xref: /freebsd/sys/dev/cxgbe/cxgbei/icl_cxgbei.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
1 /*-
2  * Copyright (c) 2012 The FreeBSD Foundation
3  * Copyright (c) 2015 Chelsio Communications, Inc.
4  * All rights reserved.
5  *
6  * This software was developed by Edward Tomasz Napierala under sponsorship
7  * from the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  */
31 
32 /*
33  * cxgbei implementation of iSCSI Common Layer kobj(9) interface.
34  */
35 
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 #include "opt_inet.h"
40 #include "opt_inet6.h"
41 
42 #ifdef TCP_OFFLOAD
43 #include <sys/param.h>
44 #include <sys/capsicum.h>
45 #include <sys/condvar.h>
46 #include <sys/conf.h>
47 #include <sys/file.h>
48 #include <sys/kernel.h>
49 #include <sys/kthread.h>
50 #include <sys/lock.h>
51 #include <sys/mbuf.h>
52 #include <sys/mutex.h>
53 #include <sys/module.h>
54 #include <sys/protosw.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
59 #include <sys/sx.h>
60 #include <sys/uio.h>
61 #include <machine/bus.h>
62 #include <vm/uma.h>
63 #include <netinet/in.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/tcp.h>
66 #include <netinet/tcp_var.h>
67 #include <netinet/toecore.h>
68 
69 #include <dev/iscsi/icl.h>
70 #include <dev/iscsi/iscsi_proto.h>
71 #include <icl_conn_if.h>
72 
73 #include "common/common.h"
74 #include "tom/t4_tom.h"
75 #include "cxgbei.h"
76 
77 SYSCTL_NODE(_kern_icl, OID_AUTO, cxgbei, CTLFLAG_RD, 0, "Chelsio iSCSI offload");
78 static int coalesce = 1;
79 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, coalesce, CTLFLAG_RWTUN,
80 	&coalesce, 0, "Try to coalesce PDUs before sending");
81 static int partial_receive_len = 128 * 1024;
82 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, partial_receive_len, CTLFLAG_RWTUN,
83     &partial_receive_len, 0, "Minimum read size for partially received "
84     "data segment");
85 static int sendspace = 1048576;
86 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, sendspace, CTLFLAG_RWTUN,
87     &sendspace, 0, "Default send socket buffer size");
88 static int recvspace = 1048576;
89 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, recvspace, CTLFLAG_RWTUN,
90     &recvspace, 0, "Default receive socket buffer size");
91 
92 static uma_zone_t icl_transfer_zone;
93 
94 static volatile u_int icl_cxgbei_ncons;
95 
96 #define ICL_CONN_LOCK(X)		mtx_lock(X->ic_lock)
97 #define ICL_CONN_UNLOCK(X)		mtx_unlock(X->ic_lock)
98 #define ICL_CONN_LOCK_ASSERT(X)		mtx_assert(X->ic_lock, MA_OWNED)
99 #define ICL_CONN_LOCK_ASSERT_NOT(X)	mtx_assert(X->ic_lock, MA_NOTOWNED)
100 
101 struct icl_pdu *icl_cxgbei_new_pdu(int);
102 void icl_cxgbei_new_pdu_set_conn(struct icl_pdu *, struct icl_conn *);
103 
104 static icl_conn_new_pdu_t	icl_cxgbei_conn_new_pdu;
105 icl_conn_pdu_free_t	icl_cxgbei_conn_pdu_free;
106 static icl_conn_pdu_data_segment_length_t
107 				    icl_cxgbei_conn_pdu_data_segment_length;
108 static icl_conn_pdu_append_data_t	icl_cxgbei_conn_pdu_append_data;
109 static icl_conn_pdu_get_data_t	icl_cxgbei_conn_pdu_get_data;
110 static icl_conn_pdu_queue_t	icl_cxgbei_conn_pdu_queue;
111 static icl_conn_handoff_t	icl_cxgbei_conn_handoff;
112 static icl_conn_free_t		icl_cxgbei_conn_free;
113 static icl_conn_close_t		icl_cxgbei_conn_close;
114 static icl_conn_task_setup_t	icl_cxgbei_conn_task_setup;
