xref: /freebsd/sys/dev/cxgbe/tom/t4_cpl_io.c (revision 314cb279959b08811543612a715e47266f685c7b)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
35 
36 #ifdef TCP_OFFLOAD
37 #include <sys/param.h>
38 #include <sys/aio.h>
39 #include <sys/file.h>
40 #include <sys/kernel.h>
41 #include <sys/ktr.h>
42 #include <sys/module.h>
43 #include <sys/proc.h>
44 #include <sys/protosw.h>
45 #include <sys/domain.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sglist.h>
49 #include <sys/taskqueue.h>
50 #include <netinet/in.h>
51 #include <netinet/in_pcb.h>
52 #include <netinet/ip.h>
53 #include <netinet/ip6.h>
54 #define TCPSTATES
55 #include <netinet/tcp_fsm.h>
56 #include <netinet/tcp_seq.h>
57 #include <netinet/tcp_var.h>
58 #include <netinet/toecore.h>
59 
60 #include <security/mac/mac_framework.h>
61 
62 #include <vm/vm.h>
63 #include <vm/vm_extern.h>
64 #include <vm/pmap.h>
65 #include <vm/vm_map.h>
66 #include <vm/vm_page.h>
67 
68 #include <dev/iscsi/iscsi_proto.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
send_flowc_wr(struct toepcb * toep,struct tcpcb * tp)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 (toep->params.tc_idx != -1) {
100 		MPASS(toep->params.tc_idx >= 0 &&
101 		    toep->params.tc_idx < sc->params.nsched_cls);
102 		nparams++;
103 	}
104 
105 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
106 
107 	wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq);
108 	if (wr == NULL) {
109 		/* XXX */
110 		panic("%s: allocation failure.", __func__);
111 	}
112 	flowc = wrtod(wr);
113 	memset(flowc, 0, wr->wr_len);
114 
115 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
116 	    V_FW_FLOWC_WR_NPARAMS(nparams));
117 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
118 	    V_FW_WR_FLOWID(toep->tid));
119 
120 #define FLOWC_PARAM(__m, __v) \
121 	do { \
122 		flowc->mnemval[paramidx].mnemonic = FW_FLOWC_MNEM_##__m; \
123 		flowc->mnemval[paramidx].val = htobe32(__v); \
124 		paramidx++; \
125 	} while (0)
126 
127 	paramidx = 0;
128 
129 	FLOWC_PARAM(PFNVFN, pfvf);
130 	FLOWC_PARAM(CH, pi->tx_chan);
131 	FLOWC_PARAM(PORT, pi->tx_chan);
132 	FLOWC_PARAM(IQID, toep->ofld_rxq->iq.abs_id);
133 	FLOWC_PARAM(SNDBUF, toep->params.sndbuf);
134 	if (tp) {
135 		FLOWC_PARAM(MSS, toep->params.emss);
136 		FLOWC_PARAM(SNDNXT, tp->snd_nxt);
137 		FLOWC_PARAM(RCVNXT, tp->rcv_nxt);
138 	} else
139 		FLOWC_PARAM(MSS, 512);
140 	CTR6(KTR_CXGBE,
141 	    "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x",
142 	    __func__, toep->tid, toep->params.emss, toep->params.sndbuf,
143 	    tp ? tp->snd_nxt : 0, tp ? tp->rcv_nxt : 0);
144 
145 	if (toep->params.tc_idx != -1)
146 		FLOWC_PARAM(SCHEDCLASS, toep->params.tc_idx);
147 #undef FLOWC_PARAM
148 
149 	KASSERT(paramidx == nparams, ("nparams mismatch"));
150 
151 	txsd->tx_credits = howmany(flowclen, 16);
152 	txsd->plen = 0;
153 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
154 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
155 	toep->tx_credits -= txsd->tx_credits;
156 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
157 		toep->txsd_pidx = 0;
158 	toep->txsd_avail--;
159 
160 	toep->flags |= TPF_FLOWC_WR_SENT;
161         t4_wrq_tx(sc, wr);
162 }
163 
164 #ifdef RATELIMIT
165 /*
166  * Input is Bytes/second (so_max_pacing_rate), chip counts in Kilobits/second.
167  */
168 static int
update_tx_rate_limit(struct adapter * sc,struct toepcb * toep,u_int Bps)169 update_tx_rate_limit(struct adapter *sc, struct toepcb *toep, u_int Bps)
170 {
171 	int tc_idx, rc;
172 	const u_int kbps = (u_int) (uint64_t)Bps * 8ULL / 1000;
173 	const int port_id = toep->vi->pi->port_id;
174 
175 	CTR3(KTR_CXGBE, "%s: tid %u, rate %uKbps", __func__, toep->tid, kbps);
176 
177 	if (kbps == 0) {
178 		/* unbind */
179 		tc_idx = -1;
180 	} else {
181 		rc = t4_reserve_cl_rl_kbps(sc, port_id, kbps, &tc_idx);
182 		if (rc != 0)
183 			return (rc);
184 		MPASS(tc_idx >= 0 && tc_idx < sc->params.nsched_cls);
185 	}
186 
187 	if (toep->params.tc_idx != tc_idx) {
188 		struct wrqe *wr;
189 		struct fw_flowc_wr *flowc;
190 		int nparams = 1, flowclen, flowclen16;
191 		struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
192 
193 		flowclen = sizeof(*flowc) + nparams * sizeof(struct
194 		    fw_flowc_mnemval);
195 		flowclen16 = howmany(flowclen, 16);
196 		if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0 ||
197 		    (wr = alloc_wrqe(roundup2(flowclen, 16),
198 		    &toep->ofld_txq->wrq)) == NULL) {
199 			if (tc_idx >= 0)
200 				t4_release_cl_rl(sc, port_id, tc_idx);
201 			return (ENOMEM);
202 		}
203 
204 		flowc = wrtod(wr);
205 		memset(flowc, 0, wr->wr_len);
206 
207 		flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
208 		    V_FW_FLOWC_WR_NPARAMS(nparams));
209 		flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
210 		    V_FW_WR_FLOWID(toep->tid));
211 
212 		flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS;
213 		if (tc_idx == -1)
214 			flowc->mnemval[0].val = htobe32(0xff);
215 		else
216 			flowc->mnemval[0].val = htobe32(tc_idx);
217 
218 		txsd->tx_credits = flowclen16;
219 		txsd->plen = 0;
220 		toep->tx_credits -= txsd->tx_credits;
221 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
222 			toep->txsd_pidx = 0;
223 		toep->txsd_avail--;
224 		t4_wrq_tx(sc, wr);
225 	}
226 
227 	if (toep->params.tc_idx >= 0)
228 		t4_release_cl_rl(sc, port_id, toep->params.tc_idx);
229 	toep->params.tc_idx = tc_idx;
230 
231 	return (0);
232 }
233 #endif
234 
235 void
send_reset(struct adapter * sc,struct toepcb * toep,uint32_t snd_nxt)236 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
237 {
238 	struct wrqe *wr;
239 	struct cpl_abort_req *req;
240 	int tid = toep->tid;
241 	struct inpcb *inp = toep->inp;
242 	struct tcpcb *tp = intotcpcb(inp);	/* don't use if INP_DROPPED */
243 
244 	INP_WLOCK_ASSERT(inp);
245 
246 	CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
247 	    __func__, toep->tid,
248 	    inp->inp_flags & INP_DROPPED ? "inp dropped" :
249 	    tcpstates[tp->t_state],
250 	    toep->flags, inp->inp_flags,
251 	    toep->flags & TPF_ABORT_SHUTDOWN ?
252 	    " (abort already in progress)" : "");
253 
254 	if (toep->flags & TPF_ABORT_SHUTDOWN)
255 		return;	/* abort already in progress */
256 
257 	toep->flags |= TPF_ABORT_SHUTDOWN;
258 
259 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
260 	    ("%s: flowc_wr not sent for tid %d.", __func__, tid));
261 
262 	wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq);
263 	if (wr == NULL) {
264 		/* XXX */
265 		panic("%s: allocation failure.", __func__);
266 	}
267 	req = wrtod(wr);
268 
269 	INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
270 	if (inp->inp_flags & INP_DROPPED)
271 		req->rsvd0 = htobe32(snd_nxt);
272 	else
273 		req->rsvd0 = htobe32(tp->snd_nxt);
274 	req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
275 	req->cmd = CPL_ABORT_SEND_RST;
276 
277 	/*
278 	 * XXX: What's the correct way to tell that the inp hasn't been detached
279 	 * from its socket?  Should I even be flushing the snd buffer here?
280 	 */
281 	if ((inp->inp_flags & INP_DROPPED) == 0) {
282 		struct socket *so = inp->inp_socket;
283 
284 		if (so != NULL)	/* because I'm not sure.  See comment above */
285 			sbflush(&so->so_snd);
286 	}
287 
288 	t4_l2t_send(sc, wr, toep->l2te);
289 }
290 
291 /*
292  * Called when a connection is established to translate the TCP options
293  * reported by HW to FreeBSD's native format.
