close */
1565 	INP_WLOCK(inp);
1566 
1567 	tp = intotcpcb(inp);
1568 
1569 	CTR6(KTR_CXGBE,
1570 	    "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1571 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1572 	    inp->inp_flags, cpl->status);
1573 
1574 	/*
1575 	 * If we'd initiated an abort earlier the reply to it is responsible for
1576 	 * cleaning up resources.  Otherwise we tear everything down right here
1577 	 * right now.  We owe the T4 a CPL_ABORT_RPL no matter what.
1578 	 */
1579 	if (toep->flags & TPF_ABORT_SHUTDOWN) {
1580 		INP_WUNLOCK(inp);
1581 		goto done;
1582 	}
1583 	toep->flags |= TPF_ABORT_SHUTDOWN;
1584 
1585 	if ((inp->inp_flags & INP_DROPPED) == 0) {
1586 		struct socket *so = inp->inp_socket;
1587 
1588 		if (so != NULL)
1589 			so_error_set(so, abort_status_to_errno(tp,
1590 			    cpl->status));
1591 		tp = tcp_close(tp);
1592 		if (tp == NULL)
1593 			INP_WLOCK(inp);	/* re-acquire */
1594 	}
1595 
1596 	final_cpl_received(toep);
1597 done:
1598 	NET_EPOCH_EXIT(et);
1599 	CURVNET_RESTORE();
1600 	send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1601 	return (0);
1602 }
1603 
1604 /*
1605  * Reply to the CPL_ABORT_REQ (send_reset)
1606  */
1607 static int
1608 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1609 {
1610 	struct adapter *sc = iq->adapter;
1611 	const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1612 	unsigned int tid = GET_TID(cpl);
1613 	struct toepcb *toep = lookup_tid(sc, tid);
1614 	struct inpcb *inp = toep->inp;
1615 #ifdef INVARIANTS
1616 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1617 #endif
1618 
1619 	KASSERT(opcode == CPL_ABORT_RPL_RSS,
1620 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1621 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1622 
1623 	if (toep->flags & TPF_SYNQE)
1624 		return (do_abort_rpl_synqe(iq, rss, m));
1625 
1626 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1627 
1628 	CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1629 	    __func__, tid, toep, inp, cpl->status);
1630 
1631 	KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1632 	    ("%s: wasn't expecting abort reply", __func__));
1633 
1634 	INP_WLOCK(inp);
1635 	final_cpl_received(toep);
1636 
1637 	return (0);
1638 }
1639 
1640 static int
1641 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1642 {
1643 	struct adapter *sc = iq->adapter;
1644 	const struct cpl_rx_data *cpl = mtod(m, const void *);
1645 	unsigned int tid = GET_TID(cpl);
1646 	struct toepcb *toep = lookup_tid(sc, tid);
1647 	struct inpcb *inp = toep->inp;
1648 	struct tcpcb *tp;
1649 	struct socket *so;
1650 	struct sockbuf *sb;
1651 	struct epoch_tracker et;
1652 	int len, rx_credits;
1653 	uint32_t ddp_placed = 0;
1654 
1655 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1656 		/*
1657 		 * do_pass_establish must have run before do_rx_data and if this
1658 		 * is still a synqe instead of a toepcb then the connection must
1659 		 * be getting aborted.