115 static icl_conn_task_done_t	icl_cxgbei_conn_task_done;
116 static icl_conn_transfer_setup_t	icl_cxgbei_conn_transfer_setup;
117 static icl_conn_transfer_done_t	icl_cxgbei_conn_transfer_done;
118 
119 static kobj_method_t icl_cxgbei_methods[] = {
120 	KOBJMETHOD(icl_conn_new_pdu, icl_cxgbei_conn_new_pdu),
121 	KOBJMETHOD(icl_conn_pdu_free, icl_cxgbei_conn_pdu_free),
122 	KOBJMETHOD(icl_conn_pdu_data_segment_length,
123 	    icl_cxgbei_conn_pdu_data_segment_length),
124 	KOBJMETHOD(icl_conn_pdu_append_data, icl_cxgbei_conn_pdu_append_data),
125 	KOBJMETHOD(icl_conn_pdu_get_data, icl_cxgbei_conn_pdu_get_data),
126 	KOBJMETHOD(icl_conn_pdu_queue, icl_cxgbei_conn_pdu_queue),
127 	KOBJMETHOD(icl_conn_handoff, icl_cxgbei_conn_handoff),
128 	KOBJMETHOD(icl_conn_free, icl_cxgbei_conn_free),
129 	KOBJMETHOD(icl_conn_close, icl_cxgbei_conn_close),
130 	KOBJMETHOD(icl_conn_task_setup, icl_cxgbei_conn_task_setup),
131 	KOBJMETHOD(icl_conn_task_done, icl_cxgbei_conn_task_done),
132 	KOBJMETHOD(icl_conn_transfer_setup, icl_cxgbei_conn_transfer_setup),
133 	KOBJMETHOD(icl_conn_transfer_done, icl_cxgbei_conn_transfer_done),
134 	{ 0, 0 }
135 };
136 
137 DEFINE_CLASS(icl_cxgbei, icl_cxgbei_methods, sizeof(struct icl_cxgbei_conn));
138 
139 #if 0
140 /*
141  * Subtract another 256 for AHS from MAX_DSL if AHS could be used.
142  */
143 #define CXGBEI_MAX_PDU 16224
144 #define CXGBEI_MAX_DSL (CXGBEI_MAX_PDU - sizeof(struct iscsi_bhs) - 8)
145 #endif
146 #define CXGBEI_MAX_DSL 8192
147 #define CXGBEI_MAX_PDU (CXGBEI_MAX_DSL + sizeof(struct iscsi_bhs) + 8)
148 
149 void
150 icl_cxgbei_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip)
151 {
152 #ifdef INVARIANTS
153 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
154 #endif
155 
156 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
157 	MPASS(ic == ip->ip_conn);
158 	MPASS(ip->ip_bhs_mbuf != NULL);
159 
160 	m_freem(ip->ip_ahs_mbuf);
161 	m_freem(ip->ip_data_mbuf);
162 	m_freem(ip->ip_bhs_mbuf);	/* storage for icl_cxgbei_pdu itself */
163 
164 #ifdef DIAGNOSTIC
165 	if (__predict_true(ic != NULL))
166 		refcount_release(&ic->ic_outstanding_pdus);
167 #endif
168 }
169 
170 struct icl_pdu *
171 icl_cxgbei_new_pdu(int flags)
172 {
173 	struct icl_cxgbei_pdu *icp;
174 	struct icl_pdu *ip;
175 	struct mbuf *m;
176 	uintptr_t a;
177 
178 	m = m_gethdr(flags, MT_DATA);
179 	if (__predict_false(m == NULL))
180 		return (NULL);
181 
182 	a = roundup2(mtod(m, uintptr_t), _Alignof(struct icl_cxgbei_pdu));
183 	icp = (struct icl_cxgbei_pdu *)a;
184 	bzero(icp, sizeof(*icp));
185 
186 	icp->icp_signature = CXGBEI_PDU_SIGNATURE;
187 	ip = &icp->ip;
188 	ip->ip_bhs_mbuf = m;
189 
190 	a = roundup2((uintptr_t)(icp + 1), _Alignof(struct iscsi_bhs *));
191 	ip->ip_bhs = (struct iscsi_bhs *)a;
192 #ifdef INVARIANTS
193 	/* Everything must fit entirely in the mbuf. */
194 	a = (uintptr_t)(ip->ip_bhs + 1);
195 	MPASS(a <= (uintptr_t)m + MSIZE);
196 #endif
197 	bzero(ip->ip_bhs, sizeof(*ip->ip_bhs));
198 
199 	m->m_data = (void *)ip->ip_bhs;
200 	m->m_len = sizeof(struct iscsi_bhs);
201 	m->m_pkthdr.len = m->m_len;
202 
203 	return (ip);
204 }
205 
206 void
207 icl_cxgbei_new_pdu_set_conn(struct icl_pdu *ip, struct icl_conn *ic)
208 {
209 
210 	ip->ip_conn = ic;
211 #ifdef DIAGNOSTIC
212 	refcount_acquire(&ic->ic_outstanding_pdus);
213 #endif
214 }
215 
216 /*
217  * Allocate icl_pdu with empty BHS to fill up by the caller.