294  */
295 static void
assign_rxopt(struct tcpcb * tp,uint16_t opt)296 assign_rxopt(struct tcpcb *tp, uint16_t opt)
297 {
298 	struct toepcb *toep = tp->t_toe;
299 	struct inpcb *inp = tptoinpcb(tp);
300 	struct adapter *sc = td_adapter(toep->td);
301 
302 	INP_LOCK_ASSERT(inp);
303 
304 	toep->params.mtu_idx = G_TCPOPT_MSS(opt);
305 	tp->t_maxseg = sc->params.mtus[toep->params.mtu_idx];
306 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
307 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
308 	else
309 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
310 
311 	toep->params.emss = tp->t_maxseg;
312 	if (G_TCPOPT_TSTAMP(opt)) {
313 		toep->params.tstamp = 1;
314 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
315 		tp->t_flags |= TF_RCVD_TSTMP;	/* timestamps ok */
316 		tp->ts_recent = 0;		/* hmmm */
317 		tp->ts_recent_age = tcp_ts_getticks();
318 	} else
319 		toep->params.tstamp = 0;
320 
321 	if (G_TCPOPT_SACK(opt)) {
322 		toep->params.sack = 1;
323 		tp->t_flags |= TF_SACK_PERMIT;	/* should already be set */
324 	} else {
325 		toep->params.sack = 0;
326 		tp->t_flags &= ~TF_SACK_PERMIT;	/* sack disallowed by peer */
327 	}
328 
329 	if (G_TCPOPT_WSCALE_OK(opt))
330 		tp->t_flags |= TF_RCVD_SCALE;
331 
332 	/* Doing window scaling? */
333 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
334 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
335 		tp->rcv_scale = tp->request_r_scale;
336 		tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
337 	} else
338 		toep->params.wscale = 0;
339 
340 	CTR6(KTR_CXGBE,
341 	    "assign_rxopt: tid %d, mtu_idx %u, emss %u, ts %u, sack %u, wscale %u",
342 	    toep->tid, toep->params.mtu_idx, toep->params.emss,
343 	    toep->params.tstamp, toep->params.sack, toep->params.wscale);
344 }
345 
346 /*
347  * Completes some final bits of initialization for just established connections
348  * and changes their state to TCPS_ESTABLISHED.
349  *
350  * The ISNs are from the exchange of SYNs.
351  */
352 void
make_established(struct toepcb * toep,uint32_t iss,uint32_t irs,uint16_t opt)353 make_established(struct toepcb *toep, uint32_t iss, uint32_t irs, uint16_t opt)
354 {
355 	struct inpcb *inp = toep->inp;
356 	struct socket *so = inp->inp_socket;
357 	struct tcpcb *tp = intotcpcb(inp);
358 	uint16_t tcpopt = be16toh(opt);
359 
360 	INP_WLOCK_ASSERT(inp);
361 	KASSERT(tp->t_state == TCPS_SYN_SENT ||
362 	    tp->t_state == TCPS_SYN_RECEIVED,
363 	    ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
364 
365 	CTR6(KTR_CXGBE, "%s: tid %d, so %p, inp %p, tp %p, toep %p",
366 	    __func__, toep->tid, so, inp, tp, toep);
367 
368 	tcp_state_change(tp, TCPS_ESTABLISHED);
369 	tp->t_starttime = ticks;
370 	TCPSTAT_INC(tcps_connects);
371 
372 	tp->irs = irs;
373 	tcp_rcvseqinit(tp);
374 	tp->rcv_wnd = (u_int)toep->params.opt0_bufsize << 10;
375 	tp->rcv_adv += tp->rcv_wnd;
376 	tp->last_ack_sent = tp->rcv_nxt;
377 
378 	tp->iss = iss;
379 	tcp_sendseqinit(tp);
380 	tp->snd_una = iss + 1;
381 	tp->snd_nxt = iss + 1;
382 	tp->snd_max = iss + 1;
383 
384 	assign_rxopt(tp, tcpopt);
385 	send_flowc_wr(toep, tp);
386 
387 	soisconnected(so);
388 }
389 
390 int
send_rx_credits(struct adapter * sc,struct toepcb * toep,int credits)391 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
392 {
393 	struct wrqe *wr;
394 	struct cpl_rx_data_ack *req;
395 	uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
396 
397 	KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
398 
399 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
400 	if (wr == NULL)
401 		return (0);
402 	req = wrtod(wr);
403 
404 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
405 	req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
406 
407 	t4_wrq_tx(sc, wr);
408 	return (credits);
409 }
410 
411 void
t4_rcvd_locked(struct toedev * tod,struct tcpcb * tp)412 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp)
413 {
414 	struct adapter *sc = tod->tod_softc;
415 	struct inpcb *inp = tptoinpcb(tp);
416 	struct socket *so = inp->inp_socket;
417 	struct sockbuf *sb = &so->so_rcv;
418 	struct toepcb *toep = tp->t_toe;
419 	int rx_credits;
420 
421 	INP_WLOCK_ASSERT(inp);
422 	SOCKBUF_LOCK_ASSERT(sb);
423 
424 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
425 	if (rx_credits > 0 &&
426 	    (tp->rcv_wnd <= 32 * 1024 || rx_credits >= 64 * 1024 ||
427 	    (rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
428 	    sbused(sb) + tp->rcv_wnd < sb->sb_lowat)) {
429 		rx_credits = send_rx_credits(sc, toep, rx_credits);
430 		tp->rcv_wnd += rx_credits;
431 		tp->rcv_adv += rx_credits;
432 	}
433 }
434 
435 void
t4_rcvd(struct toedev * tod,struct tcpcb * tp)436 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
437 {
438 	struct inpcb *inp = tptoinpcb(tp);
439 	struct socket *so = inp->inp_socket;
440 	struct sockbuf *sb = &so->so_rcv;
441 
442 	SOCKBUF_LOCK(sb);
443 	t4_rcvd_locked(tod, tp);
444 	SOCKBUF_UNLOCK(sb);
445 }
446 
447 /*
448  * Close a connection by sending a CPL_CLOSE_CON_REQ message.
449  */
450 int
t4_close_conn(struct adapter * sc,struct toepcb * toep)451 t4_close_conn(struct adapter *sc, struct toepcb *toep)
452 {
453 	struct wrqe *wr;
454 	struct cpl_close_con_req *req;
455 	unsigned int tid = toep->tid;
456 
457 	CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
458 	    toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
459 
460 	if (toep->flags & TPF_FIN_SENT)
461 		return (0);
462 
463 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
464 	    ("%s: flowc_wr not sent for tid %u.", __func__, tid));
465 
466 	wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq);
467 	if (wr == NULL) {
468 		/* XXX */
469 		panic("%s: allocation failure.", __func__);
470 	}
471 	req = wrtod(wr);
472 
473         req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
474 	    V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
475 	req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
476 	    V_FW_WR_FLOWID(tid));
477         req->wr.wr_lo = cpu_to_be64(0);
478         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
479 	req->rsvd = 0;
480 
481 	toep->flags |= TPF_FIN_SENT;
482 	toep->flags &= ~TPF_SEND_FIN;
483 	t4_l2t_send(sc, wr, toep->l2te);
484 
485 	return (0);
486 }
487 
488 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
489 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
490 #define MIN_ISO_TX_CREDITS  (howmany(sizeof(struct cpl_tx_data_iso), 16))
491 #define MIN_TX_CREDITS(iso)						\
492 	(MIN_OFLD_TX_CREDITS + ((iso) ? MIN_ISO_TX_CREDITS : 0))
493 
494 /* Maximum amount of immediate data we could stuff in a WR */
495 static inline int
max_imm_payload(int tx_credits,int iso)496 max_imm_payload(int tx_credits, int iso)
497 {
498 	const int iso_cpl_size = iso ? sizeof(struct cpl_tx_data_iso) : 0;
499 	const int n = 1;	/* Use no more than one desc for imm. data WR */
500 
501 	KASSERT(tx_credits >= 0 &&
502 		tx_credits <= MAX_OFLD_TX_CREDITS,
503 		("%s: %d credits", __func__, tx_credits));
504 
505 	if (tx_credits < MIN_TX_CREDITS(iso))
506 		return (0);
507 
508 	if (tx_credits >= (n * EQ_ESIZE) / 16)
509 		return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr) -
510 		    iso_cpl_size);
511 	else
512 		return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr) -
513 		    iso_cpl_size);
514 }
515 
516 /* Maximum number of SGL entries we could stuff in a WR */
517 static inline int
max_dsgl_nsegs(int tx_credits,int iso)518 max_dsgl_nsegs(int tx_credits, int iso)
519 {
520 	int nseg = 1;	/* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
521 	int sge_pair_credits = tx_credits - MIN_TX_CREDITS(iso);
522 
523 	KASSERT(tx_credits >= 0 &&
524 		tx_credits <= MAX_OFLD_TX_CREDITS,
525 		("%s: %d credits", __func__, tx_credits));
526 
527 	if (tx_credits < MIN_TX_CREDITS(iso))
528 		return (0);
529 
530 	nseg += 2 * (sge_pair_credits * 16 / 24);
531 	if ((sge_pair_credits * 16) % 24 == 16)
532 		nseg++;
533 
534 	return (nseg);
535 }
536 
537 static inline void
write_tx_wr(void * dst,struct toepcb * toep,int fw_wr_opcode,unsigned int immdlen,unsigned int plen,uint8_t credits,int shove,int ulp_submode)538 write_tx_wr(void *dst, struct toepcb *toep, int fw_wr_opcode,
539     unsigned int immdlen, unsigned int plen, uint8_t credits, int shove,
540     int ulp_submode)
541 {
542 	struct fw_ofld_tx_data_wr *txwr = dst;
543 
544 	txwr->op_to_immdlen = htobe32(V_WR_OP(fw_wr_opcode) |
545 	    V_FW_WR_IMMDLEN(immdlen));
546 	txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
547 	    V_FW_WR_LEN16(credits));
548 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ulp_mode(toep)) |
549 	    V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
550 	txwr->plen = htobe32(plen);
551 
552 	if (toep->params.tx_align > 0) {
553 		if (plen < 2 * toep->params.emss)
554 			txwr->lsodisable_to_flags |=
555 			    htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
556 		else
557 			txwr->lsodisable_to_flags |=
558 			    htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
559 				(toep->params.nagle == 0 ? 0 :
560 				F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
561 	}
562 }
563 
564 /*
565  * Generate a DSGL from a starting mbuf.  The total number of segments and the
566  * maximum segments in any one mbuf are provided.