1660 		 */
1661 		MPASS(toep->flags & TPF_ABORT_SHUTDOWN);
1662 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1663 		    toep, toep->flags);
1664 		m_freem(m);
1665 		return (0);
1666 	}
1667 
1668 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1669 
1670 	/* strip off CPL header */
1671 	m_adj(m, sizeof(*cpl));
1672 	len = m->m_pkthdr.len;
1673 
1674 	INP_WLOCK(inp);
1675 	if (inp->inp_flags & INP_DROPPED) {
1676 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1677 		    __func__, tid, len, inp->inp_flags);
1678 		INP_WUNLOCK(inp);
1679 		m_freem(m);
1680 		return (0);
1681 	}
1682 
1683 	tp = intotcpcb(inp);
1684 
1685 	if (__predict_false(ulp_mode(toep) == ULP_MODE_TLS &&
1686 	   toep->flags & TPF_TLS_RECEIVE)) {
1687 		/* Received "raw" data on a TLS socket. */
1688 		CTR3(KTR_CXGBE, "%s: tid %u, raw TLS data (%d bytes)",
1689 		    __func__, tid, len);
1690 		do_rx_data_tls(cpl, toep, m);
1691 		return (0);
1692 	}
1693 
1694 	if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1695 		ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1696 
1697 	tp->rcv_nxt += len;
1698 	if (tp->rcv_wnd < len) {
1699 		KASSERT(ulp_mode(toep) == ULP_MODE_RDMA,
1700 				("%s: negative window size", __func__));
1701 	}
1702 
1703 	tp->rcv_wnd -= len;
1704 	tp->t_rcvtime = ticks;
1705 
1706 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1707 		DDP_LOCK(toep);
1708 	so = inp_inpcbtosocket(inp);
1709 	sb = &so->so_rcv;
1710 	SOCKBUF_LOCK(sb);
1711 
1712 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1713 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1714 		    __func__, tid, len);
1715 		m_freem(m);
1716 		SOCKBUF_UNLOCK(sb);
1717 		if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1718 			DDP_UNLOCK(toep);
1719 		INP_WUNLOCK(inp);
1720 
1721 		CURVNET_SET(toep->vnet);
1722 		NET_EPOCH_ENTER(et);
1723 		INP_WLOCK(inp);
1724 		tp = tcp_drop(tp, ECONNRESET);
1725 		if (tp)
1726 			INP_WUNLOCK(inp);
1727 		NET_EPOCH_EXIT(et);
1728 		CURVNET_RESTORE();
1729 
1730 		return (0);
1731 	}
1732 
1733 	/* receive buffer autosize */
1734 	MPASS(toep->vnet == so->so_vnet);
1735 	CURVNET_SET(toep->vnet);
1736 	if (sb->sb_flags & SB_AUTOSIZE &&
1737 	    V_tcp_do_autorcvbuf &&
1738 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1739 	    len > (sbspace(sb) / 8 * 7)) {
1740 		unsigned int hiwat = sb->sb_hiwat;
1741 		unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
1742 		    V_tcp_autorcvbuf_max);
1743 
1744 		if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
1745 			sb->sb_flags &= ~SB_AUTOSIZE;
1746 	}
1747 
1748 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1749 		int changed = !(toep->ddp.flags & DDP_ON) ^ cpl->ddp_off;
1750 
1751 		if (toep->ddp.waiting_count != 0 || toep->ddp.active_count != 0)
1752 			CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)",
1753 			    __func__, tid, len);
1754 
1755 		if (changed) {
1756 			if (toep->ddp.flags & DDP_SC_REQ)
1757 				toep->ddp.flags ^= DDP_ON | DDP_SC_REQ;
1758 			else {
1759 				KASSERT(cpl->ddp_off == 1,
1760 				    ("%s: DDP switched on by itself.",
1761 				    __func__));
1762 
1763 				/* Fell out of DDP mode */
1764 				toep->ddp.flags &= ~DDP_ON;
1765 				CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
1766 				    __func__);
1767 
1768 				insert_ddp_data(toep, ddp_placed);
1769 			}
1770 		}
1771 
1772 		if (toep->ddp.flags & DDP_ON) {
1773 			/*
1774 			 * CPL_RX_DATA with DDP on can only be an indicate.
1775 			 * Start posting queued AIO requests via DDP.  The
1776 			 * payload that arrived in this indicate is appended
1777 			 * to the socket buffer as usual.