218  */
219 static struct icl_pdu *
220 icl_cxgbei_conn_new_pdu(struct icl_conn *ic, int flags)
221 {
222 	struct icl_pdu *ip;
223 
224 	ip = icl_cxgbei_new_pdu(flags);
225 	if (__predict_false(ip == NULL))
226 		return (NULL);
227 	icl_cxgbei_new_pdu_set_conn(ip, ic);
228 
229 	return (ip);
230 }
231 
232 static size_t
233 icl_pdu_data_segment_length(const struct icl_pdu *request)
234 {
235 	uint32_t len = 0;
236 
237 	len += request->ip_bhs->bhs_data_segment_len[0];
238 	len <<= 8;
239 	len += request->ip_bhs->bhs_data_segment_len[1];
240 	len <<= 8;
241 	len += request->ip_bhs->bhs_data_segment_len[2];
242 
243 	return (len);
244 }
245 
246 size_t
247 icl_cxgbei_conn_pdu_data_segment_length(struct icl_conn *ic,
248     const struct icl_pdu *request)
249 {
250 
251 	return (icl_pdu_data_segment_length(request));
252 }
253 
254 static uint32_t
255 icl_conn_build_tasktag(struct icl_conn *ic, uint32_t tag)
256 {
257 	return tag;
258 }
259 
260 static struct mbuf *
261 finalize_pdu(struct icl_cxgbei_conn *icc, struct icl_cxgbei_pdu *icp)
262 {
263 	struct icl_pdu *ip = &icp->ip;
264 	uint8_t ulp_submode, padding;
265 	struct mbuf *m, *last;
266 	struct iscsi_bhs *bhs;
267 
268 	/*
269 	 * Fix up the data segment mbuf first.
270 	 */
271 	m = ip->ip_data_mbuf;
272 	ulp_submode = icc->ulp_submode;
273 	if (m) {
274 		last = m_last(m);
275 
276 		/*
277 		 * Round up the data segment to a 4B boundary.  Pad with 0 if
278 		 * necessary.  There will definitely be room in the mbuf.
279 		 */
280 		padding = roundup2(ip->ip_data_len, 4) - ip->ip_data_len;
281 		if (padding) {
282 			bzero(mtod(last, uint8_t *) + last->m_len, padding);
283 			last->m_len += padding;
284 		}
285 	} else {
286 		MPASS(ip->ip_data_len == 0);
287 		ulp_submode &= ~ULP_CRC_DATA;
288 		padding = 0;
289 	}
290 
291 	/*
292 	 * Now the header mbuf that has the BHS.