567  */
568 static void
write_tx_sgl(void * dst,struct mbuf * start,struct mbuf * stop,int nsegs,int n)569 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
570 {
571 	struct mbuf *m;
572 	struct ulptx_sgl *usgl = dst;
573 	int i, j, rc;
574 	struct sglist sg;
575 	struct sglist_seg segs[n];
576 
577 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
578 
579 	sglist_init(&sg, n, segs);
580 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
581 	    V_ULPTX_NSGE(nsegs));
582 
583 	i = -1;
584 	for (m = start; m != stop; m = m->m_next) {
585 		if (m->m_flags & M_EXTPG)
586 			rc = sglist_append_mbuf_epg(&sg, m,
587 			    mtod(m, vm_offset_t), m->m_len);
588 		else
589 			rc = sglist_append(&sg, mtod(m, void *), m->m_len);
590 		if (__predict_false(rc != 0))
591 			panic("%s: sglist_append %d", __func__, rc);
592 
593 		for (j = 0; j < sg.sg_nseg; i++, j++) {
594 			if (i < 0) {
595 				usgl->len0 = htobe32(segs[j].ss_len);
596 				usgl->addr0 = htobe64(segs[j].ss_paddr);
597 			} else {
598 				usgl->sge[i / 2].len[i & 1] =
599 				    htobe32(segs[j].ss_len);
600 				usgl->sge[i / 2].addr[i & 1] =
601 				    htobe64(segs[j].ss_paddr);
602 			}
603 #ifdef INVARIANTS
604 			nsegs--;
605 #endif
606 		}
607 		sglist_reset(&sg);
608 	}
609 	if (i & 1)
610 		usgl->sge[i / 2].len[1] = htobe32(0);
611 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
612 	    __func__, nsegs, start, stop));
613 }
614 
615 /*
616  * Max number of SGL entries an offload tx work request can have.  This is 41
617  * (1 + 40) for a full 512B work request.
618  * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
619  */
620 #define OFLD_SGL_LEN (41)
621 
622 /*
623  * Send data and/or a FIN to the peer.
624  *
625  * The socket's so_snd buffer consists of a stream of data starting with sb_mb
626  * and linked together with m_next.  sb_sndptr, if set, is the last mbuf that
627  * was transmitted.
628  *
629  * drop indicates the number of bytes that should be dropped from the head of
630  * the send buffer.  It is an optimization that lets do_fw4_ack avoid creating
631  * contention on the send buffer lock (before this change it used to do
632  * sowwakeup and then t4_push_frames right after that when recovering from tx
633  * stalls).  When drop is set this function MUST drop the bytes and wake up any
634  * writers.
635  */
636 void
t4_push_frames(struct adapter * sc,struct toepcb * toep,int drop)637 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
638 {
639 	struct mbuf *sndptr, *m, *sb_sndptr;
640 	struct fw_ofld_tx_data_wr *txwr;
641 	struct wrqe *wr;
642 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
643 	struct inpcb *inp = toep->inp;
644 	struct tcpcb *tp = intotcpcb(inp);
645 	struct socket *so = inp->inp_socket;
646 	struct sockbuf *sb = &so->so_snd;
647 	int tx_credits, shove, compl, sowwakeup;
648 	struct ofld_tx_sdesc *txsd;
649 	bool nomap_mbuf_seen;
650 
651 	INP_WLOCK_ASSERT(inp);
652 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
653 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
654 
655 	KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
656 	    ulp_mode(toep) == ULP_MODE_TCPDDP ||
657 	    ulp_mode(toep) == ULP_MODE_TLS ||
658 	    ulp_mode(toep) == ULP_MODE_RDMA,
659 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
660 
661 #ifdef VERBOSE_TRACES
662 	CTR5(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
663 	    __func__, toep->tid, toep->flags, tp->t_flags, drop);
664 #endif
665 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
666 		return;
667 
668 #ifdef RATELIMIT
669 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
670 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
671 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
672 	}
673 #endif
674 
675 	/*
676 	 * This function doesn't resume by itself.  Someone else must clear the
677 	 * flag and call this function.
678 	 */
679 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
680 		KASSERT(drop == 0,
681 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
682 		return;
683 	}
684 
685 	txsd = &toep->txsd[toep->txsd_pidx];
686 	do {
687 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
688 		max_imm = max_imm_payload(tx_credits, 0);
689 		max_nsegs = max_dsgl_nsegs(tx_credits, 0);
690 
691 		SOCKBUF_LOCK(sb);
692 		sowwakeup = drop;
693 		if (drop) {
694 			sbdrop_locked(sb, drop);
695 			drop = 0;
696 		}
697 		sb_sndptr = sb->sb_sndptr;
698 		sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
699 		plen = 0;
700 		nsegs = 0;
701 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
702 		nomap_mbuf_seen = false;
703 		for (m = sndptr; m != NULL; m = m->m_next) {
704 			int n;
705 
706 			if ((m->m_flags & M_NOTAVAIL) != 0)
707 				break;
708 			if (m->m_flags & M_EXTPG) {
709 #ifdef KERN_TLS
710 				if (m->m_epg_tls != NULL) {
711 					toep->flags |= TPF_KTLS;
712 					if (plen == 0) {
713 						SOCKBUF_UNLOCK(sb);
714 						t4_push_ktls(sc, toep, 0);
715 						return;
716 					}
717 					break;
718 				}
719 #endif
720 				n = sglist_count_mbuf_epg(m,
721 				    mtod(m, vm_offset_t), m->m_len);
722 			} else
723 				n = sglist_count(mtod(m, void *), m->m_len);
724 
725 			nsegs += n;
726 			plen += m->m_len;
727 
728 			/* This mbuf sent us _over_ the nsegs limit, back out */
729 			if (plen > max_imm && nsegs > max_nsegs) {
730 				nsegs -= n;
731 				plen -= m->m_len;
732 				if (plen == 0) {
733 					/* Too few credits */
734 					toep->flags |= TPF_TX_SUSPENDED;
735 					if (sowwakeup) {
736 						if (!TAILQ_EMPTY(
737 						    &toep->aiotx_jobq))
738 							t4_aiotx_queue_toep(so,
739 							    toep);
740 						sowwakeup_locked(so);
741 					} else
742 						SOCKBUF_UNLOCK(sb);
743 					SOCKBUF_UNLOCK_ASSERT(sb);
744 					return;
745 				}
746 				break;
747 			}
748 
749 			if (m->m_flags & M_EXTPG)
750 				nomap_mbuf_seen = true;
751 			if (max_nsegs_1mbuf < n)
752 				max_nsegs_1mbuf = n;
753 			sb_sndptr = m;	/* new sb->sb_sndptr if all goes well */
754 
755 			/* This mbuf put us right at the max_nsegs limit */
756 			if (plen > max_imm && nsegs == max_nsegs) {
757 				m = m->m_next;
758 				break;
759 			}
760 		}
761 
762 		if (sbused(sb) > sb->sb_hiwat * 5 / 8 &&
763 		    toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
764 			compl = 1;
765 		else
766 			compl = 0;
767 
768 		if (sb->sb_flags & SB_AUTOSIZE &&
769 		    V_tcp_do_autosndbuf &&
770 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
771 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
772 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
773 			    V_tcp_autosndbuf_max);
774 
775 			if (!sbreserve_locked(so, SO_SND, newsize, NULL))
776 				sb->sb_flags &= ~SB_AUTOSIZE;
777 			else
778 				sowwakeup = 1;	/* room available */
779 		}
780 		if (sowwakeup) {
781 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
782 				t4_aiotx_queue_toep(so, toep);
783 			sowwakeup_locked(so);
784 		} else
785 			SOCKBUF_UNLOCK(sb);
786 		SOCKBUF_UNLOCK_ASSERT(sb);
787 
788 		/* nothing to send */
789 		if (plen == 0) {
790 			KASSERT(m == NULL || (m->m_flags & M_NOTAVAIL) != 0,
791 			    ("%s: nothing to send, but m != NULL is ready",
792 			    __func__));
793 			break;
794 		}
795 
796 		if (__predict_false(toep->flags & TPF_FIN_SENT))
797 			panic("%s: excess tx.", __func__);
798 
799 		shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
800 		if (plen <= max_imm && !nomap_mbuf_seen) {
801 
802 			/* Immediate data tx */
803 
804 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
805 					&toep->ofld_txq->wrq);
806 			if (wr == NULL) {
807 				/* XXX: how will we recover from this? */
808 				toep->flags |= TPF_TX_SUSPENDED;
809 				return;
810 			}
811 			txwr = wrtod(wr);
812 			credits = howmany(wr->wr_len, 16);
813 			write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, plen, plen,
814 			    credits, shove, 0);
815 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
816 			nsegs = 0;
817 		} else {
818 			int wr_len;
819 
820 			/* DSGL tx */
821 
822 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
823 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
824 			wr = alloc_wrqe(roundup2(wr_len, 16),
825 			    &toep->ofld_txq->wrq);
826 			if (wr == NULL) {
827 				/* XXX: how will we recover from this? */
828 				toep->flags |= TPF_TX_SUSPENDED;
829 				return;
830 			}
831 			txwr = wrtod(wr);
832 			credits = howmany(wr_len, 16);
833 			write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, 0, plen,
834 			    credits, shove, 0);
835 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
836 			    max_nsegs_1mbuf);
837 			if (wr_len & 0xf) {
838 				uint64_t *pad = (uint64_t *)
839 				    ((uintptr_t)txwr + wr_len);
840 				*pad = 0;
841 			}
842 		}
843 
844 		KASSERT(toep->tx_credits >= credits,
845 			("%s: not enough credits", __func__));
846 
847 		toep->tx_credits -= credits;
848 		toep->tx_nocompl += credits;
849 		toep->plen_nocompl += plen;
850 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
851 		    toep->tx_nocompl >= toep->tx_total / 4)
852 			compl = 1;
853 
854 		if (compl || ulp_mode(toep) == ULP_MODE_RDMA) {
855 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
856 			toep->tx_nocompl = 0;
857 			toep->plen_nocompl = 0;
858 		}
859 
860 		tp->snd_nxt += plen;
861 		tp->snd_max += plen;
862 
863 		SOCKBUF_LOCK(sb);
864 		KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
865 		sb->sb_sndptr = sb_sndptr;
866 		SOCKBUF_UNLOCK(sb);
867 
868 		toep->flags |= TPF_TX_DATA_SENT;
869 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
870 			toep->flags |= TPF_TX_SUSPENDED;
871 
872 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
873 		txsd->plen = plen;
874 		txsd->tx_credits = credits;
875 		txsd++;
876 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
877 			toep->txsd_pidx = 0;
878 			txsd = &toep->txsd[0];
879 		}
880 		toep->txsd_avail--;
881 
882 		t4_l2t_send(sc, wr, toep->l2te);
883 	} while (m != NULL && (m->m_flags & M_NOTAVAIL) == 0);
884 
885 	/* Send a FIN if requested, but only if there's no more data to send */
886 	if (m == NULL && toep->flags & TPF_SEND_FIN)
887 		t4_close_conn(sc, toep);
888 }
889 
890 static inline void
rqdrop_locked(struct mbufq * q,int plen)891 rqdrop_locked(struct mbufq *q, int plen)
892 {
893 	struct mbuf *m;
894 
895 	while (plen > 0) {
896 		m = mbufq_dequeue(q);
897 
898 		/* Too many credits. */
899 		MPASS(m != NULL);
900 		M_ASSERTPKTHDR(m);
901 
902 		/* Partial credits. */
903 		MPASS(plen >= m->m_pkthdr.len);
904 
905 		plen -= m->m_pkthdr.len;
906 		m_freem(m);
907 	}
908 }
909 
910 /*
911  * Not a bit in the TCB, but is a bit in the ulp_submode field of the
912  * CPL_TX_DATA flags field in FW_ISCSI_TX_DATA_WR.
