xref: /freebsd/sys/dev/cxgbe/cxgbei/icl_cxgbei.c (revision 2b8331622f0b212cf3bb4fc4914a501e5321d506)
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/bio.h>
45 #include <sys/capsicum.h>
46 #include <sys/condvar.h>
47 #include <sys/conf.h>
48 #include <sys/file.h>
49 #include <sys/kernel.h>
50 #include <sys/kthread.h>
51 #include <sys/ktr.h>
52 #include <sys/lock.h>
53 #include <sys/mbuf.h>
54 #include <sys/mutex.h>
55 #include <sys/module.h>
56 #include <sys/protosw.h>
57 #include <sys/socket.h>
58 #include <sys/socketvar.h>
59 #include <sys/sysctl.h>
60 #include <sys/systm.h>
61 #include <sys/sx.h>
62 #include <sys/uio.h>
63 #include <machine/bus.h>
64 #include <vm/vm.h>
65 #include <vm/vm_page.h>
66 #include <vm/pmap.h>
67 #include <netinet/in.h>
68 #include <netinet/in_pcb.h>
69 #include <netinet/tcp.h>
70 #include <netinet/tcp_var.h>
71 #include <netinet/toecore.h>
72 
73 #include <dev/iscsi/icl.h>
74 #include <dev/iscsi/iscsi_proto.h>
75 #include <icl_conn_if.h>
76 
77 #include <cam/scsi/scsi_all.h>
78 #include <cam/scsi/scsi_da.h>
79 #include <cam/ctl/ctl_io.h>
80 #include <cam/ctl/ctl.h>
81 #include <cam/ctl/ctl_backend.h>
82 #include <cam/ctl/ctl_error.h>
83 #include <cam/ctl/ctl_frontend.h>
84 #include <cam/ctl/ctl_debug.h>
85 #include <cam/ctl/ctl_ha.h>
86 #include <cam/ctl/ctl_ioctl.h>
87 
88 #include <cam/cam.h>
89 #include <cam/cam_ccb.h>
90 #include <cam/cam_xpt.h>
91 #include <cam/cam_debug.h>
92 #include <cam/cam_sim.h>
93 #include <cam/cam_xpt_sim.h>
94 #include <cam/cam_xpt_periph.h>
95 #include <cam/cam_periph.h>
96 #include <cam/cam_compat.h>
97 #include <cam/scsi/scsi_message.h>
98 
99 #include "common/common.h"
100 #include "common/t4_regs.h"
101 #include "common/t4_tcb.h"
102 #include "tom/t4_tom.h"
103 #include "cxgbei.h"
104 
105 /*
106  * Use the page pod tag for the TT hash.
107  */
108 #define	TT_HASH(icc, tt)	(G_PPOD_TAG(tt) & (icc)->cmp_hash_mask)
109 
110 struct cxgbei_ddp_state {
111 	struct ppod_reservation prsv;
112 	struct cxgbei_cmp cmp;
113 };
114 
115 static MALLOC_DEFINE(M_CXGBEI, "cxgbei", "cxgbei(4)");
116 
117 SYSCTL_NODE(_kern_icl, OID_AUTO, cxgbei, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
118     "Chelsio iSCSI offload");
119 static int first_burst_length = 8192;
120 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, first_burst_length, CTLFLAG_RWTUN,
121     &first_burst_length, 0, "First burst length");
122 static int max_burst_length = 2 * 1024 * 1024;
123 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, max_burst_length, CTLFLAG_RWTUN,
124     &max_burst_length, 0, "Maximum burst length");
125 static int sendspace = 1048576;
126 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, sendspace, CTLFLAG_RWTUN,
127     &sendspace, 0, "Default send socket buffer size");
128 static int recvspace = 1048576;
129 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, recvspace, CTLFLAG_RWTUN,
130     &recvspace, 0, "Default receive socket buffer size");
131 
132 static volatile u_int icl_cxgbei_ncons;
133 
134 static icl_conn_new_pdu_t	icl_cxgbei_conn_new_pdu;
135 static icl_conn_pdu_data_segment_length_t
136 				    icl_cxgbei_conn_pdu_data_segment_length;
137 static icl_conn_pdu_append_bio_t	icl_cxgbei_conn_pdu_append_bio;
138 static icl_conn_pdu_append_data_t	icl_cxgbei_conn_pdu_append_data;
139 static icl_conn_pdu_get_bio_t	icl_cxgbei_conn_pdu_get_bio;
140 static icl_conn_pdu_get_data_t	icl_cxgbei_conn_pdu_get_data;
141 static icl_conn_pdu_queue_t	icl_cxgbei_conn_pdu_queue;
142 static icl_conn_pdu_queue_cb_t	icl_cxgbei_conn_pdu_queue_cb;
143 static icl_conn_handoff_t	icl_cxgbei_conn_handoff;
144 static icl_conn_free_t		icl_cxgbei_conn_free;