293 	 */
294 	m = ip->ip_bhs_mbuf;
295 	MPASS(m->m_pkthdr.len == sizeof(struct iscsi_bhs));
296 	MPASS(m->m_len == sizeof(struct iscsi_bhs));
297 
298 	bhs = ip->ip_bhs;
299 	bhs->bhs_data_segment_len[2] = ip->ip_data_len;
300 	bhs->bhs_data_segment_len[1] = ip->ip_data_len >> 8;
301 	bhs->bhs_data_segment_len[0] = ip->ip_data_len >> 16;
302 
303 	/* "Convert" PDU to mbuf chain.  Do not use icp/ip after this. */
304 	m->m_pkthdr.len = sizeof(struct iscsi_bhs) + ip->ip_data_len + padding;
305 	m->m_next = ip->ip_data_mbuf;
306 	set_mbuf_ulp_submode(m, ulp_submode);
307 #ifdef INVARIANTS
308 	bzero(icp, sizeof(*icp));
309 #endif
310 #ifdef DIAGNOSTIC
311 	refcount_release(&icc->ic.ic_outstanding_pdus);
312 #endif
313 
314 	return (m);
315 }
316 
317 int
318 icl_cxgbei_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *ip,
319     const void *addr, size_t len, int flags)
320 {
321 	struct mbuf *m;
322 #ifdef INVARIANTS
323 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
324 #endif
325 
326 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
327 	MPASS(ic == ip->ip_conn);
328 	KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len));
329 
330 	m = ip->ip_data_mbuf;
331 	if (m == NULL) {
332 		m = m_getjcl(M_NOWAIT, MT_DATA, 0, MJUM16BYTES);
333 		if (__predict_false(m == NULL))
334 			return (ENOMEM);
335 
336 		ip->ip_data_mbuf = m;
337 	}
338 
339 	if (__predict_true(m_append(m, len, addr) != 0)) {
340 		ip->ip_data_len += len;
341 		MPASS(ip->ip_data_len <= CXGBEI_MAX_DSL);
342 		return (0);
343 	} else {
344 	    	if (flags & M_WAITOK) {
345 			CXGBE_UNIMPLEMENTED("fail safe append");
346 		}
347 		ip->ip_data_len = m_length(m, NULL);
348 		return (1);
349 	}
350 }
351 
352 void
353 icl_cxgbei_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip,
354     size_t off, void *addr, size_t len)
355 {
356 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
357 
358 	if (icp->pdu_flags & SBUF_ULP_FLAG_DATA_DDPED)
359 		return; /* data is DDP'ed, no need to copy */
360 	m_copydata(ip->ip_data_mbuf, off, len, addr);
361 }
362 
363 void
364 icl_cxgbei_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip)
365 {
366 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
367 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
368 	struct socket *so = ic->ic_socket;
369 	struct toepcb *toep = icc->toep;
370 	struct inpcb *inp;
371 	struct mbuf *m;
372 
373 	MPASS(ic == ip->ip_conn);
374 	MPASS(ip->ip_bhs_mbuf != NULL);
375 	/* The kernel doesn't generate PDUs with AHS. */
376 	MPASS(ip->ip_ahs_mbuf == NULL && ip->ip_ahs_len == 0);
377 
378 	ICL_CONN_LOCK_ASSERT(ic);
379 	/* NOTE: sowriteable without so_snd lock is a mostly harmless race. */
380 	if (ic->ic_disconnecting || so == NULL || !sowriteable(so)) {
381 		icl_cxgbei_conn_pdu_free(ic, ip);
382 		return;
383 	}
384 
385 	m = finalize_pdu(icc, icp);
386 	M_ASSERTPKTHDR(m);
387 	MPASS((m->m_pkthdr.len & 3) == 0);
388 	MPASS(m->m_pkthdr.len + 8 <= CXGBEI_MAX_PDU);
389 
390 	/*
391 	 * Do not get inp from toep->inp as the toepcb might have detached
392 	 * already.
393 	 */
394 	inp = sotoinpcb(so);
395 	INP_WLOCK(inp);
396 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) ||