913  */
914 #define	ULP_ISO		G_TX_ULP_SUBMODE(F_FW_ISCSI_TX_DATA_WR_ULPSUBMODE_ISO)
915 
916 static void
write_tx_data_iso(void * dst,u_int ulp_submode,uint8_t flags,uint16_t mss,int len,int npdu)917 write_tx_data_iso(void *dst, u_int ulp_submode, uint8_t flags, uint16_t mss,
918     int len, int npdu)
919 {
920 	struct cpl_tx_data_iso *cpl;
921 	unsigned int burst_size;
922 	unsigned int last;
923 
924 	/*
925 	 * The firmware will set the 'F' bit on the last PDU when
926 	 * either condition is true:
927 	 *
928 	 * - this large PDU is marked as the "last" slice
929 	 *
930 	 * - the amount of data payload bytes equals the burst_size
931 	 *
932 	 * The strategy used here is to always set the burst_size
933 	 * artificially high (len includes the size of the template
934 	 * BHS) and only set the "last" flag if the original PDU had
935 	 * 'F' set.
936 	 */
937 	burst_size = len;
938 	last = !!(flags & CXGBE_ISO_F);
939 
940 	cpl = (struct cpl_tx_data_iso *)dst;
941 	cpl->op_to_scsi = htonl(V_CPL_TX_DATA_ISO_OP(CPL_TX_DATA_ISO) |
942 	    V_CPL_TX_DATA_ISO_FIRST(1) | V_CPL_TX_DATA_ISO_LAST(last) |
943 	    V_CPL_TX_DATA_ISO_CPLHDRLEN(0) |
944 	    V_CPL_TX_DATA_ISO_HDRCRC(!!(ulp_submode & ULP_CRC_HEADER)) |
945 	    V_CPL_TX_DATA_ISO_PLDCRC(!!(ulp_submode & ULP_CRC_DATA)) |
946 	    V_CPL_TX_DATA_ISO_IMMEDIATE(0) |
947 	    V_CPL_TX_DATA_ISO_SCSI(CXGBE_ISO_TYPE(flags)));
948 
949 	cpl->ahs_len = 0;
950 	cpl->mpdu = htons(DIV_ROUND_UP(mss, 4));
951 	cpl->burst_size = htonl(DIV_ROUND_UP(burst_size, 4));
952 	cpl->len = htonl(len);
953 	cpl->reserved2_seglen_offset = htonl(0);
954 	cpl->datasn_offset = htonl(0);
955 	cpl->buffer_offset = htonl(0);
956 	cpl->reserved3 = 0;
957 }
958 
959 static struct wrqe *
write_iscsi_mbuf_wr(struct toepcb * toep,struct mbuf * sndptr)960 write_iscsi_mbuf_wr(struct toepcb *toep, struct mbuf *sndptr)
961 {
962 	struct mbuf *m;
963 	struct fw_ofld_tx_data_wr *txwr;
964 	struct cpl_tx_data_iso *cpl_iso;
965 	void *p;
966 	struct wrqe *wr;
967 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
968 	u_int adjusted_plen, imm_data, ulp_submode;
969 	struct inpcb *inp = toep->inp;
970 	struct tcpcb *tp = intotcpcb(inp);
971 	int tx_credits, shove, npdu, wr_len;
972 	uint16_t iso_mss;
973 	static const u_int ulp_extra_len[] = {0, 4, 4, 8};
974 	bool iso, nomap_mbuf_seen;
975 
976 	M_ASSERTPKTHDR(sndptr);
977 
978 	tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
979 	if (mbuf_raw_wr(sndptr)) {
980 		plen = sndptr->m_pkthdr.len;
981 		KASSERT(plen <= SGE_MAX_WR_LEN,
982 		    ("raw WR len %u is greater than max WR len", plen));
983 		if (plen > tx_credits * 16)
984 			return (NULL);
985 
986 		wr = alloc_wrqe(roundup2(plen, 16), &toep->ofld_txq->wrq);
987 		if (__predict_false(wr == NULL))
988 			return (NULL);
989 
990 		m_copydata(sndptr, 0, plen, wrtod(wr));
991 		return (wr);
992 	}
993 
994 	iso = mbuf_iscsi_iso(sndptr);
995 	max_imm = max_imm_payload(tx_credits, iso);
996 	max_nsegs = max_dsgl_nsegs(tx_credits, iso);
997 	iso_mss = mbuf_iscsi_iso_mss(sndptr);
998 
999 	plen = 0;
1000 	nsegs = 0;
1001 	max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
1002 	nomap_mbuf_seen = false;
1003 	for (m = sndptr; m != NULL; m = m->m_next) {
1004 		int n;
1005 
1006 		if (m->m_flags & M_EXTPG)
1007 			n = sglist_count_mbuf_epg(m, mtod(m, vm_offset_t),
1008 			    m->m_len);
1009 		else
1010 			n = sglist_count(mtod(m, void *), m->m_len);
1011 
1012 		nsegs += n;
1013 		plen += m->m_len;
1014 
1015 		/*
1016 		 * This mbuf would send us _over_ the nsegs limit.
1017 		 * Suspend tx because the PDU can't be sent out.
1018 		 */
1019 		if ((nomap_mbuf_seen || plen > max_imm) && nsegs > max_nsegs)
1020 			return (NULL);
1021 
1022 		if (m->m_flags & M_EXTPG)
1023 			nomap_mbuf_seen = true;
1024 		if (max_nsegs_1mbuf < n)
1025 			max_nsegs_1mbuf = n;
1026 	}
1027 
1028 	if (__predict_false(toep->flags & TPF_FIN_SENT))
1029 		panic("%s: excess tx.", __func__);
1030 
1031 	/*
1032 	 * We have a PDU to send.  All of it goes out in one WR so 'm'
1033 	 * is NULL.  A PDU's length is always a multiple of 4.
1034 	 */
1035 	MPASS(m == NULL);
1036 	MPASS((plen & 3) == 0);
1037 	MPASS(sndptr->m_pkthdr.len == plen);
1038 
1039 	shove = !(tp->t_flags & TF_MORETOCOME);
1040 
1041 	/*
1042 	 * plen doesn't include header and data digests, which are
1043 	 * generated and inserted in the right places by the TOE, but
1044 	 * they do occupy TCP sequence space and need to be accounted
1045 	 * for.
1046 	 */
1047 	ulp_submode = mbuf_ulp_submode(sndptr);
1048 	MPASS(ulp_submode < nitems(ulp_extra_len));
1049 	npdu = iso ? howmany(plen - ISCSI_BHS_SIZE, iso_mss) : 1;
1050 	adjusted_plen = plen + ulp_extra_len[ulp_submode] * npdu;
1051 	if (iso)
1052 		adjusted_plen += ISCSI_BHS_SIZE * (npdu - 1);
1053 	wr_len = sizeof(*txwr);
1054 	if (iso)
1055 		wr_len += sizeof(struct cpl_tx_data_iso);
1056 	if (plen <= max_imm && !nomap_mbuf_seen) {
1057 		/* Immediate data tx */
1058 		imm_data = plen;
1059 		wr_len += plen;
1060 		nsegs = 0;
1061 	} else {
1062 		/* DSGL tx */
1063 		imm_data = 0;
1064 		wr_len += sizeof(struct ulptx_sgl) +
1065 		    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1066 	}
1067 
1068 	wr = alloc_wrqe(roundup2(wr_len, 16), &toep->ofld_txq->wrq);
1069 	if (wr == NULL) {
1070 		/* XXX: how will we recover from this? */
1071 		return (NULL);
1072 	}
1073 	txwr = wrtod(wr);
1074 	credits = howmany(wr->wr_len, 16);
1075 
1076 	if (iso) {
1077 		write_tx_wr(txwr, toep, FW_ISCSI_TX_DATA_WR,
1078 		    imm_data + sizeof(struct cpl_tx_data_iso),
1079 		    adjusted_plen, credits, shove, ulp_submode | ULP_ISO);
1080 		cpl_iso = (struct cpl_tx_data_iso *)(txwr + 1);
1081 		MPASS(plen == sndptr->m_pkthdr.len);
1082 		write_tx_data_iso(cpl_iso, ulp_submode,
1083 		    mbuf_iscsi_iso_flags(sndptr), iso_mss, plen, npdu);
1084 		p = cpl_iso + 1;
1085 	} else {
1086 		write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, imm_data,
1087 		    adjusted_plen, credits, shove, ulp_submode);
1088 		p = txwr + 1;
1089 	}
1090 
1091 	if (imm_data != 0) {
1092 		m_copydata(sndptr, 0, plen, p);
1093 	} else {
1094 		write_tx_sgl(p, sndptr, m, nsegs, max_nsegs_1mbuf);
1095 		if (wr_len & 0xf) {
1096 			uint64_t *pad = (uint64_t *)((uintptr_t)txwr + wr_len);
1097 			*pad = 0;
1098 		}
1099 	}
1100 
1101 	KASSERT(toep->tx_credits >= credits,
1102 	    ("%s: not enough credits: credits %u "
1103 		"toep->tx_credits %u tx_credits %u nsegs %u "
1104 		"max_nsegs %u iso %d", __func__, credits,
1105 		toep->tx_credits, tx_credits, nsegs, max_nsegs, iso));
1106 
1107 	tp->snd_nxt += adjusted_plen;
1108 	tp->snd_max += adjusted_plen;
1109 
1110 	counter_u64_add(toep->ofld_txq->tx_iscsi_pdus, npdu);
1111 	counter_u64_add(toep->ofld_txq->tx_iscsi_octets, plen);
1112 	if (iso)
1113 		counter_u64_add(toep->ofld_txq->tx_iscsi_iso_wrs, 1);
1114 
1115 	return (wr);
1116 }
1117 
1118 void
t4_push_pdus(struct adapter * sc,struct toepcb * toep,int drop)1119 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
1120 {
1121 	struct mbuf *sndptr, *m;
1122 	struct fw_wr_hdr *wrhdr;
1123 	struct wrqe *wr;
1124 	u_int plen, credits;
1125 	struct inpcb *inp = toep->inp;
1126 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
1127 	struct mbufq *pduq = &toep->ulp_pduq;
1128 
1129 	INP_WLOCK_ASSERT(inp);
1130 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1131 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1132 	KASSERT(ulp_mode(toep) == ULP_MODE_ISCSI,
1133 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1134 
1135 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1136 		return;
1137 
1138 	/*
1139 	 * This function doesn't resume by itself.  Someone else must clear the
1140 	 * flag and call this function.