1778 			 */
1779 			handle_ddp_indicate(toep);
1780 		}
1781 	}
1782 
1783 	sbappendstream_locked(sb, m, 0);
1784 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
1785 	if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) {
1786 		rx_credits = send_rx_credits(sc, toep, rx_credits);
1787 		tp->rcv_wnd += rx_credits;
1788 		tp->rcv_adv += rx_credits;
1789 	}
1790 
1791 	if (ulp_mode(toep) == ULP_MODE_TCPDDP && toep->ddp.waiting_count > 0 &&
1792 	    sbavail(sb) != 0) {
1793 		CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__,
1794 		    tid);
1795 		ddp_queue_toep(toep);
1796 	}
1797 	if (toep->flags & TPF_TLS_STARTING)
1798 		tls_received_starting_data(sc, toep, sb, len);
1799 	sorwakeup_locked(so);
1800 	SOCKBUF_UNLOCK_ASSERT(sb);
1801 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1802 		DDP_UNLOCK(toep);
1803 
1804 	INP_WUNLOCK(inp);
1805 	CURVNET_RESTORE();
1806 	return (0);
1807 }
1808 
1809 static int
1810 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1811 {
1812 	struct adapter *sc = iq->adapter;
1813 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1814 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1815 	struct toepcb *toep = lookup_tid(sc, tid);
1816 	struct inpcb *inp;
1817 	struct tcpcb *tp;
1818 	struct socket *so;
1819 	uint8_t credits = cpl->credits;
1820 	struct ofld_tx_sdesc *txsd;
1821 	int plen;
1822 #ifdef INVARIANTS
1823 	unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1824 #endif
1825 
1826 	/*
1827 	 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1828 	 * now this comes back carrying the credits for the flowc.
1829 	 */
1830 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1831 		KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1832 		    ("%s: credits for a synq entry %p", __func__, toep));
1833 		return (0);
1834 	}
1835 
1836 	inp = toep->inp;
1837 
1838 	KASSERT(opcode == CPL_FW4_ACK,
1839 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1840 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1841 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1842 
1843 	INP_WLOCK(inp);
1844 
1845 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1846 		INP_WUNLOCK(inp);
1847 		return (0);
1848 	}
1849 
1850 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1851 	    ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1852 
1853 	tp = intotcpcb(inp);
1854 
1855 	if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1856 		tcp_seq snd_una = be32toh(cpl->snd_una);
1857 
1858 #ifdef INVARIANTS
1859 		if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1860 			log(LOG_ERR,
1861 			    "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1862 			    __func__, snd_una, toep->tid, tp->snd_una);
1863 		}
1864 #endif
1865 
1866 		if (tp->snd_una != snd_una) {
1867 			tp->snd_una = snd_una;
1868 			tp->ts_recent_age = tcp_ts_getticks();
1869 		}
1870 	}
1871 
1872 #ifdef VERBOSE_TRACES
1873 	CTR3(KTR_CXGBE, "%s: tid %d credits %u", __func__, tid, credits);
1874 #endif
1875 	so = inp->inp_socket;
1876 	txsd = &toep->txsd[toep->txsd_cidx];
1877 	plen = 0;
1878 	while (credits) {
1879 		KASSERT(credits >= txsd->tx_credits,
1880 		    ("%s: too many (or partial) credits", __func__));
1881 		credits -= txsd->tx_credits;
1882 		toep->tx_credits += txsd->tx_credits;
1883 		plen += txsd->plen;
1884 		txsd++;
1885 		toep->txsd_avail++;
1886 		KASSERT(toep->txsd_avail <= toep->txsd_total,
1887 		    ("%s: txsd avail > total", __func__));
1888 		if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1889 			txsd = &toep->txsd[0];
1890 			toep->txsd_cidx = 0;
1891 		}
1892 	}
1893 
1894 	if (toep->tx_credits == toep->tx_total) {
1895 		toep->tx_nocompl = 0;
1896 		toep->plen_nocompl = 0;
1897 	}
1898 
1899 	if (toep->flags & TPF_TX_SUSPENDED &&
1900 	    toep->tx_credits >= toep->tx_total / 4) {
1901 #ifdef VERBOSE_TRACES
1902 		CTR2(KTR_CXGBE, "%s: tid %d calling t4_push_frames", __func__,
1903 		    tid);
1904 #endif
1905 		toep->flags &= ~TPF_TX_SUSPENDED;
1906 		CURVNET_SET(toep->vnet);
1907 		t4_push_data(sc, toep, plen);
1908 		CURVNET_RESTORE();
1909 	} else if (plen > 0) {
1910 		struct sockbuf *sb = &so->so_snd;
1911 		int sbu;
1912 
1913 		SOCKBUF_LOCK(sb);
1914 		sbu = sbused(sb);
1915 		if (ulp_mode(toep) == ULP_MODE_ISCSI) {
1916 			if (__predict_false(sbu > 0)) {
1917 				/*
1918 				 * The data transmitted before the
1919 				 * tid's ULP mode changed to ISCSI is
1920 				 * still in so_snd.  Incoming credits
1921 				 * should account for so_snd first.
1922 				 */
1923 				sbdrop_locked(sb, min(sbu, plen));
1924 				plen -= min(sbu, plen);
1925 			}
1926 			sowwakeup_locked(so);	/* unlocks so_snd */
1927 			rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1928 		} else {
1929 #ifdef VERBOSE_TRACES
1930 			CTR3(KTR_CXGBE, "%s: tid %d dropped %d bytes", __func__,
1931 			    tid, plen);
1932 #endif
1933 			sbdrop_locked(sb, plen);
1934 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
1935 				t4_aiotx_queue_toep(so, toep);
1936 			sowwakeup_locked(so);	/* unlocks so_snd */
1937 		}
1938 		SOCKBUF_UNLOCK_ASSERT(sb);
1939 	}
1940 
1941 	INP_WUNLOCK(inp);
1942 
1943 	return (0);
1944 }
1945 
1946 void
1947 t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, struct toepcb *toep,
1948     uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie)
1949 {
1950 	struct wrqe *wr;
1951 	struct cpl_set_tcb_field *req;
1952 	struct ofld_tx_sdesc *txsd;
1953 
1954 	MPASS((cookie & ~M_COOKIE) == 0);
1955 	if (reply) {
1956 		MPASS(cookie != CPL_COOKIE_RESERVED);
1957 	}
1958 
1959 	wr = alloc_wrqe(sizeof(*req), wrq);
1960 	if (wr == NULL) {
1961 		/* XXX */
1962 		panic("%s: allocation failure.", __func__);
1963 	}
1964 	req = wrtod(wr);
1965 
1966 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1967 	req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1968 	if (reply == 0)
1969 		req->reply_ctrl |= htobe16(F_NO_REPLY);
1970 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie));
1971 	req->mask = htobe64(mask);
1972 	req->val = htobe64(val);
1973 	if (wrq->eq.type == EQ_OFLD) {
1974 		txsd = &toep->txsd[toep->txsd_pidx];
1975 		txsd->tx_credits = howmany(sizeof(*req), 16);
1976 		txsd->plen = 0;
1977 		KASSERT(toep->tx_credits >= txsd->tx_credits &&
1978 		    toep->txsd_avail > 0,
1979 		    ("%s: not enough credits (%d)", __func__,
1980 		    toep->tx_credits));
1981 		toep->tx_credits -= txsd->tx_credits;
1982 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1983 			toep->txsd_pidx = 0;
1984 		toep->txsd_avail--;
1985 	}
1986 
1987 	t4_wrq_tx(sc, wr);
1988 }
1989 
1990 void
1991 t4_init_cpl_io_handlers(void)
1992 {
1993 
1994 	t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close);
1995 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl);
1996 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req);
1997 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl,
1998 	    CPL_COOKIE_TOM);
1999 	t4_register_cpl_handler(CPL_RX_DATA, do_rx_data);
2000 	t4_register_shared_cpl_handler(CPL_FW4_ACK, do_fw4_ack, CPL_COOKIE_TOM);
2001 }
2002 
2003 void
2004 t4_uninit_cpl_io_handlers(void)
2005 {
2006 
2007 	t4_register_cpl_handler(CPL_PEER_CLOSE, NULL);
2008 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, NULL);
2009 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, NULL);
2010 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, NULL, CPL_COOKIE_TOM);
2011 	t4_register_cpl_handler(CPL_RX_DATA, NULL);
2012 	t4_register_shared_cpl_handler(CPL_FW4_ACK, NULL, CPL_COOKIE_TOM);
2013 }
2014 
2015 /*
2016  * Use the 'backend1' field in AIO jobs to hold an error that should
2017  * be reported when the job is completed, the 'backend3' field to
2018  * store the amount of data sent by the AIO job so far, and the
2019  * 'backend4' field to hold a reference count on the job.
2020  *
2021  * Each unmapped mbuf holds a reference on the job as does the queue
2022  * so long as the job is queued.
2023  */
2024 #define	aio_error	backend1
2025 #define	aio_sent	backend3
2026 #define	aio_refs	backend4
2027 
2028 #ifdef VERBOSE_TRACES
2029 static int
2030 jobtotid(struct kaiocb *job)
2031 {
2032 	struct socket *so;
2033 	struct tcpcb *tp;
2034 	struct toepcb *toep;
2035 
2036 	so = job->fd_file->f_data;
2037 	tp = sototcpcb(so);
2038 	toep = tp->t_toe;
2039 	return (toep->tid);
2040 }
2041 #endif
2042 
2043 static void
2044 aiotx_free_job(struct kaiocb *job)
2045 {
2046 	long status;
2047 	int error;
2048 
2049 	if (refcount_release(&job->aio_refs) == 0)
2050 		return;
2051 
2052 	error = (intptr_t)job->aio_error;
2053 	status = job->aio_sent;
2054 #ifdef VERBOSE_TRACES
2055 	CTR5(KTR_CXGBE, "%s: tid %d completed %p len %ld, error %d", __func__,
2056 	    jobtotid(job), job, status, error);
2057 #endif
2058 	if (error != 0 && status != 0)
2059 		error = 0;
2060 	if (error == ECANCELED)
2061 		aio_cancel(job);
2062 	else if (error)
2063 		aio_complete(job, -1, error);
2064 	else {
2065 		job->msgsnd = 1;
2066 		aio_complete(job, status, 0);
2067 	}
2068 }
2069 
2070 static void
2071 aiotx_free_pgs(struct mbuf *m)
2072 {
2073 	struct kaiocb *job;
2074 	vm_page_t pg;
2075 
2076 	M_ASSERTEXTPG(m);
2077 	job = m->m_ext.ext_arg1;
2078 #ifdef VERBOSE_TRACES
2079 	CTR3(KTR_CXGBE, "%s: completed %d bytes for tid %d", __func__,
2080 	    m->m_len, jobtotid(job));
2081 #endif
2082 
2083 	for (int i = 0; i < m->m_epg_npgs; i++) {
2084 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
2085 		vm_page_unwire(pg, PQ_ACTIVE);
2086 	}
2087 
2088 	aiotx_free_job(job);
2089 }
2090 
2091 /*
2092  * Allocate a chain of unmapped mbufs describing the next 'len' bytes
2093  * of an AIO job.
2094  */
2095 static struct mbuf *
2096 alloc_aiotx_mbuf(struct kaiocb *job, int len)
2097 {
2098 	struct vmspace *vm;
2099 	vm_page_t pgs[MBUF_PEXT_MAX_PGS];
2100 	struct mbuf *m, *top, *last;
2101 	vm_map_t map;
2102 	vm_offset_t start;
2103 	int i, mlen, npages, pgoff;
2104 
2105 	KASSERT(job->aio_sent + len <= job->uaiocb.aio_nbytes,
2106 	    ("%s(%p, %d): request to send beyond end of buffer", __func__,
2107 	    job, len));
2108 
2109 	/*
2110 	 * The AIO subsystem will cancel and drain all requests before
2111 	 * permitting a process to exit or exec, so p_vmspace should
2112 	 * be stable here.
2113 	 */
2114 	vm = job->userproc->p_vmspace;
2115 	map = &vm->vm_map;
2116 	start = (uintptr_t)job->uaiocb.aio_buf + job->aio_sent;
2117 	pgoff = start & PAGE_MASK;
2118 
2119 	top = NULL;
2120 	last = NULL;
2121 	while (len > 0) {
2122 		mlen = imin(len, MBUF_PEXT_MAX_PGS * PAGE_SIZE - pgoff);
2123 		KASSERT(mlen == len || ((start + mlen) & PAGE_MASK) == 0,
2124 		    ("%s: next start (%#jx + %#x) is not page aligned",
2125 		    __func__, (uintmax_t)start, mlen));
2126 
2127 		npages = vm_fault_quick_hold_pages(map, start, mlen,
2128 		    VM_PROT_WRITE, pgs, nitems(pgs));
2129 		if (npages < 0)
2130 			break;
2131 
2132 		m = mb_alloc_ext_pgs(M_WAITOK, aiotx_free_pgs);
2133 		if (m == NULL) {
2134 			vm_page_unhold_pages(pgs, npages);
2135 			break;
2136 		}
2137 
2138 		m->m_epg_1st_off = pgoff;
2139 		m->m_epg_npgs = npages;
2140 		if (npages == 1) {
2141 			KASSERT(mlen + pgoff <= PAGE_SIZE,
2142 			    ("%s: single page is too large (off %d len %d)",
2143 			    __func__, pgoff, mlen));
2144 			m->m_epg_last_len = mlen;
2145 		} else {
2146 			m->m_epg_last_len = mlen - (PAGE_SIZE - pgoff) -
2147 			    (npages - 2) * PAGE_SIZE;
2148 		}
2149 		for (i = 0; i < npages; i++)
2150 			m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pgs[i]);
2151 
2152 		m->m_len = mlen;
2153 		m->m_ext.ext_size = npages * PAGE_SIZE;
2154 		m->m_ext.ext_arg1 = job;
2155 		refcount_acquire(&job->aio_refs);
2156 
2157 #ifdef VERBOSE_TRACES
2158 		CTR5(KTR_CXGBE, "%s: tid %d, new mbuf %p for job %p, npages %d",
2159 		    __func__, jobtotid(job), m, job, npages);
2160 #endif
2161 
2162 		if (top == NULL)
2163 			top = m;
2164 		else
2165 			last->m_next = m;
2166 		last = m;
2167 
2168 		len -= mlen;
2169 		start += mlen;
2170 		pgoff = 0;
2171 	}
2172 
2173 	return (top);
2174 }
2175 
2176 static void
2177 t4_aiotx_process_job(struct toepcb *toep, struct socket *so, struct kaiocb *job)
2178 {
2179 	struct sockbuf *sb;
2180 	struct inpcb *inp;
2181 	struct tcpcb *tp;
2182 	struct mbuf *m;
2183 	int error, len;
2184 	bool moretocome, sendmore;
2185 
2186 	sb = &so->so_snd;
2187 	SOCKBUF_UNLOCK(sb);
2188 	m = NULL;
2189 
2190 #ifdef MAC
2191 	error = mac_socket_check_send(job->fd_file->f_cred, so);
2192 	if (error != 0)
2193 		goto out;
2194 #endif
2195 
2196 	/* Inline sosend_generic(). */
2197 
2198 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
2199 	MPASS(error == 0);
2200 
2201 sendanother:
2202 	SOCKBUF_LOCK(sb);
2203 	if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
2204 		SOCKBUF_UNLOCK(sb);
2205 		SOCK_IO_SEND_UNLOCK(so);
2206 		if ((so->so_options & SO_NOSIGPIPE) == 0) {
2207 			PROC_LOCK(job->userproc);
2208 			kern_psignal(job->userproc, SIGPIPE);
2209 			PROC_UNLOCK(job->userproc);
2210 		}
2211 		error = EPIPE;
2212 		goto out;
2213 	}
2214 	if (so->so_error) {
2215 		error = so->so_error;
2216 		so->so_error = 0;
2217 		SOCKBUF_UNLOCK(sb);
2218 		SOCK_IO_SEND_UNLOCK(so);
2219 		goto out;
2220 	}
2221 	if ((so->so_state & SS_ISCONNECTED) == 0) {
2222 		SOCKBUF_UNLOCK(sb);
2223 		SOCK_IO_SEND_UNLOCK(so);
2224 		error = ENOTCONN;
2225 		goto out;
2226 	}
2227 	if (sbspace(sb) < sb->sb_lowat) {
2228 		MPASS(job->aio_sent == 0 || !(so->so_state & SS_NBIO));
2229 
2230 		/*
2231 		 * Don't block if there is too little room in the socket
2232 		 * buffer.  Instead, requeue the request.
2233 		 */
2234 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2235 			SOCKBUF_UNLOCK(sb);
2236 			SOCK_IO_SEND_UNLOCK(so);
2237 			error = ECANCELED;
2238 			goto out;
2239 		}
2240 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2241 		SOCKBUF_UNLOCK(sb);
2242 		SOCK_IO_SEND_UNLOCK(so);
2243 		goto out;
2244 	}
2245 
2246 	/*
2247 	 * Write as much data as the socket permits, but no more than a
2248 	 * a single sndbuf at a time.
2249 	 */
2250 	len = sbspace(sb);
2251 	if (len > job->uaiocb.aio_nbytes - job->aio_sent) {
2252 		len = job->uaiocb.aio_nbytes - job->aio_sent;
2253 		moretocome = false;
2254 	} else
2255 		moretocome = true;
2256 	if (len > toep->params.sndbuf) {
2257 		len = toep->params.sndbuf;
2258 		sendmore = true;
2259 	} else
2260 		sendmore = false;
2261 
2262 	if (!TAILQ_EMPTY(&toep->aiotx_jobq))
2263 		moretocome = true;
2264 	SOCKBUF_UNLOCK(sb);
2265 	MPASS(len != 0);
2266 
2267 	m = alloc_aiotx_mbuf(job, len);
2268 	if (m == NULL) {
2269 		SOCK_IO_SEND_UNLOCK(so);
2270 		error = EFAULT;
2271 		goto out;
2272 	}
2273 
2274 	/* Inlined tcp_usr_send(). */
2275 
2276 	inp = toep->inp;
2277 	INP_WLOCK(inp);
2278 	if (inp->inp_flags & INP_DROPPED) {
2279 		INP_WUNLOCK(inp);
2280 		SOCK_IO_SEND_UNLOCK(so);
2281 		error = ECONNRESET;
2282 		goto out;
2283 	}
2284 
2285 	job->aio_sent += m_length(m, NULL);
2286 
2287 	sbappendstream(sb, m, 0);
2288 	m = NULL;
2289 
2290 	if (!(inp->inp_flags & INP_DROPPED)) {
2291 		tp = intotcpcb(inp);
2292 		if (moretocome)
2293 			tp->t_flags |= TF_MORETOCOME;
2294 		error = tcp_output(tp);
2295 		if (error < 0) {
2296 			INP_UNLOCK_ASSERT(inp);
2297 			SOCK_IO_SEND_UNLOCK(so);
2298 			error = -error;
2299 			goto out;
2300 		}
2301 		if (moretocome)
2302 			tp->t_flags &= ~TF_MORETOCOME;
2303 	}
2304 
2305 	INP_WUNLOCK(inp);
2306 	if (sendmore)
2307 		goto sendanother;
2308 	SOCK_IO_SEND_UNLOCK(so);
2309 
2310 	if (error)
2311 		goto out;
2312 
2313 	/*
2314 	 * If this is a blocking socket and the request has not been
2315 	 * fully completed, requeue it until the socket is ready
2316 	 * again.
2317 	 */
2318 	if (job->aio_sent < job->uaiocb.aio_nbytes &&
2319 	    !(so->so_state & SS_NBIO)) {
2320 		SOCKBUF_LOCK(sb);
2321 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2322 			SOCKBUF_UNLOCK(sb);
2323 			error = ECANCELED;
2324 			goto out;
2325 		}
2326 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2327 		return;
2328 	}
2329 
2330 	/*
2331 	 * If the request will not be requeued, drop the queue's
2332 	 * reference to the job.  Any mbufs in flight should still
2333 	 * hold a reference, but this drops the reference that the
2334 	 * queue owns while it is waiting to queue mbufs to the
2335 	 * socket.