397 	    __predict_false((toep->flags & TPF_ATTACHED) == 0))
398 		m_freem(m);
399 	else {
400 		mbufq_enqueue(&toep->ulp_pduq, m);
401 		t4_push_pdus(icc->sc, toep, 0);
402 	}
403 	INP_WUNLOCK(inp);
404 }
405 
406 static struct icl_conn *
407 icl_cxgbei_new_conn(const char *name, struct mtx *lock)
408 {
409 	struct icl_cxgbei_conn *icc;
410 	struct icl_conn *ic;
411 
412 	refcount_acquire(&icl_cxgbei_ncons);
413 
414 	icc = (struct icl_cxgbei_conn *)kobj_create(&icl_cxgbei_class, M_CXGBE,
415 	    M_WAITOK | M_ZERO);
416 	icc->icc_signature = CXGBEI_CONN_SIGNATURE;
417 	STAILQ_INIT(&icc->rcvd_pdus);
418 
419 	ic = &icc->ic;
420 	ic->ic_lock = lock;
421 
422 	/* XXXNP: review.  Most of these icl_conn fields aren't really used */
423 	STAILQ_INIT(&ic->ic_to_send);
424 	cv_init(&ic->ic_send_cv, "icl_cxgbei_tx");
425 	cv_init(&ic->ic_receive_cv, "icl_cxgbei_rx");
426 #ifdef DIAGNOSTIC
427 	refcount_init(&ic->ic_outstanding_pdus, 0);
428 #endif
429 	ic->ic_max_data_segment_length = CXGBEI_MAX_DSL;
430 	ic->ic_name = name;
431 	ic->ic_offload = "cxgbei";
432 
433 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
434 
435 	return (ic);
436 }
437 
438 void
439 icl_cxgbei_conn_free(struct icl_conn *ic)
440 {
441 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
442 
443 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
444 
445 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
446 
447 	cv_destroy(&ic->ic_send_cv);
448 	cv_destroy(&ic->ic_receive_cv);
449 
450 	kobj_delete((struct kobj *)icc, M_CXGBE);
451 	refcount_release(&icl_cxgbei_ncons);
452 }
453 
454 static int
455 icl_cxgbei_setsockopt(struct icl_conn *ic, struct socket *so)
456 {
457 	size_t minspace;
458 	struct sockopt opt;
459 	int error, one = 1;
460 
461 	/*
462 	 * For sendspace, this is required because the current code cannot
463 	 * send a PDU in pieces; thus, the minimum buffer size is equal
464 	 * to the maximum PDU size.  "+4" is to account for possible padding.
465 	 *
466 	 * What we should actually do here is to use autoscaling, but set
467 	 * some minimal buffer size to "minspace".  I don't know a way to do
468 	 * that, though.
469 	 */
470 	minspace = sizeof(struct iscsi_bhs) + ic->ic_max_data_segment_length +
471 	    ISCSI_HEADER_DIGEST_SIZE + ISCSI_DATA_DIGEST_SIZE + 4;
472 	if (sendspace < minspace)
473 		sendspace = minspace;
474 	if (recvspace < minspace)
475 		recvspace = minspace;
476 
477 	error = soreserve(so, sendspace, recvspace);
478 	if (error != 0) {
479 		icl_cxgbei_conn_close(ic);
480 		return (error);
481 	}
482 	SOCKBUF_LOCK(&so->so_snd);
483 	so->so_snd.sb_flags |= SB_AUTOSIZE;
484 	SOCKBUF_UNLOCK(&so->so_snd);
485 	SOCKBUF_LOCK(&so->so_rcv);
486 	so->so_rcv.sb_flags |= SB_AUTOSIZE;
487 	SOCKBUF_UNLOCK(&so->so_rcv);
488 
489 	/*
490 	 * Disable Nagle.
491 	 */
492 	bzero(&opt, sizeof(opt));
493 	opt.sopt_dir = SOPT_SET;
494 	opt.sopt_level = IPPROTO_TCP;
495 	opt.sopt_name = TCP_NODELAY;
496 	opt.sopt_val = &one;
497 	opt.sopt_valsize = sizeof(one);
498 	error = sosetopt(so, &opt);
499 	if (error != 0) {
500 		icl_cxgbei_conn_close(ic);
501 		return (error);
502 	}
503 
504 	return (0);
505 }
506 
507 /*
508  * Request/response structure used to find out the adapter offloading a socket.
509  */
510 struct find_ofld_adapter_rr {
511 	struct socket *so;
512 	struct adapter *sc;	/* result */
513 };
514 
515 static void
516 find_offload_adapter(struct adapter *sc, void *arg)
517 {
518 	struct find_ofld_adapter_rr *fa = arg;
519 	struct socket *so = fa->so;
520 	struct tom_data *td = sc->tom_softc;
521 	struct tcpcb *tp;
522 	struct inpcb *inp;
523 
524 	/* Non-TCP were filtered out earlier. */
525 	MPASS(so->so_proto->pr_protocol == IPPROTO_TCP);
526 
527 	if (fa->sc != NULL)
528 		return;	/* Found already. */
529 
530 	if (td == NULL)
531 		return;	/* TOE not enabled on this adapter. */
532 
533 	inp = sotoinpcb(so);
534 	INP_WLOCK(inp);
535 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
536 		tp = intotcpcb(inp);
537 		if (tp->t_flags & TF_TOE && tp->tod == &td->tod)
538 			fa->sc = sc;	/* Found. */
539 	}
540 	INP_WUNLOCK(inp);
541 }
542 
543 /* XXXNP: move this to t4_tom. */
544 static void
545 send_iscsi_flowc_wr(struct adapter *sc, struct toepcb *toep, int maxlen)
546 {
547 	struct wrqe *wr;
548 	struct fw_flowc_wr *flowc;
549 	const u_int nparams = 1;
550 	u_int flowclen;
551 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
552 
553 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
554 
555 	wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
556 	if (wr == NULL) {
557 		/* XXX */
558 		panic("%s: allocation failure.", __func__);
559 	}
560 	flowc = wrtod(wr);
561 	memset(flowc, 0, wr->wr_len);
562 
563 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
564 	    V_FW_FLOWC_WR_NPARAMS(nparams));
565 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
566 	    V_FW_WR_FLOWID(toep->tid));
567 
568 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_TXDATAPLEN_MAX;
569 	flowc->mnemval[0].val = htobe32(maxlen);
570 
571 	txsd->tx_credits = howmany(flowclen, 16);
572 	txsd->plen = 0;
573 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
574 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
575 	toep->tx_credits -= txsd->tx_credits;
576 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
577 		toep->txsd_pidx = 0;
578 	toep->txsd_avail--;
579 
580         t4_wrq_tx(sc, wr);
581 }
582 
583 static void
584 set_ulp_mode_iscsi(struct adapter *sc, struct toepcb *toep, int hcrc, int dcrc)
585 {
586 	uint64_t val = 0;
587 
588 	if (hcrc)
589 		val |= ULP_CRC_HEADER;
590 	if (dcrc)
591 		val |= ULP_CRC_DATA;
592 	val <<= 4;
593 	val |= ULP_MODE_ISCSI;
594 
595 	CTR4(KTR_CXGBE, "%s: tid %u, ULP_MODE_ISCSI, CRC hdr=%d data=%d",
596 	    __func__, toep->tid, hcrc, dcrc);
597 
598 	t4_set_tcb_field(sc, toep, 1, 0, 0xfff, val);
599 }
600 
601 /*
602  * XXXNP: Who is responsible for cleaning up the socket if this returns with an
603  * error?  Review all error paths.
604  *
605  * XXXNP: What happens to the socket's fd reference if the operation is
606  * successful, and how does that affect the socket's life cycle?
607  */
608 int
609 icl_cxgbei_conn_handoff(struct icl_conn *ic, int fd)
610 {
611 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
612 	struct find_ofld_adapter_rr fa;
613 	struct file *fp;
614 	struct socket *so;
615 	struct inpcb *inp;
616 	struct tcpcb *tp;
617 	struct toepcb *toep;
618 	cap_rights_t rights;
619 	int error;
620 
621 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
622 	ICL_CONN_LOCK_ASSERT_NOT(ic);
623 
624 	/*
625 	 * Steal the socket from userland.
626 	 */
627 	error = fget(curthread, fd,
628 	    cap_rights_init(&rights, CAP_SOCK_CLIENT), &fp);
629 	if (error != 0)
630 		return (error);
631 	if (fp->f_type != DTYPE_SOCKET) {
632 		fdrop(fp, curthread);
633 		return (EINVAL);
634 	}
635 	so = fp->f_data;
636 	if (so->so_type != SOCK_STREAM ||
637 	    so->so_proto->pr_protocol != IPPROTO_TCP) {
638 		fdrop(fp, curthread);
639 		return (EINVAL);
640 	}
641 
642 	ICL_CONN_LOCK(ic);
643 	if (ic->ic_socket != NULL) {
644 		ICL_CONN_UNLOCK(ic);
645 		fdrop(fp, curthread);
646 		return (EBUSY);
647 	}
648 	ic->ic_disconnecting = false;
649 	ic->ic_socket = so;
650 	fp->f_ops = &badfileops;
651 	fp->f_data = NULL;
652 	fdrop(fp, curthread);
653 	ICL_CONN_UNLOCK(ic);
654 
655 	/* Find the adapter offloading this socket. */
656 	fa.sc = NULL;
657 	fa.so = so;
658 	t4_iterate(find_offload_adapter, &fa);
659 	if (fa.sc == NULL)
660 		return (EINVAL);
661 	icc->sc = fa.sc;
662 
663 	error = icl_cxgbei_setsockopt(ic, so);
664 	if (error)
665 		return (error);
666 
667 	inp = sotoinpcb(so);
668 	INP_WLOCK(inp);
669 	tp = intotcpcb(inp);
670 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))
671 		error = EBUSY;
672 	else {
673 		/*
674 		 * socket could not have been "unoffloaded" if here.
675 		 */
676 		MPASS(tp->t_flags & TF_TOE);
677 		MPASS(tp->tod != NULL);
678 		MPASS(tp->t_toe != NULL);
679 		toep = tp->t_toe;
680 		MPASS(toep->vi->pi->adapter == icc->sc);
681 		icc->toep = toep;
682 		icc->cwt = cxgbei_select_worker_thread(icc);
683 		icc->ulp_submode = 0;
684 		if (ic->ic_header_crc32c)
685 			icc->ulp_submode |= ULP_CRC_HEADER;
686 		if (ic->ic_data_crc32c)
687 			icc->ulp_submode |= ULP_CRC_DATA;
688 		so->so_options |= SO_NO_DDP;
689 		toep->ulp_mode = ULP_MODE_ISCSI;
690 		toep->ulpcb = icc;
691 
692 		send_iscsi_flowc_wr(icc->sc, toep, CXGBEI_MAX_PDU);
693 		set_ulp_mode_iscsi(icc->sc, toep, ic->ic_header_crc32c,
694 		    ic->ic_data_crc32c);
695 		error = 0;
696 	}
697 	INP_WUNLOCK(inp);
698 
699 	return (error);
700 }
701 
702 void
703 icl_cxgbei_conn_close(struct icl_conn *ic)
704 {
705 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
706 	struct icl_pdu *ip;
707 	struct socket *so;
708 	struct sockbuf *sb;
709 	struct inpcb *inp;
710 	struct toepcb *toep = icc->toep;
711 
712 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
713 	ICL_CONN_LOCK_ASSERT_NOT(ic);
714 
715 	ICL_CONN_LOCK(ic);
716 	so = ic->ic_socket;
717 	if (ic->ic_disconnecting || so == NULL) {
718 		CTR4(KTR_CXGBE, "%s: icc %p (disconnecting = %d), so %p",
719 		    __func__, icc, ic->ic_disconnecting, so);
720 		ICL_CONN_UNLOCK(ic);
721 		return;
722 	}
723 	ic->ic_disconnecting = true;
724 
725 	/* These are unused in this driver right now. */
726 	MPASS(STAILQ_EMPTY(&ic->ic_to_send));
727 	MPASS(ic->ic_receive_pdu == NULL);
728 
729 #ifdef DIAGNOSTIC
730 	KASSERT(ic->ic_outstanding_pdus == 0,
731 	    ("destroying session with %d outstanding PDUs",
732 	     ic->ic_outstanding_pdus));
733 #endif
734 	ICL_CONN_UNLOCK(ic);
735 
736 	CTR3(KTR_CXGBE, "%s: tid %d, icc %p", __func__, toep ? toep->tid : -1,
737 	    icc);
738 	inp = sotoinpcb(so);
739 	sb = &so->so_rcv;
740 	INP_WLOCK(inp);
741 	if (toep != NULL) {	/* NULL if connection was never offloaded. */
742 		toep->ulpcb = NULL;
743 		mbufq_drain(&toep->ulp_pduq);
744 		SOCKBUF_LOCK(sb);
745 		if (icc->rx_flags & RXF_ACTIVE) {
746 			volatile u_int *p = &icc->rx_flags;
747 
748 			SOCKBUF_UNLOCK(sb);
749 			INP_WUNLOCK(inp);
750 
751 			while (*p & RXF_ACTIVE)
752 				pause("conclo", 1);
753 
754 			INP_WLOCK(inp);
755 			SOCKBUF_LOCK(sb);
756 		}
757 
758 		while (!STAILQ_EMPTY(&icc->rcvd_pdus)) {
759 			ip = STAILQ_FIRST(&icc->rcvd_pdus);
760 			STAILQ_REMOVE_HEAD(&icc->rcvd_pdus, ip_next);
761 			icl_cxgbei_conn_pdu_free(ic, ip);
762 		}
763 		SOCKBUF_UNLOCK(sb);
764 	}
765 	INP_WUNLOCK(inp);
766 
767 	ICL_CONN_LOCK(ic);
768 	ic->ic_socket = NULL;
769 	ICL_CONN_UNLOCK(ic);
770 
771 	/*
772 	 * XXXNP: we should send RST instead of FIN when PDUs held in various
773 	 * queues were purged instead of delivered reliably but soabort isn't
774 	 * really general purpose and wouldn't do the right thing here.
775 	 */
776 	soclose(so);
777 }
778 
779 int
780 icl_cxgbei_conn_task_setup(struct icl_conn *ic, struct ccb_scsiio *csio,
781     uint32_t *task_tagp, void **prvp)
782 {
783 	void *prv;
784 
785 	*task_tagp = icl_conn_build_tasktag(ic, *task_tagp);
786 
787 	prv = uma_zalloc(icl_transfer_zone, M_NOWAIT | M_ZERO);
788 	if (prv == NULL)
789 		return (ENOMEM);
790 
791 	*prvp = prv;
792 
793 	cxgbei_conn_task_reserve_itt(ic, prvp, csio, task_tagp);
794 
795 	return (0);
796 }
797 
798 void
799 icl_cxgbei_conn_task_done(struct icl_conn *ic, void *prv)
800 {
801 
802 	cxgbei_cleanup_task(ic, prv);
803 	uma_zfree(icl_transfer_zone, prv);
804 }
805 
806 int
807 icl_cxgbei_conn_transfer_setup(struct icl_conn *ic, union ctl_io *io,
808     uint32_t *transfer_tag, void **prvp)
809 {
810 	void *prv;
811 
812 	*transfer_tag = icl_conn_build_tasktag(ic, *transfer_tag);
813 
814 	prv = uma_zalloc(icl_transfer_zone, M_NOWAIT | M_ZERO);
815 	if (prv == NULL)
816 		return (ENOMEM);
817 
818 	*prvp = prv;
819 
820 	cxgbei_conn_transfer_reserve_ttt(ic, prvp, io, transfer_tag);
821 
822 	return (0);
823 }
824 
825 void
826 icl_cxgbei_conn_transfer_done(struct icl_conn *ic, void *prv)
827 {
828 	cxgbei_cleanup_task(ic, prv);
829 	uma_zfree(icl_transfer_zone, prv);
830 }
831 
832 static int
833 icl_cxgbei_limits(size_t *limitp)
834 {
835 
836 	*limitp = CXGBEI_MAX_DSL;
837 
838 	return (0);
839 }
840 
841 static int
842 icl_cxgbei_load(void)
843 {
844 	int error;
845 
846 	icl_transfer_zone = uma_zcreate("icl_transfer",
847 	    16 * 1024, NULL, NULL, NULL, NULL,
848 	    UMA_ALIGN_PTR, 0);
849 
850 	refcount_init(&icl_cxgbei_ncons, 0);
851 
852 	error = icl_register("cxgbei", 100, icl_cxgbei_limits,
853 	    icl_cxgbei_new_conn);
854 	KASSERT(error == 0, ("failed to register"));
855 
856 	return (error);
857 }
858 
859 static int
860 icl_cxgbei_unload(void)
861 {
862 
863 	if (icl_cxgbei_ncons != 0)
864 		return (EBUSY);
865 
866 	icl_unregister("cxgbei");
867 
868 	uma_zdestroy(icl_transfer_zone);
869 
870 	return (0);
871 }
872 
873 static int
874 icl_cxgbei_modevent(module_t mod, int what, void *arg)
875 {
876 
877 	switch (what) {
878 	case MOD_LOAD:
879 		return (icl_cxgbei_load());
880 	case MOD_UNLOAD:
881 		return (icl_cxgbei_unload());
882 	default:
883 		return (EINVAL);
884 	}
885 }
886 
887 moduledata_t icl_cxgbei_data = {
888 	"icl_cxgbei",
889 	icl_cxgbei_modevent,
890 	0
891 };
892 
893 DECLARE_MODULE(icl_cxgbei, icl_cxgbei_data, SI_SUB_DRIVERS, SI_ORDER_MIDDLE);
894 MODULE_DEPEND(icl_cxgbei, icl, 1, 1, 1);
895 MODULE_VERSION(icl_cxgbei, 1);
896 #endif
897