1141 	 */
1142 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1143 		KASSERT(drop == 0,
1144 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1145 		return;
1146 	}
1147 
1148 	if (drop) {
1149 		struct socket *so = inp->inp_socket;
1150 		struct sockbuf *sb = &so->so_snd;
1151 		int sbu;
1152 
1153 		/*
1154 		 * An unlocked read is ok here as the data should only
1155 		 * transition from a non-zero value to either another
1156 		 * non-zero value or zero.  Once it is zero it should
1157 		 * stay zero.
1158 		 */
1159 		if (__predict_false(sbused(sb)) > 0) {
1160 			SOCKBUF_LOCK(sb);
1161 			sbu = sbused(sb);
1162 			if (sbu > 0) {
1163 				/*
1164 				 * The data transmitted before the
1165 				 * tid's ULP mode changed to ISCSI is
1166 				 * still in so_snd.  Incoming credits
1167 				 * should account for so_snd first.
1168 				 */
1169 				sbdrop_locked(sb, min(sbu, drop));
1170 				drop -= min(sbu, drop);
1171 			}
1172 			sowwakeup_locked(so);	/* unlocks so_snd */
1173 		}
1174 		rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
1175 	}
1176 
1177 	while ((sndptr = mbufq_first(pduq)) != NULL) {
1178 		wr = write_iscsi_mbuf_wr(toep, sndptr);
1179 		if (wr == NULL) {
1180 			toep->flags |= TPF_TX_SUSPENDED;
1181 			return;
1182 		}
1183 
1184 		plen = sndptr->m_pkthdr.len;
1185 		credits = howmany(wr->wr_len, 16);
1186 		KASSERT(toep->tx_credits >= credits,
1187 			("%s: not enough credits", __func__));
1188 
1189 		m = mbufq_dequeue(pduq);
1190 		MPASS(m == sndptr);
1191 		mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
1192 
1193 		toep->tx_credits -= credits;
1194 		toep->tx_nocompl += credits;
1195 		toep->plen_nocompl += plen;
1196 
1197 		/*
1198 		 * Ensure there are enough credits for a full-sized WR
1199 		 * as page pod WRs can be full-sized.
1200 		 */
1201 		if (toep->tx_credits <= SGE_MAX_WR_LEN * 5 / 4 &&
1202 		    toep->tx_nocompl >= toep->tx_total / 4) {
1203 			wrhdr = wrtod(wr);
1204 			wrhdr->hi |= htobe32(F_FW_WR_COMPL);
1205 			toep->tx_nocompl = 0;
1206 			toep->plen_nocompl = 0;
1207 		}
1208 
1209 		toep->flags |= TPF_TX_DATA_SENT;
1210 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
1211 			toep->flags |= TPF_TX_SUSPENDED;
1212 
1213 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1214 		txsd->plen = plen;
1215 		txsd->tx_credits = credits;
1216 		txsd++;
1217 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1218 			toep->txsd_pidx = 0;
1219 			txsd = &toep->txsd[0];
1220 		}
1221 		toep->txsd_avail--;
1222 
1223 		t4_l2t_send(sc, wr, toep->l2te);
1224 	}
1225 
1226 	/* Send a FIN if requested, but only if there are no more PDUs to send */
1227 	if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
1228 		t4_close_conn(sc, toep);
1229 }
1230 
1231 static inline void
t4_push_data(struct adapter * sc,struct toepcb * toep,int drop)1232 t4_push_data(struct adapter *sc, struct toepcb *toep, int drop)
1233 {
1234 
1235 	if (ulp_mode(toep) == ULP_MODE_ISCSI)
1236 		t4_push_pdus(sc, toep, drop);
1237 	else if (toep->flags & TPF_KTLS)
1238 		t4_push_ktls(sc, toep, drop);
1239 	else
1240 		t4_push_frames(sc, toep, drop);
1241 }
1242 
1243 int
t4_tod_output(struct toedev * tod,struct tcpcb * tp)1244 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
1245 {
1246 	struct adapter *sc = tod->tod_softc;
1247 #ifdef INVARIANTS
1248 	struct inpcb *inp = tptoinpcb(tp);
1249 #endif
1250 	struct toepcb *toep = tp->t_toe;
1251 
1252 	INP_WLOCK_ASSERT(inp);
1253 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1254 	    ("%s: inp %p dropped.", __func__, inp));
1255 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1256 
1257 	t4_push_data(sc, toep, 0);
1258 
1259 	return (0);
1260 }
1261 
1262 int
t4_send_fin(struct toedev * tod,struct tcpcb * tp)1263 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
1264 {
1265 	struct adapter *sc = tod->tod_softc;
1266 #ifdef INVARIANTS
1267 	struct inpcb *inp = tptoinpcb(tp);
1268 #endif
1269 	struct toepcb *toep = tp->t_toe;
1270 
1271 	INP_WLOCK_ASSERT(inp);
1272 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1273 	    ("%s: inp %p dropped.", __func__, inp));
1274 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1275 
1276 	toep->flags |= TPF_SEND_FIN;
1277 	if (tp->t_state >= TCPS_ESTABLISHED)
1278 		t4_push_data(sc, toep, 0);
1279 
1280 	return (0);
1281 }
1282 
1283 int
t4_send_rst(struct toedev * tod,struct tcpcb * tp)1284 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1285 {
1286 	struct adapter *sc = tod->tod_softc;
1287 #if defined(INVARIANTS)
1288 	struct inpcb *inp = tptoinpcb(tp);
1289 #endif
1290 	struct toepcb *toep = tp->t_toe;
1291 
1292 	INP_WLOCK_ASSERT(inp);
1293 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1294 	    ("%s: inp %p dropped.", __func__, inp));
1295 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1296 
1297 	/* hmmmm */
1298 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1299 	    ("%s: flowc for tid %u [%s] not sent already",
1300 	    __func__, toep->tid, tcpstates[tp->t_state]));
1301 
1302 	send_reset(sc, toep, 0);
1303 	return (0);
1304 }
1305 
1306 /*
1307  * Peer has sent us a FIN.
1308  */
1309 static int
do_peer_close(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)1310 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1311 {
1312 	struct adapter *sc = iq->adapter;
1313 	const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1314 	unsigned int tid = GET_TID(cpl);
1315 	struct toepcb *toep = lookup_tid(sc, tid);
1316 	struct inpcb *inp = toep->inp;
1317 	struct tcpcb *tp = NULL;
1318 	struct socket *so;
1319 	struct epoch_tracker et;
1320 #ifdef INVARIANTS
1321 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1322 #endif
1323 
1324 	KASSERT(opcode == CPL_PEER_CLOSE,
1325 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1326 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1327 
1328 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1329 		/*
1330 		 * do_pass_establish must have run before do_peer_close and if
1331 		 * this is still a synqe instead of a toepcb then the connection
1332 		 * must be getting aborted.
1333 		 */
1334 		MPASS(toep->flags & TPF_ABORT_SHUTDOWN);
1335 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1336 		    toep, toep->flags);
1337 		return (0);
1338 	}
1339 
1340 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1341 
1342 	CURVNET_SET(toep->vnet);
1343 	NET_EPOCH_ENTER(et);
1344 	INP_WLOCK(inp);
1345 	tp = intotcpcb(inp);
1346 
1347 	CTR6(KTR_CXGBE,
1348 	    "%s: tid %u (%s), toep_flags 0x%x, ddp_flags 0x%x, inp %p",
1349 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1350 	    toep->ddp.flags, inp);
1351 
1352 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1353 		goto done;
1354 
1355 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1356 		DDP_LOCK(toep);
1357 		if (__predict_false(toep->ddp.flags &
1358 		    (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)))
1359 			handle_ddp_close(toep, tp, cpl->rcv_nxt);
1360 		DDP_UNLOCK(toep);
1361 	}
1362 	so = inp->inp_socket;
1363 	socantrcvmore(so);
1364 
1365 	if (ulp_mode(toep) == ULP_MODE_RDMA ||
1366 	    (ulp_mode(toep) == ULP_MODE_ISCSI && chip_id(sc) >= CHELSIO_T6)) {
1367 		/*
1368 		 * There might be data received via DDP before the FIN
1369 		 * not reported to the driver.  Just assume the
1370 		 * sequence number in the CPL is correct as the
1371 		 * sequence number of the FIN.
1372 		 */
1373 	} else {
1374 		KASSERT(tp->rcv_nxt + 1 == be32toh(cpl->rcv_nxt),
1375 		    ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1376 		    be32toh(cpl->rcv_nxt)));
1377 	}
1378 
1379 	tp->rcv_nxt = be32toh(cpl->rcv_nxt);
1380 
1381 	switch (tp->t_state) {
1382 	case TCPS_SYN_RECEIVED:
1383 		tp->t_starttime = ticks;
1384 		/* FALLTHROUGH */
1385 
1386 	case TCPS_ESTABLISHED:
1387 		tcp_state_change(tp, TCPS_CLOSE_WAIT);
1388 		break;
1389 
1390 	case TCPS_FIN_WAIT_1:
1391 		tcp_state_change(tp, TCPS_CLOSING);
1392 		break;
1393 
1394 	case TCPS_FIN_WAIT_2:
1395 		restore_so_proto(so, inp->inp_vflag & INP_IPV6);
1396 		t4_pcb_detach(NULL, tp);
1397 		tcp_twstart(tp);
1398 		INP_UNLOCK_ASSERT(inp);	 /* safe, we have a ref on the inp */
1399 		NET_EPOCH_EXIT(et);
1400 		CURVNET_RESTORE();
1401 
1402 		INP_WLOCK(inp);
1403 		final_cpl_received(toep);
1404 		return (0);
1405 
1406 	default:
1407 		log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1408 		    __func__, tid, tp->t_state);
1409 	}
1410 done:
1411 	INP_WUNLOCK(inp);
1412 	NET_EPOCH_EXIT(et);
1413 	CURVNET_RESTORE();
1414 	return (0);
1415 }
1416 
1417 /*
1418  * Peer has ACK'd our FIN.
1419  */
1420 static int
do_close_con_rpl(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)1421 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1422     struct mbuf *m)
1423 {
1424 	struct adapter *sc = iq->adapter;
1425 	const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1426 	unsigned int tid = GET_TID(cpl);
1427 	struct toepcb *toep = lookup_tid(sc, tid);
1428 	struct inpcb *inp = toep->inp;
1429 	struct tcpcb *tp = NULL;
1430 	struct socket *so = NULL;
1431 	struct epoch_tracker et;
1432 #ifdef INVARIANTS
1433 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1434 #endif
1435 
1436 	KASSERT(opcode == CPL_CLOSE_CON_RPL,
1437 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1438 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1439 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1440 
1441 	CURVNET_SET(toep->vnet);
1442 	NET_EPOCH_ENTER(et);
1443 	INP_WLOCK(inp);
1444 	tp = intotcpcb(inp);
1445 
1446 	CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1447 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1448 
1449 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1450 		goto done;
1451 
1452 	so = inp->inp_socket;
1453 	tp->snd_una = be32toh(cpl->snd_nxt) - 1;	/* exclude FIN */
1454 
1455 	switch (tp->t_state) {
1456 	case TCPS_CLOSING:	/* see TCPS_FIN_WAIT_2 in do_peer_close too */
1457 		restore_so_proto(so, inp->inp_vflag & INP_IPV6);
1458 		t4_pcb_detach(NULL, tp);
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
send_abort_rpl(struct adapter * sc,struct sge_ofld_txq * ofld_txq,int tid,int rst_status)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
abort_status_to_errno(struct tcpcb * tp,unsigned int abort_reason)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
do_abort_req(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)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
do_abort_rpl(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)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
do_rx_data(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)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;
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 if (cpl->ddp_off == 1) {
1759 				/* Fell out of DDP mode */
1760 				toep->ddp.flags &= ~DDP_ON;
1761 				CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
1762 				    __func__);
1763 
1764 				insert_ddp_data(toep, ddp_placed);
1765 			} else {
1766 				/*
1767 				 * Data was received while still
1768 				 * ULP_MODE_NONE, just fall through.
1769 				 */
1770 			}
1771 		}
1772 
1773 		if (toep->ddp.flags & DDP_ON) {
1774 			/*
1775 			 * CPL_RX_DATA with DDP on can only be an indicate.
1776 			 * Start posting queued AIO requests via DDP.  The
1777 			 * payload that arrived in this indicate is appended
1778 			 * to the socket buffer as usual.
1779 			 */
1780 			handle_ddp_indicate(toep);
1781 		}
1782 	}
1783 
1784 	sbappendstream_locked(sb, m, 0);
1785 	t4_rcvd_locked(&toep->td->tod, tp);
1786 
1787 	if (ulp_mode(toep) == ULP_MODE_TCPDDP &&
1788 	    (toep->ddp.flags & DDP_AIO) != 0 && toep->ddp.waiting_count > 0 &&
1789 	    sbavail(sb) != 0) {
1790 		CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__,
1791 		    tid);
1792 		ddp_queue_toep(toep);
1793 	}
1794 	if (toep->flags & TPF_TLS_STARTING)
1795 		tls_received_starting_data(sc, toep, sb, len);
1796 	sorwakeup_locked(so);
1797 	SOCKBUF_UNLOCK_ASSERT(sb);
1798 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1799 		DDP_UNLOCK(toep);
1800 
1801 	INP_WUNLOCK(inp);
1802 	CURVNET_RESTORE();
1803 	return (0);
1804 }
1805 
1806 static int
do_fw4_ack(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)1807 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1808 {
1809 	struct adapter *sc = iq->adapter;
1810 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1811 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1812 	struct toepcb *toep = lookup_tid(sc, tid);
1813 	struct inpcb *inp;
1814 	struct tcpcb *tp;
1815 	struct socket *so;
1816 	uint8_t credits = cpl->credits;
1817 	struct ofld_tx_sdesc *txsd;
1818 	int plen;
1819 #ifdef INVARIANTS
1820 	unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1821 #endif
1822 
1823 	/*
1824 	 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1825 	 * now this comes back carrying the credits for the flowc.
1826 	 */
1827 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1828 		KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1829 		    ("%s: credits for a synq entry %p", __func__, toep));
1830 		return (0);
1831 	}
1832 
1833 	inp = toep->inp;
1834 
1835 	KASSERT(opcode == CPL_FW4_ACK,
1836 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1837 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1838 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1839 
1840 	INP_WLOCK(inp);
1841 
1842 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1843 		INP_WUNLOCK(inp);
1844 		return (0);
1845 	}
1846 
1847 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1848 	    ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1849 
1850 	tp = intotcpcb(inp);
1851 
1852 	if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1853 		tcp_seq snd_una = be32toh(cpl->snd_una);
1854 
1855 #ifdef INVARIANTS
1856 		if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1857 			log(LOG_ERR,
1858 			    "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1859 			    __func__, snd_una, toep->tid, tp->snd_una);
1860 		}
1861 #endif
1862 
1863 		if (tp->snd_una != snd_una) {
1864 			tp->snd_una = snd_una;
1865 			tp->ts_recent_age = tcp_ts_getticks();
1866 		}
1867 	}
1868 
1869 #ifdef VERBOSE_TRACES
1870 	CTR3(KTR_CXGBE, "%s: tid %d credits %u", __func__, tid, credits);
1871 #endif
1872 	so = inp->inp_socket;
1873 	txsd = &toep->txsd[toep->txsd_cidx];
1874 	plen = 0;
1875 	while (credits) {
1876 		KASSERT(credits >= txsd->tx_credits,
1877 		    ("%s: too many (or partial) credits", __func__));
1878 		credits -= txsd->tx_credits;
1879 		toep->tx_credits += txsd->tx_credits;
1880 		plen += txsd->plen;
1881 		txsd++;
1882 		toep->txsd_avail++;
1883 		KASSERT(toep->txsd_avail <= toep->txsd_total,
1884 		    ("%s: txsd avail > total", __func__));
1885 		if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1886 			txsd = &toep->txsd[0];
1887 			toep->txsd_cidx = 0;
1888 		}
1889 	}
1890 
1891 	if (toep->tx_credits == toep->tx_total) {
1892 		toep->tx_nocompl = 0;
1893 		toep->plen_nocompl = 0;
1894 	}
1895 
1896 	if (toep->flags & TPF_TX_SUSPENDED &&
1897 	    toep->tx_credits >= toep->tx_total / 4) {
1898 #ifdef VERBOSE_TRACES
1899 		CTR2(KTR_CXGBE, "%s: tid %d calling t4_push_frames", __func__,
1900 		    tid);
1901 #endif
1902 		toep->flags &= ~TPF_TX_SUSPENDED;
1903 		CURVNET_SET(toep->vnet);
1904 		t4_push_data(sc, toep, plen);
1905 		CURVNET_RESTORE();
1906 	} else if (plen > 0) {
1907 		struct sockbuf *sb = &so->so_snd;
1908 		int sbu;
1909 
1910 		SOCKBUF_LOCK(sb);
1911 		sbu = sbused(sb);
1912 		if (ulp_mode(toep) == ULP_MODE_ISCSI) {
1913 			if (__predict_false(sbu > 0)) {
1914 				/*
1915 				 * The data transmitted before the
1916 				 * tid's ULP mode changed to ISCSI is
1917 				 * still in so_snd.  Incoming credits
1918 				 * should account for so_snd first.