145 static icl_conn_close_t		icl_cxgbei_conn_close;
146 static icl_conn_task_setup_t	icl_cxgbei_conn_task_setup;
147 static icl_conn_task_done_t	icl_cxgbei_conn_task_done;
148 static icl_conn_transfer_setup_t	icl_cxgbei_conn_transfer_setup;
149 static icl_conn_transfer_done_t	icl_cxgbei_conn_transfer_done;
150 
151 static kobj_method_t icl_cxgbei_methods[] = {
152 	KOBJMETHOD(icl_conn_new_pdu, icl_cxgbei_conn_new_pdu),
153 	KOBJMETHOD(icl_conn_pdu_free, icl_cxgbei_conn_pdu_free),
154 	KOBJMETHOD(icl_conn_pdu_data_segment_length,
155 	    icl_cxgbei_conn_pdu_data_segment_length),
156 	KOBJMETHOD(icl_conn_pdu_append_bio, icl_cxgbei_conn_pdu_append_bio),
157 	KOBJMETHOD(icl_conn_pdu_append_data, icl_cxgbei_conn_pdu_append_data),
158 	KOBJMETHOD(icl_conn_pdu_get_bio, icl_cxgbei_conn_pdu_get_bio),
159 	KOBJMETHOD(icl_conn_pdu_get_data, icl_cxgbei_conn_pdu_get_data),
160 	KOBJMETHOD(icl_conn_pdu_queue, icl_cxgbei_conn_pdu_queue),
161 	KOBJMETHOD(icl_conn_pdu_queue_cb, icl_cxgbei_conn_pdu_queue_cb),
162 	KOBJMETHOD(icl_conn_handoff, icl_cxgbei_conn_handoff),
163 	KOBJMETHOD(icl_conn_free, icl_cxgbei_conn_free),
164 	KOBJMETHOD(icl_conn_close, icl_cxgbei_conn_close),
165 	KOBJMETHOD(icl_conn_task_setup, icl_cxgbei_conn_task_setup),
166 	KOBJMETHOD(icl_conn_task_done, icl_cxgbei_conn_task_done),
167 	KOBJMETHOD(icl_conn_transfer_setup, icl_cxgbei_conn_transfer_setup),
168 	KOBJMETHOD(icl_conn_transfer_done, icl_cxgbei_conn_transfer_done),
169 	{ 0, 0 }
170 };
171 
172 DEFINE_CLASS(icl_cxgbei, icl_cxgbei_methods, sizeof(struct icl_cxgbei_conn));
173 
174 void
175 icl_cxgbei_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip)
176 {
177 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
178 
179 	KASSERT(icp->ref_cnt != 0, ("freeing deleted PDU"));
180 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
181 	MPASS(ic == ip->ip_conn);
182 
183 	m_freem(ip->ip_ahs_mbuf);
184 	m_freem(ip->ip_data_mbuf);
185 	m_freem(ip->ip_bhs_mbuf);
186 
187 	KASSERT(ic != NULL || icp->ref_cnt == 1,
188 	    ("orphaned PDU has oustanding references"));
189 
190 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) != 1)
191 		return;
192 
193 	free(icp, M_CXGBEI);
194 #ifdef DIAGNOSTIC
195 	if (__predict_true(ic != NULL))
196 		refcount_release(&ic->ic_outstanding_pdus);
197 #endif
198 }
199 
200 static void
201 icl_cxgbei_pdu_call_cb(struct icl_pdu *ip)
202 {
203 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
204 
205 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
206 
207 	if (icp->cb != NULL)
208 		icp->cb(ip, icp->error);
209 #ifdef DIAGNOSTIC
210 	if (__predict_true(ip->ip_conn != NULL))
211 		refcount_release(&ip->ip_conn->ic_outstanding_pdus);
212 #endif
213 	free(icp, M_CXGBEI);
214 }
215 
216 static void
217 icl_cxgbei_pdu_done(struct icl_pdu *ip, int error)
218 {
219 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
220 
221 	if (error != 0)
222 		icp->error = error;
223 
224 	m_freem(ip->ip_ahs_mbuf);
225 	ip->ip_ahs_mbuf = NULL;
226 	m_freem(ip->ip_data_mbuf);
227 	ip->ip_data_mbuf = NULL;
228 	m_freem(ip->ip_bhs_mbuf);
229 	ip->ip_bhs_mbuf = NULL;
230 
231 	/*
232 	 * All other references to this PDU should have been dropped
233 	 * by the m_freem() of ip_data_mbuf.
234 	 */
235 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1)
236 		icl_cxgbei_pdu_call_cb(ip);
237 	else
238 		__assert_unreachable();
239 }
240 
241 static void
242 icl_cxgbei_mbuf_done(struct mbuf *mb)
243 {
244 
245 	struct icl_cxgbei_pdu *icp = (struct icl_cxgbei_pdu *)mb->m_ext.ext_arg1;
246 
247 	/*
248 	 * NB: mb_free_mext() might leave ref_cnt as 1 without
249 	 * decrementing it if it hits the fast path in the ref_cnt
250 	 * check.
251 	 */
252 	icl_cxgbei_pdu_call_cb(&icp->ip);
253 }
254 
255 struct icl_pdu *
256 icl_cxgbei_new_pdu(int flags)
257 {
258 	struct icl_cxgbei_pdu *icp;
259 	struct icl_pdu *ip;
260 	struct mbuf *m;
261 
262 	icp = malloc(sizeof(*icp), M_CXGBEI, flags | M_ZERO);
263 	if (__predict_false(icp == NULL))
264 		return (NULL);
265 
266 	icp->icp_signature = CXGBEI_PDU_SIGNATURE;
267 	icp->ref_cnt = 1;
268 	ip = &icp->ip;
269 
270 	m = m_gethdr(flags, MT_DATA);
271 	if (__predict_false(m == NULL)) {
272 		free(icp, M_CXGBEI);
273 		return (NULL);
274 	}
275 
276 	ip->ip_bhs_mbuf = m;
277 	ip->ip_bhs = mtod(m, struct iscsi_bhs *);
278 	memset(ip->ip_bhs, 0, sizeof(*ip->ip_bhs));
279 	m->m_len = sizeof(struct iscsi_bhs);
280 	m->m_pkthdr.len = m->m_len;
281 
282 	return (ip);
283 }
284 
285 void
286 icl_cxgbei_new_pdu_set_conn(struct icl_pdu *ip, struct icl_conn *ic)
287 {
288 
289 	ip->ip_conn = ic;
290 #ifdef DIAGNOSTIC
291 	refcount_acquire(&ic->ic_outstanding_pdus);
292 #endif
293 }
294 
295 /*
296  * Allocate icl_pdu with empty BHS to fill up by the caller.
297  */
298 static struct icl_pdu *
299 icl_cxgbei_conn_new_pdu(struct icl_conn *ic, int flags)
300 {
301 	struct icl_pdu *ip;
302 
303 	ip = icl_cxgbei_new_pdu(flags);
304 	if (__predict_false(ip == NULL))
305 		return (NULL);
306 	icl_cxgbei_new_pdu_set_conn(ip, ic);
307 
308 	return (ip);
309 }
310 
311 static size_t
312 icl_pdu_data_segment_length(const struct icl_pdu *request)
313 {
314 	uint32_t len = 0;
315 
316 	len += request->ip_bhs->bhs_data_segment_len[0];
317 	len <<= 8;
318 	len += request->ip_bhs->bhs_data_segment_len[1];
319 	len <<= 8;
320 	len += request->ip_bhs->bhs_data_segment_len[2];
321 
322 	return (len);
323 }
324 
325 size_t
326 icl_cxgbei_conn_pdu_data_segment_length(struct icl_conn *ic,
327     const struct icl_pdu *request)
328 {
329 
330 	return (icl_pdu_data_segment_length(request));
331 }
332 
333 static struct mbuf *
334 finalize_pdu(struct icl_cxgbei_conn *icc, struct icl_cxgbei_pdu *icp)
335 {
336 	struct icl_pdu *ip = &icp->ip;
337 	uint8_t ulp_submode, padding;
338 	struct mbuf *m, *last;
339 	struct iscsi_bhs *bhs;
340 	int data_len;
341 
342 	/*
343 	 * Fix up the data segment mbuf first.
344 	 */
345 	m = ip->ip_data_mbuf;
346 	ulp_submode = icc->ulp_submode;
347 	if (m != NULL) {
348 		last = m_last(m);
349 
350 		/*
351 		 * Round up the data segment to a 4B boundary.	Pad with 0 if
352 		 * necessary.  There will definitely be room in the mbuf.
353 		 */
354 		padding = roundup2(ip->ip_data_len, 4) - ip->ip_data_len;
355 		if (padding != 0) {
356 			MPASS(padding <= M_TRAILINGSPACE(last));
357 			bzero(mtod(last, uint8_t *) + last->m_len, padding);
358 			last->m_len += padding;
359 		}
360 	} else {
361 		MPASS(ip->ip_data_len == 0);
362 		ulp_submode &= ~ULP_CRC_DATA;
363 		padding = 0;
364 	}
365 
366 	/*
367 	 * Now the header mbuf that has the BHS.
368 	 */
369 	m = ip->ip_bhs_mbuf;
370 	MPASS(m->m_pkthdr.len == sizeof(struct iscsi_bhs));
371 	MPASS(m->m_len == sizeof(struct iscsi_bhs));
372 
373 	bhs = ip->ip_bhs;
374 	data_len = ip->ip_data_len;
375 	if (data_len > icc->ic.ic_max_send_data_segment_length) {
376 		struct iscsi_bhs_data_in *bhsdi;
377 		int flags;
378 
379 		KASSERT(padding == 0, ("%s: ISO with padding %d for icp %p",
380 		    __func__, padding, icp));
381 		switch (bhs->bhs_opcode) {
382 		case ISCSI_BHS_OPCODE_SCSI_DATA_OUT:
383 			flags = 1;
384 			break;
385 		case ISCSI_BHS_OPCODE_SCSI_DATA_IN:
386 			flags = 2;
387 			break;
388 		default:
389 			panic("invalid opcode %#x for ISO", bhs->bhs_opcode);
390 		}
391 		data_len = icc->ic.ic_max_send_data_segment_length;
392 		bhsdi = (struct iscsi_bhs_data_in *)bhs;
393 		if (bhsdi->bhsdi_flags & BHSDI_FLAGS_F) {
394 			/*
395 			 * Firmware will set F on the final PDU in the
396 			 * burst.
397 			 */
398 			flags |= CXGBE_ISO_F;
399 			bhsdi->bhsdi_flags &= ~BHSDI_FLAGS_F;
400 		}
401 		set_mbuf_iscsi_iso(m, true);
402 		set_mbuf_iscsi_iso_flags(m, flags);
403 		set_mbuf_iscsi_iso_mss(m, data_len);
404 	}
405 
406 	bhs->bhs_data_segment_len[2] = data_len;
407 	bhs->bhs_data_segment_len[1] = data_len >> 8;
408 	bhs->bhs_data_segment_len[0] = data_len >> 16;
409 
410 	/*
411 	 * Extract mbuf chain from PDU.
412 	 */
413 	m->m_pkthdr.len += ip->ip_data_len + padding;
414 	m->m_next = ip->ip_data_mbuf;
415 	set_mbuf_ulp_submode(m, ulp_submode);
416 	ip->ip_bhs_mbuf = NULL;
417 	ip->ip_data_mbuf = NULL;
418 	ip->ip_bhs = NULL;
419 
420 	/*
421 	 * Drop PDU reference on icp.  Additional references might
422 	 * still be held by zero-copy PDU buffers (ICL_NOCOPY).
423 	 */
424 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1)
425 		icl_cxgbei_pdu_call_cb(ip);
426 
427 	return (m);
428 }
429 
430 static void
431 icl_cxgbei_tx_main(void *arg)
432 {
433 	struct epoch_tracker et;
434 	struct icl_cxgbei_conn *icc = arg;
435 	struct icl_conn *ic = &icc->ic;
436 	struct toepcb *toep = icc->toep;
437 	struct socket *so = ic->ic_socket;
438 	struct inpcb *inp = sotoinpcb(so);
439 	struct icl_pdu *ip;
440 	struct mbuf *m;
441 	struct mbufq mq;
442 	STAILQ_HEAD(, icl_pdu) tx_pdus = STAILQ_HEAD_INITIALIZER(tx_pdus);
443 
444 	mbufq_init(&mq, INT_MAX);
445 
446 	ICL_CONN_LOCK(ic);
447 	while (__predict_true(!ic->ic_disconnecting)) {
448 		while (STAILQ_EMPTY(&icc->sent_pdus)) {
449 			icc->tx_active = false;
450 			mtx_sleep(&icc->tx_active, ic->ic_lock, 0, "-", 0);
451 			if (__predict_false(ic->ic_disconnecting))
452 				goto out;
453 			MPASS(icc->tx_active);
454 		}
455 
456 		STAILQ_SWAP(&icc->sent_pdus, &tx_pdus, icl_pdu);
457 		ICL_CONN_UNLOCK(ic);
458 
459 		while ((ip = STAILQ_FIRST(&tx_pdus)) != NULL) {
460 			STAILQ_REMOVE_HEAD(&tx_pdus, ip_next);
461 
462 			m = finalize_pdu(icc, ip_to_icp(ip));
463 			M_ASSERTPKTHDR(m);
464 			MPASS((m->m_pkthdr.len & 3) == 0);
465 
466 			mbufq_enqueue(&mq, m);
467 		}
468 
469 		ICL_CONN_LOCK(ic);
470 		if (__predict_false(ic->ic_disconnecting) ||
471 		    __predict_false(ic->ic_socket == NULL)) {
472 			mbufq_drain(&mq);
473 			break;
474 		}
475 
476 		CURVNET_SET(toep->vnet);
477 		NET_EPOCH_ENTER(et);
478 		INP_WLOCK(inp);
479 
480 		ICL_CONN_UNLOCK(ic);
481 		if (__predict_false(inp->inp_flags & INP_DROPPED) ||
482 		    __predict_false((toep->flags & TPF_ATTACHED) == 0)) {
483 			mbufq_drain(&mq);
484 		} else {
485 			mbufq_concat(&toep->ulp_pduq, &mq);
486 			t4_push_pdus(icc->sc, toep, 0);
487 		}
488 		INP_WUNLOCK(inp);
489 		NET_EPOCH_EXIT(et);
490 		CURVNET_RESTORE();
491 
492 		ICL_CONN_LOCK(ic);
493 	}
494 out:
495 	ICL_CONN_UNLOCK(ic);
496 
497 	kthread_exit();
498 }
499 
500 static void
501 icl_cxgbei_rx_main(void *arg)
502 {
503 	struct icl_cxgbei_conn *icc = arg;
504 	struct icl_conn *ic = &icc->ic;
505 	struct icl_pdu *ip;
506 	struct sockbuf *sb;
507 	STAILQ_HEAD(, icl_pdu) rx_pdus = STAILQ_HEAD_INITIALIZER(rx_pdus);
508 	bool cantrcvmore;
509 
510 	sb = &ic->ic_socket->so_rcv;
511 	SOCKBUF_LOCK(sb);
512 	while (__predict_true(!ic->ic_disconnecting)) {
513 		while (STAILQ_EMPTY(&icc->rcvd_pdus)) {
514 			icc->rx_active = false;
515 			mtx_sleep(&icc->rx_active, SOCKBUF_MTX(sb), 0, "-", 0);
516 			if (__predict_false(ic->ic_disconnecting))
517 				goto out;
518 			MPASS(icc->rx_active);
519 		}
520 
521 		if (__predict_false(sbused(sb)) != 0) {
522 			/*
523 			 * PDUs were received before the tid
524 			 * transitioned to ULP mode.  Convert
525 			 * them to icl_cxgbei_pdus and insert
526 			 * them into the head of rcvd_pdus.
527 			 */
528 			parse_pdus(icc, sb);
529 		}
530 		cantrcvmore = (sb->sb_state & SBS_CANTRCVMORE) != 0;
531 		MPASS(STAILQ_EMPTY(&rx_pdus));
532 		STAILQ_SWAP(&icc->rcvd_pdus, &rx_pdus, icl_pdu);
533 		SOCKBUF_UNLOCK(sb);
534 
535 		/* Hand over PDUs to ICL. */
536 		while ((ip = STAILQ_FIRST(&rx_pdus)) != NULL) {
537 			STAILQ_REMOVE_HEAD(&rx_pdus, ip_next);
538 			if (cantrcvmore)
539 				icl_cxgbei_pdu_done(ip, ENOTCONN);
540 			else
541 				ic->ic_receive(ip);
542 		}
543 
544 		SOCKBUF_LOCK(sb);
545 	}
546 out:
547 	/*
548 	 * Since ic_disconnecting is set before the SOCKBUF_MTX is
549 	 * locked in icl_cxgbei_conn_close, the loop above can exit
550 	 * before icl_cxgbei_conn_close can lock SOCKBUF_MTX and block
551 	 * waiting for the thread exit.
552 	 */
553 	while (!icc->rx_exiting)
554 		mtx_sleep(&icc->rx_active, SOCKBUF_MTX(sb), 0, "-", 0);
555 	SOCKBUF_UNLOCK(sb);
556 
557 	kthread_exit();
558 }
559 
560 static void
561 cxgbei_free_mext_pg(struct mbuf *m)
562 {
563 	struct icl_cxgbei_pdu *icp;
564 
565 	M_ASSERTEXTPG(m);
566 
567 	/*
568 	 * Nothing to do for the pages; they are owned by the PDU /
569 	 * I/O request.
570 	 */
571 
572 	/* Drop reference on the PDU. */
573 	icp = m->m_ext.ext_arg1;
574 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1)
575 		icl_cxgbei_pdu_call_cb(&icp->ip);
576 }
577 
578 static struct mbuf *
579 cxgbei_getm(size_t len, int flags)
580 {
581 	struct mbuf *m, *m0, *m_tail;
582 
583 	m_tail = m0 = NULL;
584 
585 	/* Allocate as jumbo mbufs of size MJUM16BYTES. */
586 	while (len >= MJUM16BYTES) {
587 		m = m_getjcl(M_NOWAIT, MT_DATA, 0, MJUM16BYTES);
588 		if (__predict_false(m == NULL)) {
589 			if ((flags & M_WAITOK) != 0) {
590 				/* Fall back to non-jumbo mbufs. */
591 				break;
592 			}
593 			return (NULL);
594 		}
595 		if (m0 == NULL) {
596 			m0 = m_tail = m;
597 		} else {
598 			m_tail->m_next = m;
599 			m_tail = m;
600 		}
601 		len -= MJUM16BYTES;
602 	}
603 
604 	/* Allocate mbuf chain for the remaining data. */
605 	if (len != 0) {
606 		m = m_getm2(NULL, len, flags, MT_DATA, 0);
607 		if (__predict_false(m == NULL)) {
608 			m_freem(m0);
609 			return (NULL);
610 		}
611 		if (m0 == NULL)
612 			m0 = m;
613 		else
614 			m_tail->m_next = m;
615 	}
616 
617 	return (m0);
618 }
619 
620 int
621 icl_cxgbei_conn_pdu_append_bio(struct icl_conn *ic, struct icl_pdu *ip,
622     struct bio *bp, size_t offset, size_t len, int flags)
623 {
624 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
625 	struct mbuf *m, *m_tail;
626 	vm_offset_t vaddr;
627 	size_t page_offset, todo, mtodo;
628 	boolean_t mapped;
629 	int i;
630 
631 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
632 	MPASS(ic == ip->ip_conn);
633 	KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len));
634 
635 	m_tail = ip->ip_data_mbuf;
636 	if (m_tail != NULL)
637 		for (; m_tail->m_next != NULL; m_tail = m_tail->m_next)
638 			;
639 
640 	MPASS(bp->bio_flags & BIO_UNMAPPED);
641 	if (offset < PAGE_SIZE - bp->bio_ma_offset) {
642 		page_offset = bp->bio_ma_offset + offset;
643 		i = 0;
644 	} else {
645 		offset -= PAGE_SIZE - bp->bio_ma_offset;
646 		for (i = 1; offset >= PAGE_SIZE; i++)
647 			offset -= PAGE_SIZE;
648 		page_offset = offset;
649 	}
650 
651 	if (flags & ICL_NOCOPY) {
652 		m = NULL;
653 		while (len > 0) {
654 			if (m == NULL) {
655 				m = mb_alloc_ext_pgs(flags & ~ICL_NOCOPY,
656 				    cxgbei_free_mext_pg);
657 				if (__predict_false(m == NULL))
658 					return (ENOMEM);
659 				atomic_add_int(&icp->ref_cnt, 1);
660 				m->m_ext.ext_arg1 = icp;
661 				m->m_epg_1st_off = page_offset;
662 			}
663 
664 			todo = MIN(len, PAGE_SIZE - page_offset);
665 
666 			m->m_epg_pa[m->m_epg_npgs] =
667 			    VM_PAGE_TO_PHYS(bp->bio_ma[i]);
668 			m->m_epg_npgs++;
669 			m->m_epg_last_len = todo;
670 			m->m_len += todo;
671 			m->m_ext.ext_size += PAGE_SIZE;
672 			MBUF_EXT_PGS_ASSERT_SANITY(m);
673 
674 			if (m->m_epg_npgs == MBUF_PEXT_MAX_PGS) {
675 				if (m_tail != NULL)
676 					m_tail->m_next = m;
677 				else
678 					ip->ip_data_mbuf = m;
679 				m_tail = m;
680 				ip->ip_data_len += m->m_len;
681 				m = NULL;
682 			}
683 
684 			page_offset = 0;
685 			len -= todo;
686 			i++;
687 		}
688 
689 		if (m != NULL) {
690 			if (m_tail != NULL)
691 				m_tail->m_next = m;
692 			else
693 				ip->ip_data_mbuf = m;
694 			ip->ip_data_len += m->m_len;
695 		}
696 		return (0);
697 	}
698 
699 	m = cxgbei_getm(len, flags);
700 	if (__predict_false(m == NULL))
701 		return (ENOMEM);
702 
703 	if (ip->ip_data_mbuf == NULL) {
704 		ip->ip_data_mbuf = m;
705 		ip->ip_data_len = len;
706 	} else {
707 		m_tail->m_next = m;
708 		ip->ip_data_len += len;
709 	}
710 
711 	while (len > 0) {
712 		todo = MIN(len, PAGE_SIZE - page_offset);
713 
714 		mapped = pmap_map_io_transient(bp->bio_ma + i, &vaddr, 1,
715 		    FALSE);
716 
717 		do {
718 			mtodo = min(todo, M_SIZE(m) - m->m_len);
719 			memcpy(mtod(m, char *) + m->m_len, (char *)vaddr +
720 			    page_offset, mtodo);
721 			m->m_len += mtodo;
722 			if (m->m_len == M_SIZE(m))
723 				m = m->m_next;
724 			page_offset += mtodo;
725 			todo -= mtodo;
726 		} while (todo > 0);
727 
728 		if (__predict_false(mapped))
729 			pmap_unmap_io_transient(bp->bio_ma + 1, &vaddr, 1,
730 			    FALSE);
731 
732 		page_offset = 0;
733 		len -= todo;
734 		i++;
735 	}
736 
737 	MPASS(ip->ip_data_len <= max(ic->ic_max_send_data_segment_length,
738 	    ic->ic_hw_isomax));
739 
740 	return (0);
741 }
742 
743 int
744 icl_cxgbei_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *ip,
745     const void *addr, size_t len, int flags)
746 {
747 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
748 	struct mbuf *m, *m_tail;
749 	const char *src;
750 
751 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
752 	MPASS(ic == ip->ip_conn);
753 	KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len));
754 
755 	m_tail = ip->ip_data_mbuf;
756 	if (m_tail != NULL)
757 		for (; m_tail->m_next != NULL; m_tail = m_tail->m_next)
758 			;
759 
760 	if (flags & ICL_NOCOPY) {
761 		m = m_get(flags & ~ICL_NOCOPY, MT_DATA);
762 		if (m == NULL) {
763 			ICL_WARN("failed to allocate mbuf");
764 			return (ENOMEM);
765 		}
766 
767 		m->m_flags |= M_RDONLY;
768 		m_extaddref(m, __DECONST(char *, addr), len, &icp->ref_cnt,
769 		    icl_cxgbei_mbuf_done, icp, NULL);
770 		m->m_len = len;
771 		if (ip->ip_data_mbuf == NULL) {
772 			ip->ip_data_mbuf = m;
773 			ip->ip_data_len = len;
774 		} else {
775 			m_tail->m_next = m;
776 			m_tail = m_tail->m_next;
777 			ip->ip_data_len += len;
778 		}
779 
780 		return (0);
781 	}
782 
783 	m = cxgbei_getm(len, flags);
784 	if (__predict_false(m == NULL))
785 		return (ENOMEM);
786 
787 	if (ip->ip_data_mbuf == NULL) {
788 		ip->ip_data_mbuf = m;
789 		ip->ip_data_len = len;
790 	} else {
791 		m_tail->m_next = m;
792 		ip->ip_data_len += len;
793 	}
794 	src = (const char *)addr;
795 	for (; m != NULL; m = m->m_next) {
796 		m->m_len = min(len, M_SIZE(m));
797 		memcpy(mtod(m, void *), src, m->m_len);
798 		src += m->m_len;
799 		len -= m->m_len;
800 	}
801 	MPASS(len == 0);
802 
803 	MPASS(ip->ip_data_len <= max(ic->ic_max_send_data_segment_length,
804 	    ic->ic_hw_isomax));
805 
806 	return (0);
807 }
808 
809 void
810 icl_cxgbei_conn_pdu_get_bio(struct icl_conn *ic, struct icl_pdu *ip,
811     size_t pdu_off, struct bio *bp, size_t bio_off, size_t len)
812 {
813 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
814 	vm_offset_t vaddr;
815 	size_t page_offset, todo;
816 	boolean_t mapped;
817 	int i;
818 
819 	if (icp->icp_flags & ICPF_RX_DDP)
820 		return; /* data is DDP'ed, no need to copy */
821 
822 	MPASS(bp->bio_flags & BIO_UNMAPPED);
823 	if (bio_off < PAGE_SIZE - bp->bio_ma_offset) {
824 		page_offset = bp->bio_ma_offset + bio_off;
825 		i = 0;
826 	} else {
827 		bio_off -= PAGE_SIZE - bp->bio_ma_offset;
828 		for (i = 1; bio_off >= PAGE_SIZE; i++)
829 			bio_off -= PAGE_SIZE;
830 		page_offset = bio_off;
831 	}
832 
833 	while (len > 0) {
834 		todo = MIN(len, PAGE_SIZE - page_offset);
835 
836 		mapped = pmap_map_io_transient(bp->bio_ma + i, &vaddr, 1,
837 		    FALSE);
838 		m_copydata(ip->ip_data_mbuf, pdu_off, todo, (char *)vaddr +
839 		    page_offset);
840 		if (__predict_false(mapped))
841 			pmap_unmap_io_transient(bp->bio_ma + 1, &vaddr, 1,
842 			    FALSE);
843 
844 		page_offset = 0;
845 		pdu_off += todo;
846 		len -= todo;
847 		i++;
848 	}
849 }
850 
851 void
852 icl_cxgbei_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip,
853     size_t off, void *addr, size_t len)
854 {
855 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
856 
857 	if (icp->icp_flags & ICPF_RX_DDP)
858 		return; /* data is DDP'ed, no need to copy */
859 	m_copydata(ip->ip_data_mbuf, off, len, addr);
860 }
861 
862 void
863 icl_cxgbei_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip)
864 {
865 	icl_cxgbei_conn_pdu_queue_cb(ic, ip, NULL);
866 }
867 
868 void
869 icl_cxgbei_conn_pdu_queue_cb(struct icl_conn *ic, struct icl_pdu *ip,
870 			     icl_pdu_cb cb)
871 {
872 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
873 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
874 	struct socket *so = ic->ic_socket;
875 
876 	MPASS(ic == ip->ip_conn);
877 	MPASS(ip->ip_bhs_mbuf != NULL);
878 	/* The kernel doesn't generate PDUs with AHS. */
879 	MPASS(ip->ip_ahs_mbuf == NULL && ip->ip_ahs_len == 0);
880 
881 	ICL_CONN_LOCK_ASSERT(ic);
882 
883 	icp->cb = cb;
884 
885 	/* NOTE: sowriteable without so_snd lock is a mostly harmless race. */
886 	if (ic->ic_disconnecting || so == NULL || !sowriteable(so)) {
887 		icl_cxgbei_pdu_done(ip, ENOTCONN);
888 		return;
889 	}
890 
891 	STAILQ_INSERT_TAIL(&icc->sent_pdus, ip, ip_next);
892 	if (!icc->tx_active) {
893 		icc->tx_active = true;
894 		wakeup(&icc->tx_active);
895 	}
896 }
897 
898 static struct icl_conn *
899 icl_cxgbei_new_conn(const char *name, struct mtx *lock)
900 {
901 	struct icl_cxgbei_conn *icc;
902 	struct icl_conn *ic;
903 
904 	refcount_acquire(&icl_cxgbei_ncons);
905 
906 	icc = (struct icl_cxgbei_conn *)kobj_create(&icl_cxgbei_class, M_CXGBE,
907 	    M_WAITOK | M_ZERO);
908 	icc->icc_signature = CXGBEI_CONN_SIGNATURE;
909 	STAILQ_INIT(&icc->rcvd_pdus);
910 	STAILQ_INIT(&icc->sent_pdus);
911 
912 	icc->cmp_table = hashinit(64, M_CXGBEI, &icc->cmp_hash_mask);
913 	mtx_init(&icc->cmp_lock, "cxgbei_cmp", NULL, MTX_DEF);
914 
915 	ic = &icc->ic;
916 	ic->ic_lock = lock;
917 
918 #ifdef DIAGNOSTIC
919 	refcount_init(&ic->ic_outstanding_pdus, 0);
920 #endif
921 	ic->ic_name = name;
922 	ic->ic_offload = "cxgbei";
923 	ic->ic_unmapped = true;
924 
925 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
926 
927 	return (ic);
928 }
929 
930 void
931 icl_cxgbei_conn_free(struct icl_conn *ic)
932 {
933 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
934 
935 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
936 
937 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
938 
939 	mtx_destroy(&icc->cmp_lock);
940 	hashdestroy(icc->cmp_table, M_CXGBEI, icc->cmp_hash_mask);
941 	kobj_delete((struct kobj *)icc, M_CXGBE);
942 	refcount_release(&icl_cxgbei_ncons);
943 }
944 
945 static int
946 icl_cxgbei_setsockopt(struct icl_conn *ic, struct socket *so, int sspace,
947     int rspace)
948 {
949 	struct sockopt opt;
950 	int error, one = 1, ss, rs;
951 
952 	ss = max(sendspace, sspace);
953 	rs = max(recvspace, rspace);
954 
955 	error = soreserve(so, ss, rs);
956 	if (error != 0)
957 		return (error);
958 	SOCKBUF_LOCK(&so->so_snd);
959 	so->so_snd.sb_flags |= SB_AUTOSIZE;
960 	SOCKBUF_UNLOCK(&so->so_snd);
961 	SOCKBUF_LOCK(&so->so_rcv);
962 	so->so_rcv.sb_flags |= SB_AUTOSIZE;
963 	SOCKBUF_UNLOCK(&so->so_rcv);
964 
965 	/*
966 	 * Disable Nagle.
967 	 */
968 	bzero(&opt, sizeof(opt));
969 	opt.sopt_dir = SOPT_SET;
970 	opt.sopt_level = IPPROTO_TCP;
971 	opt.sopt_name = TCP_NODELAY;
972 	opt.sopt_val = &one;
973 	opt.sopt_valsize = sizeof(one);
974 	error = sosetopt(so, &opt);
975 	if (error != 0)
976 		return (error);
977 
978 	return (0);
979 }
980 
981 /*
982  * Request/response structure used to find out the adapter offloading a socket.
983  */
984 struct find_ofld_adapter_rr {
985 	struct socket *so;
986 	struct adapter *sc;	/* result */
987 };
988 
989 static void
990 find_offload_adapter(struct adapter *sc, void *arg)
991 {
992 	struct find_ofld_adapter_rr *fa = arg;
993 	struct socket *so = fa->so;
994 	struct tom_data *td = sc->tom_softc;
995 	struct tcpcb *tp;
996 	struct inpcb *inp;
997 
998 	/* Non-TCP were filtered out earlier. */
999 	MPASS(so->so_proto->pr_protocol == IPPROTO_TCP);
1000 
1001 	if (fa->sc != NULL)
1002 		return;	/* Found already. */
1003 
1004 	if (td == NULL)
1005 		return;	/* TOE not enabled on this adapter. */
1006 
1007 	inp = sotoinpcb(so);
1008 	INP_WLOCK(inp);
1009 	if ((inp->inp_flags & INP_DROPPED) == 0) {
1010 		tp = intotcpcb(inp);
1011 		if (tp->t_flags & TF_TOE && tp->tod == &td->tod)
1012 			fa->sc = sc;	/* Found. */
1013 	}
1014 	INP_WUNLOCK(inp);
1015 }
1016 
1017 static bool
1018 is_memfree(struct adapter *sc)
1019 {
1020 	uint32_t em;
1021 
1022 	em = t4_read_reg(sc, A_MA_TARGET_MEM_ENABLE);
1023 	if ((em & F_EXT_MEM_ENABLE) != 0)
1024 		return (false);
1025 	if (is_t5(sc) && (em & F_EXT_MEM1_ENABLE) != 0)
1026 		return (false);
1027 	return (true);
1028 }
1029 
1030 /* XXXNP: move this to t4_tom. */
1031 static void
1032 send_iscsi_flowc_wr(struct adapter *sc, struct toepcb *toep, int maxlen)
1033 {
1034 	struct wrqe *wr;
1035 	struct fw_flowc_wr *flowc;
1036 	const u_int nparams = 1;
1037 	u_int flowclen;
1038 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
1039 
1040 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
1041 
1042 	wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq);
1043 	if (wr == NULL) {
1044 		/* XXX */
1045 		panic("%s: allocation failure.", __func__);
1046 	}
1047 	flowc = wrtod(wr);
1048 	memset(flowc, 0, wr->wr_len);
1049 
1050 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
1051 	    V_FW_FLOWC_WR_NPARAMS(nparams));
1052 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
1053 	    V_FW_WR_FLOWID(toep->tid));
1054 
1055 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_TXDATAPLEN_MAX;
1056 	flowc->mnemval[0].val = htobe32(maxlen);
1057 
1058 	txsd->tx_credits = howmany(flowclen, 16);
1059 	txsd->plen = 0;
1060 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
1061 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
1062 	toep->tx_credits -= txsd->tx_credits;
1063 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
1064 		toep->txsd_pidx = 0;
1065 	toep->txsd_avail--;
1066 
1067 	t4_wrq_tx(sc, wr);
1068 }
1069 
1070 static void
1071 set_ulp_mode_iscsi(struct adapter *sc, struct toepcb *toep, u_int ulp_submode)
1072 {
1073 	uint64_t val;
1074 
1075 	CTR3(KTR_CXGBE, "%s: tid %u, ULP_MODE_ISCSI, submode=%#x",
1076 	    __func__, toep->tid, ulp_submode);
1077 
1078 	val = V_TCB_ULP_TYPE(ULP_MODE_ISCSI) | V_TCB_ULP_RAW(ulp_submode);
1079 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_ULP_TYPE,
1080 	    V_TCB_ULP_TYPE(M_TCB_ULP_TYPE) | V_TCB_ULP_RAW(M_TCB_ULP_RAW), val,
1081 	    0, 0);
1082 
1083 	val = V_TF_RX_FLOW_CONTROL_DISABLE(1ULL);
1084 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, val, val, 0, 0);
1085 }
1086 
1087 /*
1088  * XXXNP: Who is responsible for cleaning up the socket if this returns with an
1089  * error?  Review all error paths.
1090  *
1091  * XXXNP: What happens to the socket's fd reference if the operation is
1092  * successful, and how does that affect the socket's life cycle?
1093  */
1094 int
1095 icl_cxgbei_conn_handoff(struct icl_conn *ic, int fd)
1096 {
1097 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
1098 	struct find_ofld_adapter_rr fa;
1099 	struct file *fp;
1100 	struct socket *so;
1101 	struct inpcb *inp;
1102 	struct tcpcb *tp;
1103 	struct toepcb *toep;
1104 	cap_rights_t rights;
1105 	u_int max_iso_payload, max_rx_pdu_len, max_tx_pdu_len;
1106 	int error, max_iso_pdus;
1107 
1108 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
1109 	ICL_CONN_LOCK_ASSERT_NOT(ic);
1110 
1111 	/*
1112 	 * