xref: /freebsd/sys/dev/cxgbe/tom/t4_cpl_io.c (revision d869395ac4bd248da7c5bdc67afb19ca89fbeeee)
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 	KASSERT(howmany(flowclen, 16) <= MAX_OFLD_TX_SDESC_CREDITS,
152 	    ("%s: tx_credits %u too large", __func__, howmany(flowclen, 16)));
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
update_tx_rate_limit(struct adapter * sc,struct toepcb * toep,u_int Bps)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 		KASSERT(flowclen16 <= MAX_OFLD_TX_SDESC_CREDITS,
221 		    ("%s: tx_credits %u too large", __func__, flowclen16));
222 		txsd->tx_credits = flowclen16;
223 		txsd->plen = 0;
224 		toep->tx_credits -= txsd->tx_credits;
225 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
226 			toep->txsd_pidx = 0;
227 		toep->txsd_avail--;
228 		t4_wrq_tx(sc, wr);
229 	}
230 
231 	if (toep->params.tc_idx >= 0)
232 		t4_release_cl_rl(sc, port_id, toep->params.tc_idx);
233 	toep->params.tc_idx = tc_idx;
234 
235 	return (0);
236 }
237 #endif
238 
239 void
send_reset(struct adapter * sc,struct toepcb * toep,uint32_t snd_nxt)240 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt)
241 {
242 	struct wrqe *wr;
243 	struct cpl_abort_req *req;
244 	int tid = toep->tid;
245 	struct inpcb *inp = toep->inp;
246 	struct tcpcb *tp = intotcpcb(inp);	/* don't use if INP_DROPPED */
247 
248 	INP_WLOCK_ASSERT(inp);
249 
250 	CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s",
251 	    __func__, toep->tid,
252 	    inp->inp_flags & INP_DROPPED ? "inp dropped" :
253 	    tcpstates[tp->t_state],
254 	    toep->flags, inp->inp_flags,
255 	    toep->flags & TPF_ABORT_SHUTDOWN ?
256 	    " (abort already in progress)" : "");
257 
258 	if (toep->flags & TPF_ABORT_SHUTDOWN)
259 		return;	/* abort already in progress */
260 
261 	toep->flags |= TPF_ABORT_SHUTDOWN;
262 
263 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
264 	    ("%s: flowc_wr not sent for tid %d.", __func__, tid));
265 
266 	wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq);
267 	if (wr == NULL) {
268 		/* XXX */
269 		panic("%s: allocation failure.", __func__);
270 	}
271 	req = wrtod(wr);
272 
273 	INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid);
274 	if (inp->inp_flags & INP_DROPPED)
275 		req->rsvd0 = htobe32(snd_nxt);
276 	else
277 		req->rsvd0 = htobe32(tp->snd_nxt);
278 	req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT);
279 	req->cmd = CPL_ABORT_SEND_RST;
280 
281 	/*
282 	 * XXX: What's the correct way to tell that the inp hasn't been detached
283 	 * from its socket?  Should I even be flushing the snd buffer here?
284 	 */
285 	if ((inp->inp_flags & INP_DROPPED) == 0) {
286 		struct socket *so = inp->inp_socket;
287 
288 		if (so != NULL)	/* because I'm not sure.  See comment above */
289 			sbflush(&so->so_snd);
290 	}
291 
292 	t4_l2t_send(sc, wr, toep->l2te);
293 }
294 
295 /*
296  * Called when a connection is established to translate the TCP options
297  * reported by HW to FreeBSD's native format.
298  */
299 static void
assign_rxopt(struct tcpcb * tp,uint16_t opt)300 assign_rxopt(struct tcpcb *tp, uint16_t opt)
301 {
302 	struct toepcb *toep = tp->t_toe;
303 	struct inpcb *inp = tptoinpcb(tp);
304 	struct adapter *sc = td_adapter(toep->td);
305 
306 	INP_LOCK_ASSERT(inp);
307 
308 	toep->params.mtu_idx = G_TCPOPT_MSS(opt);
309 	tp->t_maxseg = sc->params.mtus[toep->params.mtu_idx];
310 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
311 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
312 	else
313 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
314 
315 	toep->params.emss = tp->t_maxseg;
316 	if (G_TCPOPT_TSTAMP(opt)) {
317 		toep->params.tstamp = 1;
318 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
319 		tp->t_flags |= TF_RCVD_TSTMP;	/* timestamps ok */
320 		tp->ts_recent = 0;		/* hmmm */
321 		tp->ts_recent_age = tcp_ts_getticks();
322 	} else
323 		toep->params.tstamp = 0;
324 
325 	if (G_TCPOPT_SACK(opt)) {
326 		toep->params.sack = 1;
327 		tp->t_flags |= TF_SACK_PERMIT;	/* should already be set */
328 	} else {
329 		toep->params.sack = 0;
330 		tp->t_flags &= ~TF_SACK_PERMIT;	/* sack disallowed by peer */
331 	}
332 
333 	if (G_TCPOPT_WSCALE_OK(opt))
334 		tp->t_flags |= TF_RCVD_SCALE;
335 
336 	/* Doing window scaling? */
337 	if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) ==
338 	    (TF_RCVD_SCALE | TF_REQ_SCALE)) {
339 		tp->rcv_scale = tp->request_r_scale;
340 		tp->snd_scale = G_TCPOPT_SND_WSCALE(opt);
341 	} else
342 		toep->params.wscale = 0;
343 
344 	CTR6(KTR_CXGBE,
345 	    "assign_rxopt: tid %d, mtu_idx %u, emss %u, ts %u, sack %u, wscale %u",
346 	    toep->tid, toep->params.mtu_idx, toep->params.emss,
347 	    toep->params.tstamp, toep->params.sack, toep->params.wscale);
348 }
349 
350 /*
351  * Completes some final bits of initialization for just established connections
352  * and changes their state to TCPS_ESTABLISHED.
353  *
354  * The ISNs are from the exchange of SYNs.
355  */
356 void
make_established(struct toepcb * toep,uint32_t iss,uint32_t irs,uint16_t opt)357 make_established(struct toepcb *toep, uint32_t iss, uint32_t irs, uint16_t opt)
358 {
359 	struct inpcb *inp = toep->inp;
360 	struct socket *so = inp->inp_socket;
361 	struct tcpcb *tp = intotcpcb(inp);
362 	uint16_t tcpopt = be16toh(opt);
363 
364 	INP_WLOCK_ASSERT(inp);
365 	KASSERT(tp->t_state == TCPS_SYN_SENT ||
366 	    tp->t_state == TCPS_SYN_RECEIVED,
367 	    ("%s: TCP state %s", __func__, tcpstates[tp->t_state]));
368 
369 	CTR6(KTR_CXGBE, "%s: tid %d, so %p, inp %p, tp %p, toep %p",
370 	    __func__, toep->tid, so, inp, tp, toep);
371 
372 	tcp_state_change(tp, TCPS_ESTABLISHED);
373 	tp->t_starttime = ticks;
374 	TCPSTAT_INC(tcps_connects);
375 
376 	tp->irs = irs;
377 	tcp_rcvseqinit(tp);
378 	tp->rcv_wnd = (u_int)toep->params.opt0_bufsize << 10;
379 	tp->rcv_adv += tp->rcv_wnd;
380 	tp->last_ack_sent = tp->rcv_nxt;
381 
382 	tp->iss = iss;
383 	tcp_sendseqinit(tp);
384 	tp->snd_una = iss + 1;
385 	tp->snd_nxt = iss + 1;
386 	tp->snd_max = iss + 1;
387 
388 	assign_rxopt(tp, tcpopt);
389 	send_flowc_wr(toep, tp);
390 
391 	soisconnected(so);
392 }
393 
394 int
send_rx_credits(struct adapter * sc,struct toepcb * toep,int credits)395 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits)
396 {
397 	struct wrqe *wr;
398 	struct cpl_rx_data_ack *req;
399 	uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
400 
401 	KASSERT(credits >= 0, ("%s: %d credits", __func__, credits));
402 
403 	wr = alloc_wrqe(sizeof(*req), toep->ctrlq);
404 	if (wr == NULL)
405 		return (0);
406 	req = wrtod(wr);
407 
408 	INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid);
409 	req->credit_dack = htobe32(dack | V_RX_CREDITS(credits));
410 
411 	t4_wrq_tx(sc, wr);
412 	return (credits);
413 }
414 
415 void
t4_rcvd_locked(struct toedev * tod,struct tcpcb * tp)416 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp)
417 {
418 	struct adapter *sc = tod->tod_softc;
419 	struct inpcb *inp = tptoinpcb(tp);
420 	struct socket *so = inp->inp_socket;
421 	struct sockbuf *sb = &so->so_rcv;
422 	struct toepcb *toep = tp->t_toe;
423 	int rx_credits;
424 
425 	INP_WLOCK_ASSERT(inp);
426 	SOCKBUF_LOCK_ASSERT(sb);
427 
428 	rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
429 	if (rx_credits > 0 &&
430 	    (tp->rcv_wnd <= 32 * 1024 || rx_credits >= 64 * 1024 ||
431 	    (rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) ||
432 	    sbused(sb) + tp->rcv_wnd < sb->sb_lowat)) {
433 		rx_credits = send_rx_credits(sc, toep, rx_credits);
434 		tp->rcv_wnd += rx_credits;
435 		tp->rcv_adv += rx_credits;
436 	}
437 }
438 
439 void
t4_rcvd(struct toedev * tod,struct tcpcb * tp)440 t4_rcvd(struct toedev *tod, struct tcpcb *tp)
441 {
442 	struct inpcb *inp = tptoinpcb(tp);
443 	struct socket *so = inp->inp_socket;
444 	struct sockbuf *sb = &so->so_rcv;
445 
446 	SOCKBUF_LOCK(sb);
447 	t4_rcvd_locked(tod, tp);
448 	SOCKBUF_UNLOCK(sb);
449 }
450 
451 /*
452  * Close a connection by sending a CPL_CLOSE_CON_REQ message.
453  */
454 int
t4_close_conn(struct adapter * sc,struct toepcb * toep)455 t4_close_conn(struct adapter *sc, struct toepcb *toep)
456 {
457 	struct wrqe *wr;
458 	struct cpl_close_con_req *req;
459 	unsigned int tid = toep->tid;
460 
461 	CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid,
462 	    toep->flags & TPF_FIN_SENT ? ", IGNORED" : "");
463 
464 	if (toep->flags & TPF_FIN_SENT)
465 		return (0);
466 
467 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
468 	    ("%s: flowc_wr not sent for tid %u.", __func__, tid));
469 
470 	wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq);
471 	if (wr == NULL) {
472 		/* XXX */
473 		panic("%s: allocation failure.", __func__);
474 	}
475 	req = wrtod(wr);
476 
477         req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) |
478 	    V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr)));
479 	req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) |
480 	    V_FW_WR_FLOWID(tid));
481         req->wr.wr_lo = cpu_to_be64(0);
482         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
483 	req->rsvd = 0;
484 
485 	toep->flags |= TPF_FIN_SENT;
486 	toep->flags &= ~TPF_SEND_FIN;
487 	t4_l2t_send(sc, wr, toep->l2te);
488 
489 	return (0);
490 }
491 
492 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
493 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16))
494 #define MIN_ISO_TX_CREDITS  (howmany(sizeof(struct cpl_tx_data_iso), 16))
495 #define MIN_TX_CREDITS(iso)						\
496 	(MIN_OFLD_TX_CREDITS + ((iso) ? MIN_ISO_TX_CREDITS : 0))
497 
498 _Static_assert(MAX_OFLD_TX_CREDITS <= MAX_OFLD_TX_SDESC_CREDITS,
499     "MAX_OFLD_TX_SDESC_CREDITS too small");
500 
501 /* Maximum amount of immediate data we could stuff in a WR */
502 static inline int
max_imm_payload(int tx_credits,int iso)503 max_imm_payload(int tx_credits, int iso)
504 {
505 	const int iso_cpl_size = iso ? sizeof(struct cpl_tx_data_iso) : 0;
506 	const int n = 1;	/* Use no more than one desc for imm. data WR */
507 
508 	KASSERT(tx_credits >= 0 &&
509 		tx_credits <= MAX_OFLD_TX_CREDITS,
510 		("%s: %d credits", __func__, tx_credits));
511 
512 	if (tx_credits < MIN_TX_CREDITS(iso))
513 		return (0);
514 
515 	if (tx_credits >= (n * EQ_ESIZE) / 16)
516 		return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr) -
517 		    iso_cpl_size);
518 	else
519 		return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr) -
520 		    iso_cpl_size);
521 }
522 
523 /* Maximum number of SGL entries we could stuff in a WR */
524 static inline int
max_dsgl_nsegs(int tx_credits,int iso)525 max_dsgl_nsegs(int tx_credits, int iso)
526 {
527 	int nseg = 1;	/* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */
528 	int sge_pair_credits = tx_credits - MIN_TX_CREDITS(iso);
529 
530 	KASSERT(tx_credits >= 0 &&
531 		tx_credits <= MAX_OFLD_TX_CREDITS,
532 		("%s: %d credits", __func__, tx_credits));
533 
534 	if (tx_credits < MIN_TX_CREDITS(iso))
535 		return (0);
536 
537 	nseg += 2 * (sge_pair_credits * 16 / 24);
538 	if ((sge_pair_credits * 16) % 24 == 16)
539 		nseg++;
540 
541 	return (nseg);
542 }
543 
544 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)545 write_tx_wr(void *dst, struct toepcb *toep, int fw_wr_opcode,
546     unsigned int immdlen, unsigned int plen, uint8_t credits, int shove,
547     int ulp_submode)
548 {
549 	struct fw_ofld_tx_data_wr *txwr = dst;
550 
551 	txwr->op_to_immdlen = htobe32(V_WR_OP(fw_wr_opcode) |
552 	    V_FW_WR_IMMDLEN(immdlen));
553 	txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) |
554 	    V_FW_WR_LEN16(credits));
555 	txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ulp_mode(toep)) |
556 	    V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove));
557 	txwr->plen = htobe32(plen);
558 
559 	if (toep->params.tx_align > 0) {
560 		if (plen < 2 * toep->params.emss)
561 			txwr->lsodisable_to_flags |=
562 			    htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE);
563 		else
564 			txwr->lsodisable_to_flags |=
565 			    htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD |
566 				(toep->params.nagle == 0 ? 0 :
567 				F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE));
568 	}
569 }
570 
571 /*
572  * Generate a DSGL from a starting mbuf.  The total number of segments and the
573  * maximum segments in any one mbuf are provided.
574  */
575 static void
write_tx_sgl(void * dst,struct mbuf * start,struct mbuf * stop,int nsegs,int n)576 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n)
577 {
578 	struct mbuf *m;
579 	struct ulptx_sgl *usgl = dst;
580 	int i, j, rc;
581 	struct sglist sg;
582 	struct sglist_seg segs[n];
583 
584 	KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
585 
586 	sglist_init(&sg, n, segs);
587 	usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
588 	    V_ULPTX_NSGE(nsegs));
589 
590 	i = -1;
591 	for (m = start; m != stop; m = m->m_next) {
592 		if (m->m_flags & M_EXTPG)
593 			rc = sglist_append_mbuf_epg(&sg, m,
594 			    mtod(m, vm_offset_t), m->m_len);
595 		else
596 			rc = sglist_append(&sg, mtod(m, void *), m->m_len);
597 		if (__predict_false(rc != 0))
598 			panic("%s: sglist_append %d", __func__, rc);
599 
600 		for (j = 0; j < sg.sg_nseg; i++, j++) {
601 			if (i < 0) {
602 				usgl->len0 = htobe32(segs[j].ss_len);
603 				usgl->addr0 = htobe64(segs[j].ss_paddr);
604 			} else {
605 				usgl->sge[i / 2].len[i & 1] =
606 				    htobe32(segs[j].ss_len);
607 				usgl->sge[i / 2].addr[i & 1] =
608 				    htobe64(segs[j].ss_paddr);
609 			}
610 #ifdef INVARIANTS
611 			nsegs--;
612 #endif
613 		}
614 		sglist_reset(&sg);
615 	}
616 	if (i & 1)
617 		usgl->sge[i / 2].len[1] = htobe32(0);
618 	KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p",
619 	    __func__, nsegs, start, stop));
620 }
621 
622 bool
t4_push_raw_wr(struct adapter * sc,struct toepcb * toep,struct mbuf * m)623 t4_push_raw_wr(struct adapter *sc, struct toepcb *toep, struct mbuf *m)
624 {
625 #ifdef INVARIANTS
626 	struct inpcb *inp = toep->inp;
627 #endif
628 	struct wrqe *wr;
629 	struct ofld_tx_sdesc *txsd;
630 	u_int credits, plen;
631 
632 	INP_WLOCK_ASSERT(inp);
633 	MPASS(mbuf_raw_wr(m));
634 	plen = m->m_pkthdr.len;
635 	credits = howmany(plen, 16);
636 	if (credits > toep->tx_credits)
637 		return (false);
638 
639 	wr = alloc_wrqe(roundup2(plen, 16), &toep->ofld_txq->wrq);
640 	if (wr == NULL)
641 		return (false);
642 
643 	m_copydata(m, 0, plen, wrtod(wr));
644 	m_freem(m);
645 
646 	toep->tx_credits -= credits;
647 	if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
648 		toep->flags |= TPF_TX_SUSPENDED;
649 
650 	KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
651 	KASSERT(credits <= MAX_OFLD_TX_SDESC_CREDITS,
652 	    ("%s: tx_credits %u too large", __func__, credits));
653 	txsd = &toep->txsd[toep->txsd_pidx];
654 	txsd->plen = 0;
655 	txsd->tx_credits = credits;
656 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
657 		toep->txsd_pidx = 0;
658 	toep->txsd_avail--;
659 
660 	t4_wrq_tx(sc, wr);
661 	return (true);
662 }
663 
664 /*
665  * Max number of SGL entries an offload tx work request can have.  This is 41
666  * (1 + 40) for a full 512B work request.
667  * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40)
668  */
669 #define OFLD_SGL_LEN (41)
670 
671 /*
672  * Send data and/or a FIN to the peer.
673  *
674  * The socket's so_snd buffer consists of a stream of data starting with sb_mb
675  * and linked together with m_next.  sb_sndptr, if set, is the last mbuf that
676  * was transmitted.
677  *
678  * drop indicates the number of bytes that should be dropped from the head of
679  * the send buffer.  It is an optimization that lets do_fw4_ack avoid creating
680  * contention on the send buffer lock (before this change it used to do
681  * sowwakeup and then t4_push_frames right after that when recovering from tx
682  * stalls).  When drop is set this function MUST drop the bytes and wake up any
683  * writers.
684  */
685 static void
t4_push_frames(struct adapter * sc,struct toepcb * toep,int drop)686 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop)
687 {
688 	struct mbuf *sndptr, *m, *sb_sndptr;
689 	struct fw_ofld_tx_data_wr *txwr;
690 	struct wrqe *wr;
691 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
692 	struct inpcb *inp = toep->inp;
693 	struct tcpcb *tp = intotcpcb(inp);
694 	struct socket *so = inp->inp_socket;
695 	struct sockbuf *sb = &so->so_snd;
696 	struct mbufq *pduq = &toep->ulp_pduq;
697 	int tx_credits, shove, compl, sowwakeup;
698 	struct ofld_tx_sdesc *txsd;
699 	bool nomap_mbuf_seen;
700 
701 	INP_WLOCK_ASSERT(inp);
702 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
703 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
704 
705 	KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
706 	    ulp_mode(toep) == ULP_MODE_TCPDDP ||
707 	    ulp_mode(toep) == ULP_MODE_TLS ||
708 	    ulp_mode(toep) == ULP_MODE_RDMA,
709 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
710 
711 #ifdef VERBOSE_TRACES
712 	CTR5(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
713 	    __func__, toep->tid, toep->flags, tp->t_flags, drop);
714 #endif
715 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
716 		return;
717 
718 #ifdef RATELIMIT
719 	if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
720 	    (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
721 		inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
722 	}
723 #endif
724 
725 	/*
726 	 * This function doesn't resume by itself.  Someone else must clear the
727 	 * flag and call this function.
728 	 */
729 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
730 		KASSERT(drop == 0,
731 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
732 		return;
733 	}
734 
735 	txsd = &toep->txsd[toep->txsd_pidx];
736 	do {
737 		tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
738 		max_imm = max_imm_payload(tx_credits, 0);
739 		max_nsegs = max_dsgl_nsegs(tx_credits, 0);
740 
741 		if (__predict_false((sndptr = mbufq_first(pduq)) != NULL)) {
742 			if (!t4_push_raw_wr(sc, toep, sndptr)) {
743 				toep->flags |= TPF_TX_SUSPENDED;
744 				return;
745 			}
746 
747 			m = mbufq_dequeue(pduq);
748 			MPASS(m == sndptr);
749 
750 			txsd = &toep->txsd[toep->txsd_pidx];
751 			continue;
752 		}
753 
754 		SOCKBUF_LOCK(sb);
755 		sowwakeup = drop;
756 		if (drop) {
757 			sbdrop_locked(sb, drop);
758 			drop = 0;
759 		}
760 		sb_sndptr = sb->sb_sndptr;
761 		sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb;
762 		plen = 0;
763 		nsegs = 0;
764 		max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
765 		nomap_mbuf_seen = false;
766 		for (m = sndptr; m != NULL; m = m->m_next) {
767 			int n;
768 
769 			if ((m->m_flags & M_NOTREADY) != 0)
770 				break;
771 			if (plen + m->m_len > MAX_OFLD_TX_SDESC_PLEN)
772 				break;
773 			if (m->m_flags & M_EXTPG) {
774 #ifdef KERN_TLS
775 				if (m->m_epg_tls != NULL) {
776 					toep->flags |= TPF_KTLS;
777 					if (plen == 0) {
778 						SOCKBUF_UNLOCK(sb);
779 						t4_push_ktls(sc, toep, 0);
780 						return;
781 					}
782 					break;
783 				}
784 #endif
785 				n = sglist_count_mbuf_epg(m,
786 				    mtod(m, vm_offset_t), m->m_len);
787 			} else
788 				n = sglist_count(mtod(m, void *), m->m_len);
789 
790 			nsegs += n;
791 			plen += m->m_len;
792 
793 			/* This mbuf sent us _over_ the nsegs limit, back out */
794 			if (plen > max_imm && nsegs > max_nsegs) {
795 				nsegs -= n;
796 				plen -= m->m_len;
797 				if (plen == 0) {
798 					/* Too few credits */
799 					toep->flags |= TPF_TX_SUSPENDED;
800 					if (sowwakeup) {
801 						if (!TAILQ_EMPTY(
802 						    &toep->aiotx_jobq))
803 							t4_aiotx_queue_toep(so,
804 							    toep);
805 						sowwakeup_locked(so);
806 					} else
807 						SOCKBUF_UNLOCK(sb);
808 					SOCKBUF_UNLOCK_ASSERT(sb);
809 					return;
810 				}
811 				break;
812 			}
813 
814 			if (m->m_flags & M_EXTPG)
815 				nomap_mbuf_seen = true;
816 			if (max_nsegs_1mbuf < n)
817 				max_nsegs_1mbuf = n;
818 			sb_sndptr = m;	/* new sb->sb_sndptr if all goes well */
819 
820 			/* This mbuf put us right at the max_nsegs limit */
821 			if (plen > max_imm && nsegs == max_nsegs) {
822 				m = m->m_next;
823 				break;
824 			}
825 		}
826 
827 		if (sbused(sb) > sb->sb_hiwat * 5 / 8 &&
828 		    toep->plen_nocompl + plen >= sb->sb_hiwat / 4)
829 			compl = 1;
830 		else
831 			compl = 0;
832 
833 		if (sb->sb_flags & SB_AUTOSIZE &&
834 		    V_tcp_do_autosndbuf &&
835 		    sb->sb_hiwat < V_tcp_autosndbuf_max &&
836 		    sbused(sb) >= sb->sb_hiwat * 7 / 8) {
837 			int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
838 			    V_tcp_autosndbuf_max);
839 
840 			if (!sbreserve_locked(so, SO_SND, newsize, NULL))
841 				sb->sb_flags &= ~SB_AUTOSIZE;
842 			else
843 				sowwakeup = 1;	/* room available */
844 		}
845 		if (sowwakeup) {
846 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
847 				t4_aiotx_queue_toep(so, toep);
848 			sowwakeup_locked(so);
849 		} else
850 			SOCKBUF_UNLOCK(sb);
851 		SOCKBUF_UNLOCK_ASSERT(sb);
852 
853 		/* nothing to send */
854 		if (plen == 0) {
855 			KASSERT(m == NULL || (m->m_flags & M_NOTREADY) != 0,
856 			    ("%s: nothing to send, but m != NULL is ready",
857 			    __func__));
858 			break;
859 		}
860 
861 		if (__predict_false(toep->flags & TPF_FIN_SENT))
862 			panic("%s: excess tx.", __func__);
863 
864 		shove = m == NULL && !(tp->t_flags & TF_MORETOCOME);
865 		if (plen <= max_imm && !nomap_mbuf_seen) {
866 
867 			/* Immediate data tx */
868 
869 			wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16),
870 					&toep->ofld_txq->wrq);
871 			if (wr == NULL) {
872 				/* XXX: how will we recover from this? */
873 				toep->flags |= TPF_TX_SUSPENDED;
874 				return;
875 			}
876 			txwr = wrtod(wr);
877 			credits = howmany(wr->wr_len, 16);
878 			write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, plen, plen,
879 			    credits, shove, 0);
880 			m_copydata(sndptr, 0, plen, (void *)(txwr + 1));
881 			nsegs = 0;
882 		} else {
883 			int wr_len;
884 
885 			/* DSGL tx */
886 
887 			wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) +
888 			    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
889 			wr = alloc_wrqe(roundup2(wr_len, 16),
890 			    &toep->ofld_txq->wrq);
891 			if (wr == NULL) {
892 				/* XXX: how will we recover from this? */
893 				toep->flags |= TPF_TX_SUSPENDED;
894 				return;
895 			}
896 			txwr = wrtod(wr);
897 			credits = howmany(wr_len, 16);
898 			write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, 0, plen,
899 			    credits, shove, 0);
900 			write_tx_sgl(txwr + 1, sndptr, m, nsegs,
901 			    max_nsegs_1mbuf);
902 			if (wr_len & 0xf) {
903 				uint64_t *pad = (uint64_t *)
904 				    ((uintptr_t)txwr + wr_len);
905 				*pad = 0;
906 			}
907 		}
908 
909 		KASSERT(toep->tx_credits >= credits,
910 			("%s: not enough credits", __func__));
911 
912 		toep->tx_credits -= credits;
913 		toep->tx_nocompl += credits;
914 		toep->plen_nocompl += plen;
915 		if (toep->tx_credits <= toep->tx_total * 3 / 8 &&
916 		    toep->tx_nocompl >= toep->tx_total / 4)
917 			compl = 1;
918 
919 		if (compl || ulp_mode(toep) == ULP_MODE_RDMA) {
920 			txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL);
921 			toep->tx_nocompl = 0;
922 			toep->plen_nocompl = 0;
923 		}
924 
925 		tp->snd_nxt += plen;
926 		tp->snd_max += plen;
927 
928 		SOCKBUF_LOCK(sb);
929 		KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__));
930 		sb->sb_sndptr = sb_sndptr;
931 		SOCKBUF_UNLOCK(sb);
932 
933 		toep->flags |= TPF_TX_DATA_SENT;
934 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
935 			toep->flags |= TPF_TX_SUSPENDED;
936 
937 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
938 		KASSERT(plen <= MAX_OFLD_TX_SDESC_PLEN,
939 		    ("%s: plen %u too large", __func__, plen));
940 		txsd->plen = plen;
941 		txsd->tx_credits = credits;
942 		txsd++;
943 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
944 			toep->txsd_pidx = 0;
945 			txsd = &toep->txsd[0];
946 		}
947 		toep->txsd_avail--;
948 
949 		t4_l2t_send(sc, wr, toep->l2te);
950 	} while (m != NULL && (m->m_flags & M_NOTREADY) == 0);
951 
952 	/* Send a FIN if requested, but only if there's no more data to send */
953 	if (m == NULL && toep->flags & TPF_SEND_FIN)
954 		t4_close_conn(sc, toep);
955 }
956 
957 static inline void
rqdrop_locked(struct mbufq * q,int plen)958 rqdrop_locked(struct mbufq *q, int plen)
959 {
960 	struct mbuf *m;
961 
962 	while (plen > 0) {
963 		m = mbufq_dequeue(q);
964 
965 		/* Too many credits. */
966 		MPASS(m != NULL);
967 		M_ASSERTPKTHDR(m);
968 
969 		/* Partial credits. */
970 		MPASS(plen >= m->m_pkthdr.len);
971 
972 		plen -= m->m_pkthdr.len;
973 		m_freem(m);
974 	}
975 }
976 
977 /*
978  * Not a bit in the TCB, but is a bit in the ulp_submode field of the
979  * CPL_TX_DATA flags field in FW_ISCSI_TX_DATA_WR.
980  */
981 #define	ULP_ISO		G_TX_ULP_SUBMODE(F_FW_ISCSI_TX_DATA_WR_ULPSUBMODE_ISO)
982 
983 static void
write_tx_data_iso(void * dst,u_int ulp_submode,uint8_t flags,uint16_t mss,int len,int npdu)984 write_tx_data_iso(void *dst, u_int ulp_submode, uint8_t flags, uint16_t mss,
985     int len, int npdu)
986 {
987 	struct cpl_tx_data_iso *cpl;
988 	unsigned int burst_size;
989 	unsigned int last;
990 
991 	/*
992 	 * The firmware will set the 'F' bit on the last PDU when
993 	 * either condition is true:
994 	 *
995 	 * - this large PDU is marked as the "last" slice
996 	 *
997 	 * - the amount of data payload bytes equals the burst_size
998 	 *
999 	 * The strategy used here is to always set the burst_size
1000 	 * artificially high (len includes the size of the template
1001 	 * BHS) and only set the "last" flag if the original PDU had
1002 	 * 'F' set.
1003 	 */
1004 	burst_size = len;
1005 	last = !!(flags & CXGBE_ISO_F);
1006 
1007 	cpl = (struct cpl_tx_data_iso *)dst;
1008 	cpl->op_to_scsi = htonl(V_CPL_TX_DATA_ISO_OP(CPL_TX_DATA_ISO) |
1009 	    V_CPL_TX_DATA_ISO_FIRST(1) | V_CPL_TX_DATA_ISO_LAST(last) |
1010 	    V_CPL_TX_DATA_ISO_CPLHDRLEN(0) |
1011 	    V_CPL_TX_DATA_ISO_HDRCRC(!!(ulp_submode & ULP_CRC_HEADER)) |
1012 	    V_CPL_TX_DATA_ISO_PLDCRC(!!(ulp_submode & ULP_CRC_DATA)) |
1013 	    V_CPL_TX_DATA_ISO_IMMEDIATE(0) |
1014 	    V_CPL_TX_DATA_ISO_SCSI(CXGBE_ISO_TYPE(flags)));
1015 
1016 	cpl->ahs_len = 0;
1017 	cpl->mpdu = htons(DIV_ROUND_UP(mss, 4));
1018 	cpl->burst_size = htonl(DIV_ROUND_UP(burst_size, 4));
1019 	cpl->len = htonl(len);
1020 	cpl->reserved2_seglen_offset = htonl(0);
1021 	cpl->datasn_offset = htonl(0);
1022 	cpl->buffer_offset = htonl(0);
1023 	cpl->reserved3 = 0;
1024 }
1025 
1026 static struct wrqe *
write_iscsi_mbuf_wr(struct toepcb * toep,struct mbuf * sndptr)1027 write_iscsi_mbuf_wr(struct toepcb *toep, struct mbuf *sndptr)
1028 {
1029 	struct mbuf *m;
1030 	struct fw_ofld_tx_data_wr *txwr;
1031 	struct cpl_tx_data_iso *cpl_iso;
1032 	void *p;
1033 	struct wrqe *wr;
1034 	u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf;
1035 	u_int adjusted_plen, imm_data, ulp_submode;
1036 	struct inpcb *inp = toep->inp;
1037 	struct tcpcb *tp = intotcpcb(inp);
1038 	int tx_credits, shove, npdu, wr_len;
1039 	uint16_t iso_mss;
1040 	static const u_int ulp_extra_len[] = {0, 4, 4, 8};
1041 	bool iso, nomap_mbuf_seen;
1042 
1043 	M_ASSERTPKTHDR(sndptr);
1044 
1045 	tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
1046 	if (mbuf_raw_wr(sndptr)) {
1047 		plen = sndptr->m_pkthdr.len;
1048 		KASSERT(plen <= SGE_MAX_WR_LEN,
1049 		    ("raw WR len %u is greater than max WR len", plen));
1050 		if (plen > tx_credits * 16)
1051 			return (NULL);
1052 
1053 		wr = alloc_wrqe(roundup2(plen, 16), &toep->ofld_txq->wrq);
1054 		if (__predict_false(wr == NULL))
1055 			return (NULL);
1056 
1057 		m_copydata(sndptr, 0, plen, wrtod(wr));
1058 		return (wr);
1059 	}
1060 
1061 	iso = mbuf_iscsi_iso(sndptr);
1062 	max_imm = max_imm_payload(tx_credits, iso);
1063 	max_nsegs = max_dsgl_nsegs(tx_credits, iso);
1064 	iso_mss = mbuf_iscsi_iso_mss(sndptr);
1065 
1066 	plen = 0;
1067 	nsegs = 0;
1068 	max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
1069 	nomap_mbuf_seen = false;
1070 	for (m = sndptr; m != NULL; m = m->m_next) {
1071 		int n;
1072 
1073 		if (m->m_flags & M_EXTPG)
1074 			n = sglist_count_mbuf_epg(m, mtod(m, vm_offset_t),
1075 			    m->m_len);
1076 		else
1077 			n = sglist_count(mtod(m, void *), m->m_len);
1078 
1079 		nsegs += n;
1080 		plen += m->m_len;
1081 
1082 		/*
1083 		 * This mbuf would send us _over_ the nsegs limit.
1084 		 * Suspend tx because the PDU can't be sent out.
1085 		 */
1086 		if ((nomap_mbuf_seen || plen > max_imm) && nsegs > max_nsegs)
1087 			return (NULL);
1088 
1089 		if (m->m_flags & M_EXTPG)
1090 			nomap_mbuf_seen = true;
1091 		if (max_nsegs_1mbuf < n)
1092 			max_nsegs_1mbuf = n;
1093 	}
1094 
1095 	if (__predict_false(toep->flags & TPF_FIN_SENT))
1096 		panic("%s: excess tx.", __func__);
1097 
1098 	/*
1099 	 * We have a PDU to send.  All of it goes out in one WR so 'm'
1100 	 * is NULL.  A PDU's length is always a multiple of 4.
1101 	 */
1102 	MPASS(m == NULL);
1103 	MPASS((plen & 3) == 0);
1104 	MPASS(sndptr->m_pkthdr.len == plen);
1105 
1106 	shove = !(tp->t_flags & TF_MORETOCOME);
1107 
1108 	/*
1109 	 * plen doesn't include header and data digests, which are
1110 	 * generated and inserted in the right places by the TOE, but
1111 	 * they do occupy TCP sequence space and need to be accounted
1112 	 * for.
1113 	 */
1114 	ulp_submode = mbuf_ulp_submode(sndptr);
1115 	MPASS(ulp_submode < nitems(ulp_extra_len));
1116 	npdu = iso ? howmany(plen - ISCSI_BHS_SIZE, iso_mss) : 1;
1117 	adjusted_plen = plen + ulp_extra_len[ulp_submode] * npdu;
1118 	if (iso)
1119 		adjusted_plen += ISCSI_BHS_SIZE * (npdu - 1);
1120 	wr_len = sizeof(*txwr);
1121 	if (iso)
1122 		wr_len += sizeof(struct cpl_tx_data_iso);
1123 	if (plen <= max_imm && !nomap_mbuf_seen) {
1124 		/* Immediate data tx */
1125 		imm_data = plen;
1126 		wr_len += plen;
1127 		nsegs = 0;
1128 	} else {
1129 		/* DSGL tx */
1130 		imm_data = 0;
1131 		wr_len += sizeof(struct ulptx_sgl) +
1132 		    ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1133 	}
1134 
1135 	wr = alloc_wrqe(roundup2(wr_len, 16), &toep->ofld_txq->wrq);
1136 	if (wr == NULL) {
1137 		/* XXX: how will we recover from this? */
1138 		return (NULL);
1139 	}
1140 	txwr = wrtod(wr);
1141 	credits = howmany(wr->wr_len, 16);
1142 
1143 	if (iso) {
1144 		write_tx_wr(txwr, toep, FW_ISCSI_TX_DATA_WR,
1145 		    imm_data + sizeof(struct cpl_tx_data_iso),
1146 		    adjusted_plen, credits, shove, ulp_submode | ULP_ISO);
1147 		cpl_iso = (struct cpl_tx_data_iso *)(txwr + 1);
1148 		MPASS(plen == sndptr->m_pkthdr.len);
1149 		write_tx_data_iso(cpl_iso, ulp_submode,
1150 		    mbuf_iscsi_iso_flags(sndptr), iso_mss, plen, npdu);
1151 		p = cpl_iso + 1;
1152 	} else {
1153 		write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, imm_data,
1154 		    adjusted_plen, credits, shove, ulp_submode);
1155 		p = txwr + 1;
1156 	}
1157 
1158 	if (imm_data != 0) {
1159 		m_copydata(sndptr, 0, plen, p);
1160 	} else {
1161 		write_tx_sgl(p, sndptr, m, nsegs, max_nsegs_1mbuf);
1162 		if (wr_len & 0xf) {
1163 			uint64_t *pad = (uint64_t *)((uintptr_t)txwr + wr_len);
1164 			*pad = 0;
1165 		}
1166 	}
1167 
1168 	KASSERT(toep->tx_credits >= credits,
1169 	    ("%s: not enough credits: credits %u "
1170 		"toep->tx_credits %u tx_credits %u nsegs %u "
1171 		"max_nsegs %u iso %d", __func__, credits,
1172 		toep->tx_credits, tx_credits, nsegs, max_nsegs, iso));
1173 
1174 	tp->snd_nxt += adjusted_plen;
1175 	tp->snd_max += adjusted_plen;
1176 
1177 	counter_u64_add(toep->ofld_txq->tx_iscsi_pdus, npdu);
1178 	counter_u64_add(toep->ofld_txq->tx_iscsi_octets, plen);
1179 	if (iso)
1180 		counter_u64_add(toep->ofld_txq->tx_iscsi_iso_wrs, 1);
1181 
1182 	return (wr);
1183 }
1184 
1185 void
t4_push_pdus(struct adapter * sc,struct toepcb * toep,int drop)1186 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop)
1187 {
1188 	struct mbuf *sndptr, *m;
1189 	struct fw_wr_hdr *wrhdr;
1190 	struct wrqe *wr;
1191 	u_int plen, credits;
1192 	struct inpcb *inp = toep->inp;
1193 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
1194 	struct mbufq *pduq = &toep->ulp_pduq;
1195 
1196 	INP_WLOCK_ASSERT(inp);
1197 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1198 	    ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1199 	KASSERT(ulp_mode(toep) == ULP_MODE_ISCSI,
1200 	    ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1201 
1202 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1203 		return;
1204 
1205 	/*
1206 	 * This function doesn't resume by itself.  Someone else must clear the
1207 	 * flag and call this function.
1208 	 */
1209 	if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1210 		KASSERT(drop == 0,
1211 		    ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1212 		return;
1213 	}
1214 
1215 	if (drop) {
1216 		struct socket *so = inp->inp_socket;
1217 		struct sockbuf *sb = &so->so_snd;
1218 		int sbu;
1219 
1220 		/*
1221 		 * An unlocked read is ok here as the data should only
1222 		 * transition from a non-zero value to either another
1223 		 * non-zero value or zero.  Once it is zero it should
1224 		 * stay zero.
1225 		 */
1226 		if (__predict_false(sbused(sb)) > 0) {
1227 			SOCKBUF_LOCK(sb);
1228 			sbu = sbused(sb);
1229 			if (sbu > 0) {
1230 				/*
1231 				 * The data transmitted before the
1232 				 * tid's ULP mode changed to ISCSI is
1233 				 * still in so_snd.  Incoming credits
1234 				 * should account for so_snd first.
1235 				 */
1236 				sbdrop_locked(sb, min(sbu, drop));
1237 				drop -= min(sbu, drop);
1238 			}
1239 			sowwakeup_locked(so);	/* unlocks so_snd */
1240 		}
1241 		rqdrop_locked(&toep->ulp_pdu_reclaimq, drop);
1242 	}
1243 
1244 	while ((sndptr = mbufq_first(pduq)) != NULL) {
1245 		wr = write_iscsi_mbuf_wr(toep, sndptr);
1246 		if (wr == NULL) {
1247 			toep->flags |= TPF_TX_SUSPENDED;
1248 			return;
1249 		}
1250 
1251 		plen = sndptr->m_pkthdr.len;
1252 		credits = howmany(wr->wr_len, 16);
1253 		KASSERT(toep->tx_credits >= credits,
1254 			("%s: not enough credits", __func__));
1255 
1256 		m = mbufq_dequeue(pduq);
1257 		MPASS(m == sndptr);
1258 		mbufq_enqueue(&toep->ulp_pdu_reclaimq, m);
1259 
1260 		toep->tx_credits -= credits;
1261 		toep->tx_nocompl += credits;
1262 		toep->plen_nocompl += plen;
1263 
1264 		/*
1265 		 * Ensure there are enough credits for a full-sized WR
1266 		 * as page pod WRs can be full-sized.
1267 		 */
1268 		if (toep->tx_credits <= SGE_MAX_WR_LEN * 5 / 4 &&
1269 		    toep->tx_nocompl >= toep->tx_total / 4) {
1270 			wrhdr = wrtod(wr);
1271 			wrhdr->hi |= htobe32(F_FW_WR_COMPL);
1272 			toep->tx_nocompl = 0;
1273 			toep->plen_nocompl = 0;
1274 		}
1275 
1276 		toep->flags |= TPF_TX_DATA_SENT;
1277 		if (toep->tx_credits < MIN_OFLD_TX_CREDITS)
1278 			toep->flags |= TPF_TX_SUSPENDED;
1279 
1280 		KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1281 		KASSERT(plen <= MAX_OFLD_TX_SDESC_PLEN,
1282 		    ("%s: plen %u too large", __func__, plen));
1283 		txsd->plen = plen;
1284 		txsd->tx_credits = credits;
1285 		txsd++;
1286 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1287 			toep->txsd_pidx = 0;
1288 			txsd = &toep->txsd[0];
1289 		}
1290 		toep->txsd_avail--;
1291 
1292 		t4_l2t_send(sc, wr, toep->l2te);
1293 	}
1294 
1295 	/* Send a FIN if requested, but only if there are no more PDUs to send */
1296 	if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN)
1297 		t4_close_conn(sc, toep);
1298 }
1299 
1300 static inline void
t4_push_data(struct adapter * sc,struct toepcb * toep,int drop)1301 t4_push_data(struct adapter *sc, struct toepcb *toep, int drop)
1302 {
1303 
1304 	if (ulp_mode(toep) == ULP_MODE_ISCSI)
1305 		t4_push_pdus(sc, toep, drop);
1306 	else if (toep->flags & TPF_KTLS)
1307 		t4_push_ktls(sc, toep, drop);
1308 	else
1309 		t4_push_frames(sc, toep, drop);
1310 }
1311 
1312 void
t4_raw_wr_tx(struct adapter * sc,struct toepcb * toep,struct mbuf * m)1313 t4_raw_wr_tx(struct adapter *sc, struct toepcb *toep, struct mbuf *m)
1314 {
1315 #ifdef INVARIANTS
1316 	struct inpcb *inp = toep->inp;
1317 #endif
1318 
1319 	INP_WLOCK_ASSERT(inp);
1320 
1321 	/*
1322 	 * If there are other raw WRs enqueued, enqueue to preserve
1323 	 * FIFO ordering.
1324 	 */
1325 	if (!mbufq_empty(&toep->ulp_pduq)) {
1326 		mbufq_enqueue(&toep->ulp_pduq, m);
1327 		return;
1328 	}
1329 
1330 	/*
1331 	 * Cannot call t4_push_data here as that will lock so_snd and
1332 	 * some callers of this run in rx handlers with so_rcv locked.
1333 	 * Instead, just try to transmit this WR.
1334 	 */
1335 	if (!t4_push_raw_wr(sc, toep, m)) {
1336 		mbufq_enqueue(&toep->ulp_pduq, m);
1337 		toep->flags |= TPF_TX_SUSPENDED;
1338 	}
1339 }
1340 
1341 int
t4_tod_output(struct toedev * tod,struct tcpcb * tp)1342 t4_tod_output(struct toedev *tod, struct tcpcb *tp)
1343 {
1344 	struct adapter *sc = tod->tod_softc;
1345 #ifdef INVARIANTS
1346 	struct inpcb *inp = tptoinpcb(tp);
1347 #endif
1348 	struct toepcb *toep = tp->t_toe;
1349 
1350 	INP_WLOCK_ASSERT(inp);
1351 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1352 	    ("%s: inp %p dropped.", __func__, inp));
1353 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1354 
1355 	t4_push_data(sc, toep, 0);
1356 
1357 	return (0);
1358 }
1359 
1360 int
t4_send_fin(struct toedev * tod,struct tcpcb * tp)1361 t4_send_fin(struct toedev *tod, struct tcpcb *tp)
1362 {
1363 	struct adapter *sc = tod->tod_softc;
1364 #ifdef INVARIANTS
1365 	struct inpcb *inp = tptoinpcb(tp);
1366 #endif
1367 	struct toepcb *toep = tp->t_toe;
1368 
1369 	INP_WLOCK_ASSERT(inp);
1370 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1371 	    ("%s: inp %p dropped.", __func__, inp));
1372 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1373 
1374 	toep->flags |= TPF_SEND_FIN;
1375 	if (tp->t_state >= TCPS_ESTABLISHED)
1376 		t4_push_data(sc, toep, 0);
1377 
1378 	return (0);
1379 }
1380 
1381 int
t4_send_rst(struct toedev * tod,struct tcpcb * tp)1382 t4_send_rst(struct toedev *tod, struct tcpcb *tp)
1383 {
1384 	struct adapter *sc = tod->tod_softc;
1385 #if defined(INVARIANTS)
1386 	struct inpcb *inp = tptoinpcb(tp);
1387 #endif
1388 	struct toepcb *toep = tp->t_toe;
1389 
1390 	INP_WLOCK_ASSERT(inp);
1391 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1392 	    ("%s: inp %p dropped.", __func__, inp));
1393 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
1394 
1395 	/* hmmmm */
1396 	KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1397 	    ("%s: flowc for tid %u [%s] not sent already",
1398 	    __func__, toep->tid, tcpstates[tp->t_state]));
1399 
1400 	send_reset(sc, toep, 0);
1401 	return (0);
1402 }
1403 
1404 /*
1405  * Peer has sent us a FIN.
1406  */
1407 static int
do_peer_close(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)1408 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1409 {
1410 	struct adapter *sc = iq->adapter;
1411 	const struct cpl_peer_close *cpl = (const void *)(rss + 1);
1412 	unsigned int tid = GET_TID(cpl);
1413 	struct toepcb *toep = lookup_tid(sc, tid);
1414 	struct inpcb *inp = toep->inp;
1415 	struct tcpcb *tp = NULL;
1416 	struct socket *so;
1417 	struct epoch_tracker et;
1418 #ifdef INVARIANTS
1419 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1420 #endif
1421 
1422 	KASSERT(opcode == CPL_PEER_CLOSE,
1423 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1424 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1425 
1426 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1427 		/*
1428 		 * do_pass_establish must have run before do_peer_close and if
1429 		 * this is still a synqe instead of a toepcb then the connection
1430 		 * must be getting aborted.
1431 		 */
1432 		MPASS(toep->flags & TPF_ABORT_SHUTDOWN);
1433 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1434 		    toep, toep->flags);
1435 		return (0);
1436 	}
1437 
1438 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1439 
1440 	CURVNET_SET(toep->vnet);
1441 	NET_EPOCH_ENTER(et);
1442 	INP_WLOCK(inp);
1443 	tp = intotcpcb(inp);
1444 
1445 	CTR6(KTR_CXGBE,
1446 	    "%s: tid %u (%s), toep_flags 0x%x, ddp_flags 0x%x, inp %p",
1447 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1448 	    toep->ddp.flags, inp);
1449 
1450 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1451 		goto done;
1452 
1453 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1454 		DDP_LOCK(toep);
1455 		if (__predict_false(toep->ddp.flags &
1456 		    (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)))
1457 			handle_ddp_close(toep, tp, cpl->rcv_nxt);
1458 		DDP_UNLOCK(toep);
1459 	}
1460 	so = inp->inp_socket;
1461 	socantrcvmore(so);
1462 
1463 	if (ulp_mode(toep) == ULP_MODE_RDMA ||
1464 	    (ulp_mode(toep) == ULP_MODE_ISCSI && chip_id(sc) >= CHELSIO_T6)) {
1465 		/*
1466 		 * There might be data received via DDP before the FIN
1467 		 * not reported to the driver.  Just assume the
1468 		 * sequence number in the CPL is correct as the
1469 		 * sequence number of the FIN.
1470 		 */
1471 	} else {
1472 		KASSERT(tp->rcv_nxt + 1 == be32toh(cpl->rcv_nxt),
1473 		    ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt,
1474 		    be32toh(cpl->rcv_nxt)));
1475 	}
1476 
1477 	tp->rcv_nxt = be32toh(cpl->rcv_nxt);
1478 
1479 	switch (tp->t_state) {
1480 	case TCPS_SYN_RECEIVED:
1481 		tp->t_starttime = ticks;
1482 		/* FALLTHROUGH */
1483 
1484 	case TCPS_ESTABLISHED:
1485 		tcp_state_change(tp, TCPS_CLOSE_WAIT);
1486 		break;
1487 
1488 	case TCPS_FIN_WAIT_1:
1489 		tcp_state_change(tp, TCPS_CLOSING);
1490 		break;
1491 
1492 	case TCPS_FIN_WAIT_2:
1493 		restore_so_proto(so, inp->inp_vflag & INP_IPV6);
1494 		t4_pcb_detach(NULL, tp);
1495 		tcp_twstart(tp);
1496 		INP_UNLOCK_ASSERT(inp);	 /* safe, we have a ref on the inp */
1497 		NET_EPOCH_EXIT(et);
1498 		CURVNET_RESTORE();
1499 
1500 		INP_WLOCK(inp);
1501 		final_cpl_received(toep);
1502 		return (0);
1503 
1504 	default:
1505 		log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n",
1506 		    __func__, tid, tp->t_state);
1507 	}
1508 done:
1509 	INP_WUNLOCK(inp);
1510 	NET_EPOCH_EXIT(et);
1511 	CURVNET_RESTORE();
1512 	return (0);
1513 }
1514 
1515 /*
1516  * Peer has ACK'd our FIN.
1517  */
1518 static int
do_close_con_rpl(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)1519 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss,
1520     struct mbuf *m)
1521 {
1522 	struct adapter *sc = iq->adapter;
1523 	const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1);
1524 	unsigned int tid = GET_TID(cpl);
1525 	struct toepcb *toep = lookup_tid(sc, tid);
1526 	struct inpcb *inp = toep->inp;
1527 	struct tcpcb *tp = NULL;
1528 	struct socket *so = NULL;
1529 	struct epoch_tracker et;
1530 #ifdef INVARIANTS
1531 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1532 #endif
1533 
1534 	KASSERT(opcode == CPL_CLOSE_CON_RPL,
1535 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1536 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1537 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1538 
1539 	CURVNET_SET(toep->vnet);
1540 	NET_EPOCH_ENTER(et);
1541 	INP_WLOCK(inp);
1542 	tp = intotcpcb(inp);
1543 
1544 	CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x",
1545 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags);
1546 
1547 	if (toep->flags & TPF_ABORT_SHUTDOWN)
1548 		goto done;
1549 
1550 	so = inp->inp_socket;
1551 	tp->snd_una = be32toh(cpl->snd_nxt) - 1;	/* exclude FIN */
1552 
1553 	switch (tp->t_state) {
1554 	case TCPS_CLOSING:	/* see TCPS_FIN_WAIT_2 in do_peer_close too */
1555 		restore_so_proto(so, inp->inp_vflag & INP_IPV6);
1556 		t4_pcb_detach(NULL, tp);
1557 		tcp_twstart(tp);
1558 release:
1559 		INP_UNLOCK_ASSERT(inp);	/* safe, we have a ref on the  inp */
1560 		NET_EPOCH_EXIT(et);
1561 		CURVNET_RESTORE();
1562 
1563 		INP_WLOCK(inp);
1564 		final_cpl_received(toep);	/* no more CPLs expected */
1565 
1566 		return (0);
1567 	case TCPS_LAST_ACK:
1568 		if (tcp_close(tp))
1569 			INP_WUNLOCK(inp);
1570 		goto release;
1571 
1572 	case TCPS_FIN_WAIT_1:
1573 		if (so->so_rcv.sb_state & SBS_CANTRCVMORE)
1574 			soisdisconnected(so);
1575 		tcp_state_change(tp, TCPS_FIN_WAIT_2);
1576 		break;
1577 
1578 	default:
1579 		log(LOG_ERR,
1580 		    "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n",
1581 		    __func__, tid, tcpstates[tp->t_state]);
1582 	}
1583 done:
1584 	INP_WUNLOCK(inp);
1585 	NET_EPOCH_EXIT(et);
1586 	CURVNET_RESTORE();
1587 	return (0);
1588 }
1589 
1590 void
send_abort_rpl(struct adapter * sc,struct sge_ofld_txq * ofld_txq,int tid,int rst_status)1591 send_abort_rpl(struct adapter *sc, struct sge_ofld_txq *ofld_txq, int tid,
1592     int rst_status)
1593 {
1594 	struct wrqe *wr;
1595 	struct cpl_abort_rpl *cpl;
1596 
1597 	wr = alloc_wrqe(sizeof(*cpl), &ofld_txq->wrq);
1598 	if (wr == NULL) {
1599 		/* XXX */
1600 		panic("%s: allocation failure.", __func__);
1601 	}
1602 	cpl = wrtod(wr);
1603 
1604 	INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid);
1605 	cpl->cmd = rst_status;
1606 
1607 	t4_wrq_tx(sc, wr);
1608 }
1609 
1610 static int
abort_status_to_errno(struct tcpcb * tp,unsigned int abort_reason)1611 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason)
1612 {
1613 	switch (abort_reason) {
1614 	case CPL_ERR_BAD_SYN:
1615 	case CPL_ERR_CONN_RESET:
1616 		return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET);
1617 	case CPL_ERR_XMIT_TIMEDOUT:
1618 	case CPL_ERR_PERSIST_TIMEDOUT:
1619 	case CPL_ERR_FINWAIT2_TIMEDOUT:
1620 	case CPL_ERR_KEEPALIVE_TIMEDOUT:
1621 		return (ETIMEDOUT);
1622 	default:
1623 		return (EIO);
1624 	}
1625 }
1626 
1627 /*
1628  * TCP RST from the peer, timeout, or some other such critical error.
1629  */
1630 static int
do_abort_req(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)1631 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1632 {
1633 	struct adapter *sc = iq->adapter;
1634 	const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1);
1635 	unsigned int tid = GET_TID(cpl);
1636 	struct toepcb *toep = lookup_tid(sc, tid);
1637 	struct sge_ofld_txq *ofld_txq = toep->ofld_txq;
1638 	struct inpcb *inp;
1639 	struct tcpcb *tp;
1640 	struct epoch_tracker et;
1641 #ifdef INVARIANTS
1642 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1643 #endif
1644 
1645 	KASSERT(opcode == CPL_ABORT_REQ_RSS,
1646 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1647 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1648 
1649 	if (toep->flags & TPF_SYNQE)
1650 		return (do_abort_req_synqe(iq, rss, m));
1651 
1652 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1653 
1654 	if (negative_advice(cpl->status)) {
1655 		CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)",
1656 		    __func__, cpl->status, tid, toep->flags);
1657 		return (0);	/* Ignore negative advice */
1658 	}
1659 
1660 	inp = toep->inp;
1661 	CURVNET_SET(toep->vnet);
1662 	NET_EPOCH_ENTER(et);	/* for tcp_close */
1663 	INP_WLOCK(inp);
1664 
1665 	tp = intotcpcb(inp);
1666 
1667 	CTR6(KTR_CXGBE,
1668 	    "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d",
1669 	    __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags,
1670 	    inp->inp_flags, cpl->status);
1671 
1672 	/*
1673 	 * If we'd initiated an abort earlier the reply to it is responsible for
1674 	 * cleaning up resources.  Otherwise we tear everything down right here
1675 	 * right now.  We owe the T4 a CPL_ABORT_RPL no matter what.
1676 	 */
1677 	if (toep->flags & TPF_ABORT_SHUTDOWN) {
1678 		INP_WUNLOCK(inp);
1679 		goto done;
1680 	}
1681 	toep->flags |= TPF_ABORT_SHUTDOWN;
1682 
1683 	if ((inp->inp_flags & INP_DROPPED) == 0) {
1684 		struct socket *so = inp->inp_socket;
1685 
1686 		if (so != NULL)
1687 			so_error_set(so, abort_status_to_errno(tp,
1688 			    cpl->status));
1689 		tp = tcp_close(tp);
1690 		if (tp == NULL)
1691 			INP_WLOCK(inp);	/* re-acquire */
1692 	}
1693 
1694 	final_cpl_received(toep);
1695 done:
1696 	NET_EPOCH_EXIT(et);
1697 	CURVNET_RESTORE();
1698 	send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST);
1699 	return (0);
1700 }
1701 
1702 /*
1703  * Reply to the CPL_ABORT_REQ (send_reset)
1704  */
1705 static int
do_abort_rpl(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)1706 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1707 {
1708 	struct adapter *sc = iq->adapter;
1709 	const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1);
1710 	unsigned int tid = GET_TID(cpl);
1711 	struct toepcb *toep = lookup_tid(sc, tid);
1712 	struct inpcb *inp = toep->inp;
1713 #ifdef INVARIANTS
1714 	unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl)));
1715 #endif
1716 
1717 	KASSERT(opcode == CPL_ABORT_RPL_RSS,
1718 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1719 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1720 
1721 	if (toep->flags & TPF_SYNQE)
1722 		return (do_abort_rpl_synqe(iq, rss, m));
1723 
1724 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1725 
1726 	CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d",
1727 	    __func__, tid, toep, inp, cpl->status);
1728 
1729 	KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1730 	    ("%s: wasn't expecting abort reply", __func__));
1731 
1732 	INP_WLOCK(inp);
1733 	final_cpl_received(toep);
1734 
1735 	return (0);
1736 }
1737 
1738 static int
do_rx_data(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)1739 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1740 {
1741 	struct adapter *sc = iq->adapter;
1742 	const struct cpl_rx_data *cpl = mtod(m, const void *);
1743 	unsigned int tid = GET_TID(cpl);
1744 	struct toepcb *toep = lookup_tid(sc, tid);
1745 	struct inpcb *inp = toep->inp;
1746 	struct tcpcb *tp;
1747 	struct socket *so;
1748 	struct sockbuf *sb;
1749 	struct epoch_tracker et;
1750 	int len;
1751 	uint32_t ddp_placed = 0;
1752 
1753 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1754 		/*
1755 		 * do_pass_establish must have run before do_rx_data and if this
1756 		 * is still a synqe instead of a toepcb then the connection must
1757 		 * be getting aborted.
1758 		 */
1759 		MPASS(toep->flags & TPF_ABORT_SHUTDOWN);
1760 		CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid,
1761 		    toep, toep->flags);
1762 		m_freem(m);
1763 		return (0);
1764 	}
1765 
1766 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1767 
1768 	/* strip off CPL header */
1769 	m_adj(m, sizeof(*cpl));
1770 	len = m->m_pkthdr.len;
1771 
1772 	INP_WLOCK(inp);
1773 	if (inp->inp_flags & INP_DROPPED) {
1774 		CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1775 		    __func__, tid, len, inp->inp_flags);
1776 		INP_WUNLOCK(inp);
1777 		m_freem(m);
1778 		return (0);
1779 	}
1780 
1781 	tp = intotcpcb(inp);
1782 
1783 	if (__predict_false(ulp_mode(toep) == ULP_MODE_TLS &&
1784 	   toep->flags & TPF_TLS_RECEIVE)) {
1785 		/* Received "raw" data on a TLS socket. */
1786 		CTR3(KTR_CXGBE, "%s: tid %u, raw TLS data (%d bytes)",
1787 		    __func__, tid, len);
1788 		do_rx_data_tls(cpl, toep, m);
1789 		return (0);
1790 	}
1791 
1792 	if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq)))
1793 		ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt;
1794 
1795 	tp->rcv_nxt += len;
1796 	if (tp->rcv_wnd < len) {
1797 		KASSERT(ulp_mode(toep) == ULP_MODE_RDMA,
1798 				("%s: negative window size", __func__));
1799 	}
1800 
1801 	tp->rcv_wnd -= len;
1802 	tp->t_rcvtime = ticks;
1803 
1804 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1805 		DDP_LOCK(toep);
1806 	so = inp_inpcbtosocket(inp);
1807 	sb = &so->so_rcv;
1808 	SOCKBUF_LOCK(sb);
1809 
1810 	if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
1811 		CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
1812 		    __func__, tid, len);
1813 		m_freem(m);
1814 		SOCKBUF_UNLOCK(sb);
1815 		if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1816 			DDP_UNLOCK(toep);
1817 		INP_WUNLOCK(inp);
1818 
1819 		CURVNET_SET(toep->vnet);
1820 		NET_EPOCH_ENTER(et);
1821 		INP_WLOCK(inp);
1822 		tp = tcp_drop(tp, ECONNRESET);
1823 		if (tp)
1824 			INP_WUNLOCK(inp);
1825 		NET_EPOCH_EXIT(et);
1826 		CURVNET_RESTORE();
1827 
1828 		return (0);
1829 	}
1830 
1831 	/* receive buffer autosize */
1832 	MPASS(toep->vnet == so->so_vnet);
1833 	CURVNET_SET(toep->vnet);
1834 	if (sb->sb_flags & SB_AUTOSIZE &&
1835 	    V_tcp_do_autorcvbuf &&
1836 	    sb->sb_hiwat < V_tcp_autorcvbuf_max &&
1837 	    len > (sbspace(sb) / 8 * 7)) {
1838 		unsigned int hiwat = sb->sb_hiwat;
1839 		unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
1840 		    V_tcp_autorcvbuf_max);
1841 
1842 		if (!sbreserve_locked(so, SO_RCV, newsize, NULL))
1843 			sb->sb_flags &= ~SB_AUTOSIZE;
1844 	}
1845 
1846 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1847 		int changed = !(toep->ddp.flags & DDP_ON) ^ cpl->ddp_off;
1848 
1849 		if (toep->ddp.waiting_count != 0 || toep->ddp.active_count != 0)
1850 			CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)",
1851 			    __func__, tid, len);
1852 
1853 		if (changed) {
1854 			if (toep->ddp.flags & DDP_SC_REQ)
1855 				toep->ddp.flags ^= DDP_ON | DDP_SC_REQ;
1856 			else if (cpl->ddp_off == 1) {
1857 				/* Fell out of DDP mode */
1858 				toep->ddp.flags &= ~DDP_ON;
1859 				CTR1(KTR_CXGBE, "%s: fell out of DDP mode",
1860 				    __func__);
1861 
1862 				insert_ddp_data(toep, ddp_placed);
1863 			} else {
1864 				/*
1865 				 * Data was received while still
1866 				 * ULP_MODE_NONE, just fall through.
1867 				 */
1868 			}
1869 		}
1870 
1871 		if (toep->ddp.flags & DDP_ON) {
1872 			/*
1873 			 * CPL_RX_DATA with DDP on can only be an indicate.
1874 			 * Start posting queued AIO requests via DDP.  The
1875 			 * payload that arrived in this indicate is appended
1876 			 * to the socket buffer as usual.
1877 			 */
1878 			handle_ddp_indicate(toep);
1879 		}
1880 	}
1881 
1882 	sbappendstream_locked(sb, m, 0);
1883 	t4_rcvd_locked(&toep->td->tod, tp);
1884 
1885 	if (ulp_mode(toep) == ULP_MODE_TCPDDP &&
1886 	    (toep->ddp.flags & DDP_AIO) != 0 && toep->ddp.waiting_count > 0 &&
1887 	    sbavail(sb) != 0) {
1888 		CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__,
1889 		    tid);
1890 		ddp_queue_toep(toep);
1891 	}
1892 	if (toep->flags & TPF_TLS_STARTING)
1893 		tls_received_starting_data(sc, toep, sb, len);
1894 	sorwakeup_locked(so);
1895 	SOCKBUF_UNLOCK_ASSERT(sb);
1896 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1897 		DDP_UNLOCK(toep);
1898 
1899 	INP_WUNLOCK(inp);
1900 	CURVNET_RESTORE();
1901 	return (0);
1902 }
1903 
1904 static int
do_fw4_ack(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)1905 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1906 {
1907 	struct adapter *sc = iq->adapter;
1908 	const struct cpl_fw4_ack *cpl = (const void *)(rss + 1);
1909 	unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl)));
1910 	struct toepcb *toep = lookup_tid(sc, tid);
1911 	struct inpcb *inp;
1912 	struct tcpcb *tp;
1913 	struct socket *so;
1914 	uint8_t credits = cpl->credits;
1915 	struct ofld_tx_sdesc *txsd;
1916 	int plen;
1917 #ifdef INVARIANTS
1918 	unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl)));
1919 #endif
1920 
1921 	/*
1922 	 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and
1923 	 * now this comes back carrying the credits for the flowc.
1924 	 */
1925 	if (__predict_false(toep->flags & TPF_SYNQE)) {
1926 		KASSERT(toep->flags & TPF_ABORT_SHUTDOWN,
1927 		    ("%s: credits for a synq entry %p", __func__, toep));
1928 		return (0);
1929 	}
1930 
1931 	inp = toep->inp;
1932 
1933 	KASSERT(opcode == CPL_FW4_ACK,
1934 	    ("%s: unexpected opcode 0x%x", __func__, opcode));
1935 	KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__));
1936 	KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__));
1937 
1938 	INP_WLOCK(inp);
1939 
1940 	if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) {
1941 		INP_WUNLOCK(inp);
1942 		return (0);
1943 	}
1944 
1945 	KASSERT((inp->inp_flags & INP_DROPPED) == 0,
1946 	    ("%s: inp_flags 0x%x", __func__, inp->inp_flags));
1947 
1948 	tp = intotcpcb(inp);
1949 
1950 	if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) {
1951 		tcp_seq snd_una = be32toh(cpl->snd_una);
1952 
1953 #ifdef INVARIANTS
1954 		if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) {
1955 			log(LOG_ERR,
1956 			    "%s: unexpected seq# %x for TID %u, snd_una %x\n",
1957 			    __func__, snd_una, toep->tid, tp->snd_una);
1958 		}
1959 #endif
1960 
1961 		if (tp->snd_una != snd_una) {
1962 			tp->snd_una = snd_una;
1963 			tp->ts_recent_age = tcp_ts_getticks();
1964 		}
1965 	}
1966 
1967 #ifdef VERBOSE_TRACES
1968 	CTR3(KTR_CXGBE, "%s: tid %d credits %u", __func__, tid, credits);
1969 #endif
1970 	so = inp->inp_socket;
1971 	txsd = &toep->txsd[toep->txsd_cidx];
1972 	plen = 0;
1973 	while (credits) {
1974 		KASSERT(credits >= txsd->tx_credits,
1975 		    ("%s: too many (or partial) credits", __func__));
1976 		credits -= txsd->tx_credits;
1977 		toep->tx_credits += txsd->tx_credits;
1978 		plen += txsd->plen;
1979 		txsd++;
1980 		toep->txsd_avail++;
1981 		KASSERT(toep->txsd_avail <= toep->txsd_total,
1982 		    ("%s: txsd avail > total", __func__));
1983 		if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) {
1984 			txsd = &toep->txsd[0];
1985 			toep->txsd_cidx = 0;
1986 		}
1987 	}
1988 
1989 	if (toep->tx_credits == toep->tx_total) {
1990 		toep->tx_nocompl = 0;
1991 		toep->plen_nocompl = 0;
1992 	}
1993 
1994 	if (toep->flags & TPF_TX_SUSPENDED &&
1995 	    toep->tx_credits >= toep->tx_total / 4) {
1996 #ifdef VERBOSE_TRACES
1997 		CTR2(KTR_CXGBE, "%s: tid %d calling t4_push_frames", __func__,
1998 		    tid);
1999 #endif
2000 		toep->flags &= ~TPF_TX_SUSPENDED;
2001 		CURVNET_SET(toep->vnet);
2002 		t4_push_data(sc, toep, plen);
2003 		CURVNET_RESTORE();
2004 	} else if (plen > 0) {
2005 		struct sockbuf *sb = &so->so_snd;
2006 		int sbu;
2007 
2008 		SOCKBUF_LOCK(sb);
2009 		sbu = sbused(sb);
2010 		if (ulp_mode(toep) == ULP_MODE_ISCSI) {
2011 			if (__predict_false(sbu > 0)) {
2012 				/*
2013 				 * The data transmitted before the
2014 				 * tid's ULP mode changed to ISCSI is
2015 				 * still in so_snd.  Incoming credits
2016 				 * should account for so_snd first.
2017 				 */
2018 				sbdrop_locked(sb, min(sbu, plen));
2019 				plen -= min(sbu, plen);
2020 			}
2021 			sowwakeup_locked(so);	/* unlocks so_snd */
2022 			rqdrop_locked(&toep->ulp_pdu_reclaimq, plen);
2023 		} else {
2024 #ifdef VERBOSE_TRACES
2025 			CTR3(KTR_CXGBE, "%s: tid %d dropped %d bytes", __func__,
2026 			    tid, plen);
2027 #endif
2028 			sbdrop_locked(sb, plen);
2029 			if (!TAILQ_EMPTY(&toep->aiotx_jobq))
2030 				t4_aiotx_queue_toep(so, toep);
2031 			sowwakeup_locked(so);	/* unlocks so_snd */
2032 		}
2033 		SOCKBUF_UNLOCK_ASSERT(sb);
2034 	}
2035 
2036 	INP_WUNLOCK(inp);
2037 
2038 	return (0);
2039 }
2040 
2041 void
write_set_tcb_field(struct adapter * sc,void * dst,struct toepcb * toep,uint16_t word,uint64_t mask,uint64_t val,int reply,int cookie)2042 write_set_tcb_field(struct adapter *sc, void *dst, struct toepcb *toep,
2043     uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie)
2044 {
2045 	struct cpl_set_tcb_field *req = dst;
2046 
2047 	MPASS((cookie & ~M_COOKIE) == 0);
2048 	if (reply) {
2049 		MPASS(cookie != CPL_COOKIE_RESERVED);
2050 	}
2051 
2052 	INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid);
2053 	req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id));
2054 	if (reply == 0)
2055 		req->reply_ctrl |= htobe16(F_NO_REPLY);
2056 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie));
2057 	req->mask = htobe64(mask);
2058 	req->val = htobe64(val);
2059 }
2060 
2061 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)2062 t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, struct toepcb *toep,
2063     uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie)
2064 {
2065 	struct wrqe *wr;
2066 	struct ofld_tx_sdesc *txsd;
2067 	const u_int len = sizeof(struct cpl_set_tcb_field);
2068 
2069 	wr = alloc_wrqe(len, wrq);
2070 	if (wr == NULL) {
2071 		/* XXX */
2072 		panic("%s: allocation failure.", __func__);
2073 	}
2074 	write_set_tcb_field(sc, wrtod(wr), toep, word, mask, val, reply,
2075 	    cookie);
2076 
2077 	if (wrq->eq.type == EQ_OFLD) {
2078 		txsd = &toep->txsd[toep->txsd_pidx];
2079 		_Static_assert(howmany(len, 16) <= MAX_OFLD_TX_SDESC_CREDITS,
2080 		    "MAX_OFLD_TX_SDESC_CREDITS too small");
2081 		txsd->tx_credits = howmany(len, 16);
2082 		txsd->plen = 0;
2083 		KASSERT(toep->tx_credits >= txsd->tx_credits &&
2084 		    toep->txsd_avail > 0,
2085 		    ("%s: not enough credits (%d)", __func__,
2086 		    toep->tx_credits));
2087 		toep->tx_credits -= txsd->tx_credits;
2088 		if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
2089 			toep->txsd_pidx = 0;
2090 		toep->txsd_avail--;
2091 	}
2092 
2093 	t4_wrq_tx(sc, wr);
2094 }
2095 
2096 void
t4_init_cpl_io_handlers(void)2097 t4_init_cpl_io_handlers(void)
2098 {
2099 
2100 	t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close);
2101 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl);
2102 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req);
2103 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl,
2104 	    CPL_COOKIE_TOM);
2105 	t4_register_cpl_handler(CPL_RX_DATA, do_rx_data);
2106 	t4_register_shared_cpl_handler(CPL_FW4_ACK, do_fw4_ack, CPL_COOKIE_TOM);
2107 }
2108 
2109 void
t4_uninit_cpl_io_handlers(void)2110 t4_uninit_cpl_io_handlers(void)
2111 {
2112 
2113 	t4_register_cpl_handler(CPL_PEER_CLOSE, NULL);
2114 	t4_register_cpl_handler(CPL_CLOSE_CON_RPL, NULL);
2115 	t4_register_cpl_handler(CPL_ABORT_REQ_RSS, NULL);
2116 	t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, NULL, CPL_COOKIE_TOM);
2117 	t4_register_cpl_handler(CPL_RX_DATA, NULL);
2118 	t4_register_shared_cpl_handler(CPL_FW4_ACK, NULL, CPL_COOKIE_TOM);
2119 }
2120 
2121 /*
2122  * Use the 'backend1' field in AIO jobs to hold an error that should
2123  * be reported when the job is completed, the 'backend3' field to
2124  * store the amount of data sent by the AIO job so far, and the
2125  * 'backend4' field to hold a reference count on the job.
2126  *
2127  * Each unmapped mbuf holds a reference on the job as does the queue
2128  * so long as the job is queued.
2129  */
2130 #define	aio_error	backend1
2131 #define	aio_sent	backend3
2132 #define	aio_refs	backend4
2133 
2134 #ifdef VERBOSE_TRACES
2135 static int
jobtotid(struct kaiocb * job)2136 jobtotid(struct kaiocb *job)
2137 {
2138 	struct socket *so;
2139 	struct tcpcb *tp;
2140 	struct toepcb *toep;
2141 
2142 	so = job->fd_file->f_data;
2143 	tp = sototcpcb(so);
2144 	toep = tp->t_toe;
2145 	return (toep->tid);
2146 }
2147 #endif
2148 
2149 static void
aiotx_free_job(struct kaiocb * job)2150 aiotx_free_job(struct kaiocb *job)
2151 {
2152 	long status;
2153 	int error;
2154 
2155 	if (refcount_release(&job->aio_refs) == 0)
2156 		return;
2157 
2158 	error = (intptr_t)job->aio_error;
2159 	status = job->aio_sent;
2160 #ifdef VERBOSE_TRACES
2161 	CTR5(KTR_CXGBE, "%s: tid %d completed %p len %ld, error %d", __func__,
2162 	    jobtotid(job), job, status, error);
2163 #endif
2164 	if (error != 0 && status != 0)
2165 		error = 0;
2166 	if (error == ECANCELED)
2167 		aio_cancel(job);
2168 	else if (error)
2169 		aio_complete(job, -1, error);
2170 	else {
2171 		job->msgsnd = 1;
2172 		aio_complete(job, status, 0);
2173 	}
2174 }
2175 
2176 static void
aiotx_free_pgs(struct mbuf * m)2177 aiotx_free_pgs(struct mbuf *m)
2178 {
2179 	struct kaiocb *job;
2180 	vm_page_t pg;
2181 
2182 	M_ASSERTEXTPG(m);
2183 	job = m->m_ext.ext_arg1;
2184 #ifdef VERBOSE_TRACES
2185 	CTR3(KTR_CXGBE, "%s: completed %d bytes for tid %d", __func__,
2186 	    m->m_len, jobtotid(job));
2187 #endif
2188 
2189 	for (int i = 0; i < m->m_epg_npgs; i++) {
2190 		pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]);
2191 		vm_page_unwire(pg, PQ_ACTIVE);
2192 	}
2193 
2194 	aiotx_free_job(job);
2195 }
2196 
2197 /*
2198  * Allocate a chain of unmapped mbufs describing the next 'len' bytes
2199  * of an AIO job.
2200  */
2201 static struct mbuf *
alloc_aiotx_mbuf(struct kaiocb * job,int len)2202 alloc_aiotx_mbuf(struct kaiocb *job, int len)
2203 {
2204 	struct vmspace *vm;
2205 	vm_page_t pgs[MBUF_PEXT_MAX_PGS];
2206 	struct mbuf *m, *top, *last;
2207 	vm_map_t map;
2208 	vm_offset_t start;
2209 	int i, mlen, npages, pgoff;
2210 
2211 	KASSERT(job->aio_sent + len <= job->uaiocb.aio_nbytes,
2212 	    ("%s(%p, %d): request to send beyond end of buffer", __func__,
2213 	    job, len));
2214 
2215 	/*
2216 	 * The AIO subsystem will cancel and drain all requests before
2217 	 * permitting a process to exit or exec, so p_vmspace should
2218 	 * be stable here.
2219 	 */
2220 	vm = job->userproc->p_vmspace;
2221 	map = &vm->vm_map;
2222 	start = (uintptr_t)job->uaiocb.aio_buf + job->aio_sent;
2223 	pgoff = start & PAGE_MASK;
2224 
2225 	top = NULL;
2226 	last = NULL;
2227 	while (len > 0) {
2228 		mlen = imin(len, MBUF_PEXT_MAX_PGS * PAGE_SIZE - pgoff);
2229 		KASSERT(mlen == len || ((start + mlen) & PAGE_MASK) == 0,
2230 		    ("%s: next start (%#jx + %#x) is not page aligned",
2231 		    __func__, (uintmax_t)start, mlen));
2232 
2233 		npages = vm_fault_quick_hold_pages(map, start, mlen,
2234 		    VM_PROT_WRITE, pgs, nitems(pgs));
2235 		if (npages < 0)
2236 			break;
2237 
2238 		m = mb_alloc_ext_pgs(M_WAITOK, aiotx_free_pgs, M_RDONLY);
2239 		m->m_epg_1st_off = pgoff;
2240 		m->m_epg_npgs = npages;
2241 		if (npages == 1) {
2242 			KASSERT(mlen + pgoff <= PAGE_SIZE,
2243 			    ("%s: single page is too large (off %d len %d)",
2244 			    __func__, pgoff, mlen));
2245 			m->m_epg_last_len = mlen;
2246 		} else {
2247 			m->m_epg_last_len = mlen - (PAGE_SIZE - pgoff) -
2248 			    (npages - 2) * PAGE_SIZE;
2249 		}
2250 		for (i = 0; i < npages; i++)
2251 			m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pgs[i]);
2252 
2253 		m->m_len = mlen;
2254 		m->m_ext.ext_size = npages * PAGE_SIZE;
2255 		m->m_ext.ext_arg1 = job;
2256 		refcount_acquire(&job->aio_refs);
2257 
2258 #ifdef VERBOSE_TRACES
2259 		CTR5(KTR_CXGBE, "%s: tid %d, new mbuf %p for job %p, npages %d",
2260 		    __func__, jobtotid(job), m, job, npages);
2261 #endif
2262 
2263 		if (top == NULL)
2264 			top = m;
2265 		else
2266 			last->m_next = m;
2267 		last = m;
2268 
2269 		len -= mlen;
2270 		start += mlen;
2271 		pgoff = 0;
2272 	}
2273 
2274 	return (top);
2275 }
2276 
2277 static void
t4_aiotx_process_job(struct toepcb * toep,struct socket * so,struct kaiocb * job)2278 t4_aiotx_process_job(struct toepcb *toep, struct socket *so, struct kaiocb *job)
2279 {
2280 	struct sockbuf *sb;
2281 	struct inpcb *inp;
2282 	struct tcpcb *tp;
2283 	struct mbuf *m;
2284 	u_int sent;
2285 	int error, len;
2286 	bool moretocome, sendmore;
2287 
2288 	sb = &so->so_snd;
2289 	SOCKBUF_UNLOCK(sb);
2290 	m = NULL;
2291 
2292 #ifdef MAC
2293 	error = mac_socket_check_send(job->fd_file->f_cred, so);
2294 	if (error != 0)
2295 		goto out;
2296 #endif
2297 
2298 	/* Inline sosend_generic(). */
2299 
2300 	error = SOCK_IO_SEND_LOCK(so, SBL_WAIT);
2301 	MPASS(error == 0);
2302 
2303 sendanother:
2304 	SOCKBUF_LOCK(sb);
2305 	if (so->so_snd.sb_state & SBS_CANTSENDMORE) {
2306 		SOCKBUF_UNLOCK(sb);
2307 		SOCK_IO_SEND_UNLOCK(so);
2308 		if ((so->so_options & SO_NOSIGPIPE) == 0) {
2309 			PROC_LOCK(job->userproc);
2310 			kern_psignal(job->userproc, SIGPIPE);
2311 			PROC_UNLOCK(job->userproc);
2312 		}
2313 		error = EPIPE;
2314 		goto out;
2315 	}
2316 	if (so->so_error) {
2317 		error = so->so_error;
2318 		so->so_error = 0;
2319 		SOCKBUF_UNLOCK(sb);
2320 		SOCK_IO_SEND_UNLOCK(so);
2321 		goto out;
2322 	}
2323 	if ((so->so_state & SS_ISCONNECTED) == 0) {
2324 		SOCKBUF_UNLOCK(sb);
2325 		SOCK_IO_SEND_UNLOCK(so);
2326 		error = ENOTCONN;
2327 		goto out;
2328 	}
2329 	if (sbspace(sb) < sb->sb_lowat) {
2330 		MPASS(job->aio_sent == 0 || !(so->so_state & SS_NBIO));
2331 
2332 		/*
2333 		 * Don't block if there is too little room in the socket
2334 		 * buffer.  Instead, requeue the request.
2335 		 */
2336 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2337 			SOCKBUF_UNLOCK(sb);
2338 			SOCK_IO_SEND_UNLOCK(so);
2339 			error = ECANCELED;
2340 			goto out;
2341 		}
2342 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2343 		SOCKBUF_UNLOCK(sb);
2344 		SOCK_IO_SEND_UNLOCK(so);
2345 		goto out;
2346 	}
2347 
2348 	/*
2349 	 * Write as much data as the socket permits, but no more than a
2350 	 * a single sndbuf at a time.
2351 	 */
2352 	len = sbspace(sb);
2353 	if (len > job->uaiocb.aio_nbytes - job->aio_sent) {
2354 		len = job->uaiocb.aio_nbytes - job->aio_sent;
2355 		moretocome = false;
2356 	} else
2357 		moretocome = true;
2358 	if (len > toep->params.sndbuf) {
2359 		len = toep->params.sndbuf;
2360 		sendmore = true;
2361 	} else
2362 		sendmore = false;
2363 
2364 	if (!TAILQ_EMPTY(&toep->aiotx_jobq))
2365 		moretocome = true;
2366 	SOCKBUF_UNLOCK(sb);
2367 	MPASS(len != 0);
2368 
2369 	m = alloc_aiotx_mbuf(job, len);
2370 	if (m == NULL) {
2371 		SOCK_IO_SEND_UNLOCK(so);
2372 		error = EFAULT;
2373 		goto out;
2374 	}
2375 
2376 	/* Inlined tcp_usr_send(). */
2377 
2378 	inp = toep->inp;
2379 	INP_WLOCK(inp);
2380 	if (inp->inp_flags & INP_DROPPED) {
2381 		INP_WUNLOCK(inp);
2382 		SOCK_IO_SEND_UNLOCK(so);
2383 		error = ECONNRESET;
2384 		goto out;
2385 	}
2386 
2387 	sent = m_length(m, NULL);
2388 	job->aio_sent += sent;
2389 	counter_u64_add(toep->ofld_txq->tx_aio_octets, sent);
2390 
2391 	sbappendstream(sb, m, 0);
2392 	m = NULL;
2393 
2394 	if (!(inp->inp_flags & INP_DROPPED)) {
2395 		tp = intotcpcb(inp);
2396 		if (moretocome)
2397 			tp->t_flags |= TF_MORETOCOME;
2398 		error = tcp_output(tp);
2399 		if (error < 0) {
2400 			INP_UNLOCK_ASSERT(inp);
2401 			SOCK_IO_SEND_UNLOCK(so);
2402 			error = -error;
2403 			goto out;
2404 		}
2405 		if (moretocome)
2406 			tp->t_flags &= ~TF_MORETOCOME;
2407 	}
2408 
2409 	INP_WUNLOCK(inp);
2410 	if (sendmore)
2411 		goto sendanother;
2412 	SOCK_IO_SEND_UNLOCK(so);
2413 
2414 	if (error)
2415 		goto out;
2416 
2417 	/*
2418 	 * If this is a blocking socket and the request has not been
2419 	 * fully completed, requeue it until the socket is ready
2420 	 * again.
2421 	 */
2422 	if (job->aio_sent < job->uaiocb.aio_nbytes &&
2423 	    !(so->so_state & SS_NBIO)) {
2424 		SOCKBUF_LOCK(sb);
2425 		if (!aio_set_cancel_function(job, t4_aiotx_cancel)) {
2426 			SOCKBUF_UNLOCK(sb);
2427 			error = ECANCELED;
2428 			goto out;
2429 		}
2430 		TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list);
2431 		return;
2432 	}
2433 
2434 	/*
2435 	 * If the request will not be requeued, drop the queue's
2436 	 * reference to the job.  Any mbufs in flight should still
2437 	 * hold a reference, but this drops the reference that the
2438 	 * queue owns while it is waiting to queue mbufs to the
2439 	 * socket.
2440 	 */
2441 	aiotx_free_job(job);
2442 	counter_u64_add(toep->ofld_txq->tx_aio_jobs, 1);
2443 
2444 out:
2445 	if (error) {
2446 		job->aio_error = (void *)(intptr_t)error;
2447 		aiotx_free_job(job);
2448 	}
2449 	m_freem(m);
2450 	SOCKBUF_LOCK(sb);
2451 }
2452 
2453 static void
t4_aiotx_task(void * context,int pending)2454 t4_aiotx_task(void *context, int pending)
2455 {
2456 	struct toepcb *toep = context;
2457 	struct socket *so;
2458 	struct kaiocb *job;
2459 	struct epoch_tracker et;
2460 
2461 	so = toep->aiotx_so;
2462 	CURVNET_SET(toep->vnet);
2463 	NET_EPOCH_ENTER(et);
2464 	SOCKBUF_LOCK(&so->so_snd);
2465 	while (!TAILQ_EMPTY(&toep->aiotx_jobq) && sowriteable(so)) {
2466 		job = TAILQ_FIRST(&toep->aiotx_jobq);
2467 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2468 		if (!aio_clear_cancel_function(job))
2469 			continue;
2470 
2471 		t4_aiotx_process_job(toep, so, job);
2472 	}
2473 	toep->aiotx_so = NULL;
2474 	SOCKBUF_UNLOCK(&so->so_snd);
2475 	NET_EPOCH_EXIT(et);
2476 
2477 	free_toepcb(toep);
2478 	sorele(so);
2479 	CURVNET_RESTORE();
2480 }
2481 
2482 static void
t4_aiotx_queue_toep(struct socket * so,struct toepcb * toep)2483 t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep)
2484 {
2485 
2486 	SOCKBUF_LOCK_ASSERT(&toep->inp->inp_socket->so_snd);
2487 #ifdef VERBOSE_TRACES
2488 	CTR3(KTR_CXGBE, "%s: queueing aiotx task for tid %d, active = %s",
2489 	    __func__, toep->tid, toep->aiotx_so != NULL ? "true" : "false");
2490 #endif
2491 	if (toep->aiotx_so != NULL)
2492 		return;
2493 	soref(so);
2494 	toep->aiotx_so = so;
2495 	hold_toepcb(toep);
2496 	soaio_enqueue(&toep->aiotx_task);
2497 }
2498 
2499 static void
t4_aiotx_cancel(struct kaiocb * job)2500 t4_aiotx_cancel(struct kaiocb *job)
2501 {
2502 	struct socket *so;
2503 	struct sockbuf *sb;
2504 	struct tcpcb *tp;
2505 	struct toepcb *toep;
2506 
2507 	so = job->fd_file->f_data;
2508 	tp = sototcpcb(so);
2509 	toep = tp->t_toe;
2510 	MPASS(job->uaiocb.aio_lio_opcode == LIO_WRITE);
2511 	sb = &so->so_snd;
2512 
2513 	SOCKBUF_LOCK(sb);
2514 	if (!aio_cancel_cleared(job))
2515 		TAILQ_REMOVE(&toep->aiotx_jobq, job, list);
2516 	SOCKBUF_UNLOCK(sb);
2517 
2518 	job->aio_error = (void *)(intptr_t)ECANCELED;
2519 	aiotx_free_job(job);
2520 }
2521 
2522 int
t4_aio_queue_aiotx(struct socket * so,struct kaiocb * job)2523 t4_aio_queue_aiotx(struct socket *so, struct kaiocb *job)
2524 {
2525 	struct tcpcb *tp = sototcpcb(so);
2526 	struct toepcb *toep = tp->t_toe;
2527 	struct adapter *sc = td_adapter(toep->td);
2528 
2529 	/* This only handles writes. */
2530 	if (job->uaiocb.aio_lio_opcode != LIO_WRITE)
2531 		return (EOPNOTSUPP);
2532 
2533 	if (!sc->tt.tx_zcopy)
2534 		return (EOPNOTSUPP);
2535 
2536 	if (tls_tx_key(toep))
2537 		return (EOPNOTSUPP);
2538 
2539 	SOCKBUF_LOCK(&so->so_snd);
2540 #ifdef VERBOSE_TRACES
2541 	CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid);
2542 #endif
2543 	if (!aio_set_cancel_function(job, t4_aiotx_cancel))
2544 		panic("new job was cancelled");
2545 	refcount_init(&job->aio_refs, 1);
2546 	TAILQ_INSERT_TAIL(&toep->aiotx_jobq, job, list);
2547 	if (sowriteable(so))
2548 		t4_aiotx_queue_toep(so, toep);
2549 	SOCKBUF_UNLOCK(&so->so_snd);
2550 	return (0);
2551 }
2552 
2553 void
aiotx_init_toep(struct toepcb * toep)2554 aiotx_init_toep(struct toepcb *toep)
2555 {
2556 
2557 	TAILQ_INIT(&toep->aiotx_jobq);
2558 	TASK_INIT(&toep->aiotx_task, 0, t4_aiotx_task, toep);
2559 }
2560 #endif
2561