1919 				 */
1920 				sbdrop_locked(sb, min(sbu, plen));
1921 				plen -= min(sbu, plen);
1922 			}
1923 			sowwakeup_locked(so);	/* unlocks so_snd */
1924 			rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
1925 		} else {
1926 #ifdef VERBOSE_TRACES
1927 			CTR3(KTR_CXGBE, "%s: tid %d dropped %d bytes", __func__,
1928 			    tid, plen);
1929 #endif
1930 			sbdrop_locked(sb, plen);
1931 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
1932 				t4_aiotx_queue_toep(so, toep);
1933 			sowwakeup_locked(so);	/* unlocks so_snd */
1934 		}
1935 		SOCKBUF_UNLOCK_ASSERT(sb);
1936 	}
1937 
1938 	INP_WUNLOCK(inp);
1939 
1940 	return (0);
1941 }
1942 
1943 void
t4_set_tcb_field(struct adapter * sc,struct sge_wrq * wrq,struct toepcb * toep,uint16_t word,uint64_t mask,uint64_t val,int reply,int cookie)1944 t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, struct toepcb *toep,
1945     uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie)
1946 {
1947 	struct wrqe *wr;
1948 	struct cpl_set_tcb_field *req;
1949 	struct ofld_tx_sdesc *txsd;
1950 
1951 	MPASS((cookie & ~M_COOKIE) == 0);
1952 	if (reply) {
1953 		MPASS(cookie != CPL_COOKIE_RESERVED);
1954 	}
1955 
1956 	wr = alloc_wrqe(sizeof(*req), wrq);
1957 	if (wr == NULL) {
1958 		/* XXX */
1959 		panic("%s: allocation failure.", __func__);
1960 	}
1961 	req = wrtod(wr);
1962 
1963 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
1964 	req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id));
1965 	if (reply == 0)
1966 		req->reply_ctrl |= htobe16(F_NO_REPLY);
1967 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie));
1968 	req->mask = htobe64(mask);
1969 	req->val = htobe64(val);
1970 	if (wrq->eq.type == EQ_OFLD) {
1971 		txsd = &toep->txsd[toep->txsd_pidx];
1972 		txsd->tx_credits = howmany(sizeof(*req), 16);
1973 		txsd->plen = 0;
1974 		KASSERT(toep->tx_credits >= txsd->tx_credits &&
1975 		    toep->txsd_avail > 0,
1976 		    ("%s: not enough credits (%d)", __func__,
1977 		    toep->tx_credits));
1978 		toep->tx_credits -= txsd->tx_credits;
1979 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1980 			toep->txsd_pidx = 0;
1981 		toep->txsd_avail--;
1982 	}
1983 
1984 	t4_wrq_tx(sc, wr);
1985 }
1986 
1987 void
t4_init_cpl_io_handlers(void)1988 t4_init_cpl_io_handlers(void)
1989 {
1990 
1991 	t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close);
1992 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl);
1993 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req);
1994 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl,
1995 	    CPL_COOKIE_TOM);
1996 	t4_register_cpl_handler(CPL_RX_DATA, do_rx_data);
1997 	t4_register_shared_cpl_handler(CPL_FW4_ACK, do_fw4_ack, CPL_COOKIE_TOM);
1998 }
1999 
2000 void
t4_uninit_cpl_io_handlers(void)2001 t4_uninit_cpl_io_handlers(void)
2002 {
2003 
2004 	t4_register_cpl_handler(CPL_PEER_CLOSE, NULL);
2005 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, NULL);
2006 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, NULL);
2007 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, NULL, CPL_COOKIE_TOM);
2008 	t4_register_cpl_handler(CPL_RX_DATA, NULL);
2009 	t4_register_shared_cpl_handler(CPL_FW4_ACK, NULL, CPL_COOKIE_TOM);
2010 }
2011 
2012 /*
2013  * Use the 'backend1' field in AIO jobs to hold an error that should
2014  * be reported when the job is completed, the 'backend3' field to
2015  * store the amount of data sent by the AIO job so far, and the
2016  * 'backend4' field to hold a reference count on the job.
2017  *
2018  * Each unmapped mbuf holds a reference on the job as does the queue
2019  * so long as the job is queued.
2020  */
2021 #define	aio_error	backend1
2022 #define	aio_sent	backend3
2023 #define	aio_refs	backend4
2024 
2025 #ifdef VERBOSE_TRACES
2026 static int
jobtotid(struct kaiocb * job)2027 jobtotid(struct kaiocb *job)
2028 {
2029 	struct socket *so;
2030 	struct tcpcb *tp;
2031 	struct toepcb *toep;
2032 
2033 	so = job->fd_file->f_data;
2034 	tp = sototcpcb(so);
2035 	toep = tp->t_toe;
2036 	return (toep->tid);
2037 }
2038 #endif
2039 
2040 static void
aiotx_free_job(struct kaiocb * job)2041 aiotx_free_job(struct kaiocb *job)
2042 {
2043 	long status;
2044 	int error;
2045 
2046 	if (refcount_release(&job->aio_refs) == 0)
2047 		return;
2048 
2049 	error = (intptr_t)job->aio_error;
2050 	status = job->aio_sent;
2051 #ifdef VERBOSE_TRACES
2052 	CTR5(KTR_CXGBE, "%s: tid %d completed %p len %ld, error %d", __func__,
2053 	    jobtotid(job), job, status, error);
2054 #endif
2055 	if (error != 0 && status != 0)
2056 		error = 0;
2057 	if (error == ECANCELED)
2058 		aio_cancel(job);
2059 	else if (error)
2060 		aio_complete(job, -1, error);
2061 	else {
2062 		job->msgsnd = 1;
2063 		aio_complete(job, status, 0);
2064 	}
2065 }
2066 
2067 static void
aiotx_free_pgs(struct mbuf * m)2068 aiotx_free_pgs(struct mbuf *m)
2069 {
2070 	struct kaiocb *job;
2071 	vm_page_t pg;
2072 
2073 	M_ASSERTEXTPG(m);
2074 	job = m->m_ext.ext_arg1;
2075 #ifdef VERBOSE_TRACES
2076 	CTR3(KTR_CXGBE, "%s: completed %d bytes for tid %d", __func__,
2077 	    m->m_len, jobtotid(job));
2078 #endif
2079 
2080 	for (int i = 0; i < m->m_epg_npgs; i++) {
2081 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
2082 		vm_page_unwire(pg, PQ_ACTIVE);
2083 	}
2084 
2085 	aiotx_free_job(job);
2086 }
2087 
2088 /*
2089  * Allocate a chain of unmapped mbufs describing the next 'len' bytes
2090  * of an AIO job.
2091  */
2092 static struct mbuf *
alloc_aiotx_mbuf(struct kaiocb * job,int len)2093 alloc_aiotx_mbuf(struct kaiocb *job, int len)
2094 {
2095 	struct vmspace *vm;
2096 	vm_page_t pgs[MBUF_PEXT_MAX_PGS];
2097 	struct mbuf *m, *top, *last;
2098 	vm_map_t map;
2099 	vm_offset_t start;
2100 	int i, mlen, npages, pgoff;
2101 
2102 	KASSERT(job->aio_sent + len <= job->uaiocb.aio_nbytes,
2103 	    ("%s(%p, %d): request to send beyond end of buffer", __func__,
2104 	    job, len));
2105 
2106 	/*
2107 	 * The AIO subsystem will cancel and drain all requests before
2108 	 * permitting a process to exit or exec, so p_vmspace should
2109 	 * be stable here.
2110 	 */
2111 	vm = job->userproc->p_vmspace;
2112 	map = &vm->vm_map;
2113 	start = (uintptr_t)job->uaiocb.aio_buf + job->aio_sent;
2114 	pgoff = start & PAGE_MASK;
2115 
2116 	top = NULL;
2117 	last = NULL;
2118 	while (len > 0) {
2119 		mlen = imin(len, MBUF_PEXT_MAX_PGS * PAGE_SIZE - pgoff);
2120 		KASSERT(mlen == len || ((start + mlen) & PAGE_MASK) == 0,
2121 		    ("%s: next start (%#jx + %#x) is not page aligned",
2122 		    __func__, (uintmax_t)start, mlen));
2123 
2124 		npages = vm_fault_quick_hold_pages(map, start, mlen,
2125 		    VM_PROT_WRITE, pgs, nitems(pgs));
2126 		if (npages < 0)
2127 			break;
2128 
2129 		m = mb_alloc_ext_pgs(M_WAITOK, aiotx_free_pgs, M_RDONLY);
2130 		m->m_epg_1st_off = pgoff;
2131 		m->m_epg_npgs = npages;
2132 		if (npages == 1) {
2133 			KASSERT(mlen + pgoff <= PAGE_SIZE,
2134 			    ("%s: single page is too large (off %d len %d)",
2135 			    __func__, pgoff, mlen));
2136 			m->m_epg_last_len = mlen;
2137 		} else {
2138 			m->m_epg_last_len = mlen - (PAGE_SIZE - pgoff) -
2139 			    (npages - 2) * PAGE_SIZE;
2140 		}
2141 		for (i = 0; i < npages; i++)
2142 			m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pgs[i]);
2143 
2144 		m->m_len = mlen;
2145 		m->m_ext.ext_size = npages * PAGE_SIZE;
2146 		m->m_ext.ext_arg1 = job;
2147 		refcount_acquire(&job->aio_refs);
2148 
2149 #ifdef VERBOSE_TRACES
2150 		CTR5(KTR_CXGBE, "%s: tid %d, new mbuf %p for job %p, npages %d",
2151 		    __func__, jobtotid(job), m, job, npages);
2152 #endif
2153 
2154 		if (top == NULL)
2155 			top = m;
2156 		else
2157 			last->m_next = m;
2158 		last = m;
2159 
2160 		len -= mlen;
2161 		start += mlen;
2162 		pgoff = 0;
2163 	}
2164 
2165 	return (top);
2166 }
2167 
2168 static void
t4_aiotx_process_job(struct toepcb * toep,struct socket * so,struct kaiocb * job)2169 t4_aiotx_process_job(struct toepcb *toep, struct socket *so, struct kaiocb *job)
2170 {
2171 	struct sockbuf *sb;
2172 	struct inpcb *inp;
2173 	struct tcpcb *tp;
2174 	struct mbuf *m;
2175 	u_int sent;
2176 	int error, len;
2177 	bool moretocome, sendmore;
2178 
2179 	sb = &so->so_snd;
2180 	SOCKBUF_UNLOCK(sb);
2181 	m = NULL;
2182 
2183 #ifdef MAC
2184 	error = mac_socket_check_send(job->fd_file->f_cred, so);
2185 	if (error != 0)
2186 		goto out;
2187 #endif
2188 
2189 	/* Inline sosend_generic(). */
2190 
2191 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
2192 	MPASS(error == 0);
2193 
2194 sendanother:
2195 	SOCKBUF_LOCK(sb);
2196 	if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
2197 		SOCKBUF_UNLOCK(sb);
2198 		SOCK_IO_SEND_UNLOCK(so);
2199 		if ((so->so_options & SO_NOSIGPIPE) == 0) {
2200 			PROC_LOCK(job->userproc);
2201 			kern_psignal(job->userproc, SIGPIPE);
2202 			PROC_UNLOCK(job->userproc);
2203 		}
2204 		error = EPIPE;
2205 		goto out;
2206 	}
2207 	if (so->so_error) {
2208 		error = so->so_error;
2209 		so->so_error = 0;
2210 		SOCKBUF_UNLOCK(sb);
2211 		SOCK_IO_SEND_UNLOCK(so);
2212 		goto out;
2213 	}
2214 	if ((so->so_state & SS_ISCONNECTED) == 0) {
2215 		SOCKBUF_UNLOCK(sb);
2216 		SOCK_IO_SEND_UNLOCK(so);
2217 		error = ENOTCONN;
2218 		goto out;
2219 	}
2220 	if (sbspace(sb) < sb->sb_lowat) {
2221 		MPASS(job->aio_sent == 0 || !(so->so_state & SS_NBIO));
2222 
2223 		/*
2224 		 * Don't block if there is too little room in the socket
2225 		 * buffer.  Instead, requeue the request.