2336 	 */
2337 	aiotx_free_job(job);
2338 
2339 out:
2340 	if (error) {
2341 		job->aio_error = (void *)(intptr_t)error;
2342 		aiotx_free_job(job);
2343 	}
2344 	m_freem(m);
2345 	SOCKBUF_LOCK(sb);
2346 }
2347 
2348 static void
2349 t4_aiotx_task(void *context, int pending)
2350 {
2351 	struct toepcb *toep = context;
2352 	struct socket *so;
2353 	struct kaiocb *job;
2354 	struct epoch_tracker et;
2355 
2356 	so = toep->aiotx_so;
2357 	CURVNET_SET(toep->vnet);
2358 	NET_EPOCH_ENTER(et);
2359 	SOCKBUF_LOCK(&so->so_snd);
2360 	while (!TAILQ_EMPTY(&toep->aiotx_jobq) && sowriteable(so)) {
2361 		job = TAILQ_FIRST(&toep->aiotx_jobq);
2362 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2363 		if (!aio_clear_cancel_function(job))
2364 			continue;
2365 
2366 		t4_aiotx_process_job(toep, so, job);
2367 	}
2368 	toep->aiotx_so = NULL;
2369 	SOCKBUF_UNLOCK(&so->so_snd);
2370 	NET_EPOCH_EXIT(et);
2371 
2372 	free_toepcb(toep);
2373 	sorele(so);
2374 	CURVNET_RESTORE();
2375 }
2376 
2377 static void
2378 t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep)
2379 {
2380 
2381 	SOCKBUF_LOCK_ASSERT(&toep->inp->inp_socket->so_snd);
2382 #ifdef VERBOSE_TRACES
2383 	CTR3(KTR_CXGBE, "%s: queueing aiotx task for tid %d, active = %s",
2384 	    __func__, toep->tid, toep->aiotx_so != NULL ? "true" : "false");
2385 #endif
2386 	if (toep->aiotx_so != NULL)
2387 		return;
2388 	soref(so);
2389 	toep->aiotx_so = so;
2390 	hold_toepcb(toep);
2391 	soaio_enqueue(&toep->aiotx_task);
2392 }
2393 
2394 static void
2395 t4_aiotx_cancel(struct kaiocb *job)
2396 {
2397 	struct socket *so;
2398 	struct sockbuf *sb;
2399 	struct tcpcb *tp;
2400 	struct toepcb *toep;
2401 
2402 	so = job->fd_file->f_data;
2403 	tp = sototcpcb(so);
2404 	toep = tp->t_toe;
2405 	MPASS(job->uaiocb.aio_lio_opcode == LIO_WRITE);
2406 	sb = &so->so_snd;
2407 
2408 	SOCKBUF_LOCK(sb);
2409 	if (!aio_cancel_cleared(job))
2410 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2411 	SOCKBUF_UNLOCK(sb);
2412 
2413 	job->aio_error = (void *)(intptr_t)ECANCELED;
2414 	aiotx_free_job(job);
2415 }
2416 
2417 int
2418 t4_aio_queue_aiotx(struct socket *so, struct kaiocb *job)
2419 {
2420 	struct tcpcb *tp = sototcpcb(so);
2421 	struct toepcb *toep = tp->t_toe;
2422 	struct adapter *sc = td_adapter(toep->td);
2423 
2424 	/* This only handles writes. */
2425 	if (job->uaiocb.aio_lio_opcode != LIO_WRITE)
2426 		return (EOPNOTSUPP);
2427 
2428 	if (!sc->tt.tx_zcopy)
2429 		return (EOPNOTSUPP);
2430 
2431 	if (tls_tx_key(toep))
2432 		return (EOPNOTSUPP);
2433 
2434 	SOCKBUF_LOCK(&so->so_snd);
2435 #ifdef VERBOSE_TRACES
2436 	CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid);
2437 #endif
2438 	if (!aio_set_cancel_function(job, t4_aiotx_cancel))
2439 		panic("new job was cancelled");
2440 	refcount_init(&job->aio_refs, 1);
2441 	TAILQ_INSERT_TAIL(&toep->aiotx_jobq, job, list);
2442 	if (sowriteable(so))
2443 		t4_aiotx_queue_toep(so, toep);
2444 	SOCKBUF_UNLOCK(&so->so_snd);
2445 	return (0);
2446 }
2447 
2448 void
2449 aiotx_init_toep(struct toepcb *toep)
2450 {
2451 
2452 	TAILQ_INIT(&toep->aiotx_jobq);
2453 	TASK_INIT(&toep->aiotx_task, 0, t4_aiotx_task, toep);
2454 }
2455 #endif
2456