2226 		 */
2227 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2228 			SOCKBUF_UNLOCK(sb);
2229 			SOCK_IO_SEND_UNLOCK(so);
2230 			error = ECANCELED;
2231 			goto out;
2232 		}
2233 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2234 		SOCKBUF_UNLOCK(sb);
2235 		SOCK_IO_SEND_UNLOCK(so);
2236 		goto out;
2237 	}
2238 
2239 	/*
2240 	 * Write as much data as the socket permits, but no more than a
2241 	 * a single sndbuf at a time.
2242 	 */
2243 	len = sbspace(sb);
2244 	if (len > job->uaiocb.aio_nbytes - job->aio_sent) {
2245 		len = job->uaiocb.aio_nbytes - job->aio_sent;
2246 		moretocome = false;
2247 	} else
2248 		moretocome = true;
2249 	if (len > toep->params.sndbuf) {
2250 		len = toep->params.sndbuf;
2251 		sendmore = true;
2252 	} else
2253 		sendmore = false;
2254 
2255 	if (!TAILQ_EMPTY(&toep->aiotx_jobq))
2256 		moretocome = true;
2257 	SOCKBUF_UNLOCK(sb);
2258 	MPASS(len != 0);
2259 
2260 	m = alloc_aiotx_mbuf(job, len);
2261 	if (m == NULL) {
2262 		SOCK_IO_SEND_UNLOCK(so);
2263 		error = EFAULT;
2264 		goto out;
2265 	}
2266 
2267 	/* Inlined tcp_usr_send(). */
2268 
2269 	inp = toep->inp;
2270 	INP_WLOCK(inp);
2271 	if (inp->inp_flags & INP_DROPPED) {
2272 		INP_WUNLOCK(inp);
2273 		SOCK_IO_SEND_UNLOCK(so);
2274 		error = ECONNRESET;
2275 		goto out;
2276 	}
2277 
2278 	sent = m_length(m, NULL);
2279 	job->aio_sent += sent;
2280 	counter_u64_add(toep->ofld_txq->tx_aio_octets, sent);
2281 
2282 	sbappendstream(sb, m, 0);
2283 	m = NULL;
2284 
2285 	if (!(inp->inp_flags & INP_DROPPED)) {
2286 		tp = intotcpcb(inp);
2287 		if (moretocome)
2288 			tp->t_flags |= TF_MORETOCOME;
2289 		error = tcp_output(tp);
2290 		if (error < 0) {
2291 			INP_UNLOCK_ASSERT(inp);
2292 			SOCK_IO_SEND_UNLOCK(so);
2293 			error = -error;
2294 			goto out;
2295 		}
2296 		if (moretocome)
2297 			tp->t_flags &= ~TF_MORETOCOME;
2298 	}
2299 
2300 	INP_WUNLOCK(inp);
2301 	if (sendmore)
2302 		goto sendanother;
2303 	SOCK_IO_SEND_UNLOCK(so);
2304 
2305 	if (error)
2306 		goto out;
2307 
2308 	/*
2309 	 * If this is a blocking socket and the request has not been
2310 	 * fully completed, requeue it until the socket is ready
2311 	 * again.
2312 	 */
2313 	if (job->aio_sent < job->uaiocb.aio_nbytes &&
2314 	    !(so->so_state & SS_NBIO)) {
2315 		SOCKBUF_LOCK(sb);
2316 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2317 			SOCKBUF_UNLOCK(sb);
2318 			error = ECANCELED;
2319 			goto out;
2320 		}
2321 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2322 		return;
2323 	}
2324 
2325 	/*
2326 	 * If the request will not be requeued, drop the queue's
2327 	 * reference to the job.  Any mbufs in flight should still
2328 	 * hold a reference, but this drops the reference that the
2329 	 * queue owns while it is waiting to queue mbufs to the
2330 	 * socket.
2331 	 */
2332 	aiotx_free_job(job);
2333 	counter_u64_add(toep->ofld_txq->tx_aio_jobs, 1);
2334 
2335 out:
2336 	if (error) {
2337 		job->aio_error = (void *)(intptr_t)error;
2338 		aiotx_free_job(job);
2339 	}
2340 	m_freem(m);
2341 	SOCKBUF_LOCK(sb);
2342 }
2343 
2344 static void
t4_aiotx_task(void * context,int pending)2345 t4_aiotx_task(void *context, int pending)
2346 {
2347 	struct toepcb *toep = context;
2348 	struct socket *so;
2349 	struct kaiocb *job;
2350 	struct epoch_tracker et;
2351 
2352 	so = toep->aiotx_so;
2353 	CURVNET_SET(toep->vnet);
2354 	NET_EPOCH_ENTER(et);
2355 	SOCKBUF_LOCK(&so->so_snd);
2356 	while (!TAILQ_EMPTY(&toep->aiotx_jobq) && sowriteable(so)) {
2357 		job = TAILQ_FIRST(&toep->aiotx_jobq);
2358 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2359 		if (!aio_clear_cancel_function(job))
2360 			continue;
2361 
2362 		t4_aiotx_process_job(toep, so, job);
2363 	}
2364 	toep->aiotx_so = NULL;
2365 	SOCKBUF_UNLOCK(&so->so_snd);
2366 	NET_EPOCH_EXIT(et);
2367 
2368 	free_toepcb(toep);
2369 	sorele(so);
2370 	CURVNET_RESTORE();
2371 }
2372 
2373 static void
t4_aiotx_queue_toep(struct socket * so,struct toepcb * toep)2374 t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep)
2375 {
2376 
2377 	SOCKBUF_LOCK_ASSERT(&toep->inp->inp_socket->so_snd);
2378 #ifdef VERBOSE_TRACES
2379 	CTR3(KTR_CXGBE, "%s: queueing aiotx task for tid %d, active = %s",
2380 	    __func__, toep->tid, toep->aiotx_so != NULL ? "true" : "false");
2381 #endif
2382 	if (toep->aiotx_so != NULL)
2383 		return;
2384 	soref(so);
2385 	toep->aiotx_so = so;
2386 	hold_toepcb(toep);
2387 	soaio_enqueue(&toep->aiotx_task);
2388 }
2389 
2390 static void
t4_aiotx_cancel(struct kaiocb * job)2391 t4_aiotx_cancel(struct kaiocb *job)
2392 {
2393 	struct socket *so;
2394 	struct sockbuf *sb;
2395 	struct tcpcb *tp;
2396 	struct toepcb *toep;
2397 
2398 	so = job->fd_file->f_data;
2399 	tp = sototcpcb(so);
2400 	toep = tp->t_toe;
2401 	MPASS(job->uaiocb.aio_lio_opcode == LIO_WRITE);
2402 	sb = &so->so_snd;
2403 
2404 	SOCKBUF_LOCK(sb);
2405 	if (!aio_cancel_cleared(job))
2406 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2407 	SOCKBUF_UNLOCK(sb);
2408 
2409 	job->aio_error = (void *)(intptr_t)ECANCELED;
2410 	aiotx_free_job(job);
2411 }
2412 
2413 int
t4_aio_queue_aiotx(struct socket * so,struct kaiocb * job)2414 t4_aio_queue_aiotx(struct socket *so, struct kaiocb *job)
2415 {
2416 	struct tcpcb *tp = sototcpcb(so);
2417 	struct toepcb *toep = tp->t_toe;
2418 	struct adapter *sc = td_adapter(toep->td);
2419 
2420 	/* This only handles writes. */
2421 	if (job->uaiocb.aio_lio_opcode != LIO_WRITE)
2422 		return (EOPNOTSUPP);
2423 
2424 	if (!sc->tt.tx_zcopy)
2425 		return (EOPNOTSUPP);
2426 
2427 	if (tls_tx_key(toep))
2428 		return (EOPNOTSUPP);
2429 
2430 	SOCKBUF_LOCK(&so->so_snd);
2431 #ifdef VERBOSE_TRACES
2432 	CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid);
2433 #endif
2434 	if (!aio_set_cancel_function(job, t4_aiotx_cancel))
2435 		panic("new job was cancelled");
2436 	refcount_init(&job->aio_refs, 1);
2437 	TAILQ_INSERT_TAIL(&toep->aiotx_jobq, job, list);
2438 	if (sowriteable(so))
2439 		t4_aiotx_queue_toep(so, toep);
2440 	SOCKBUF_UNLOCK(&so->so_snd);
2441 	return (0);
2442 }
2443 
2444 void
aiotx_init_toep(struct toepcb * toep)2445 aiotx_init_toep(struct toepcb *toep)
2446 {
2447 
2448 	TAILQ_INIT(&toep->aiotx_jobq);
2449 	TASK_INIT(&toep->aiotx_task, 0, t4_aiotx_task, toep);
2450 }
2451 #endif
2452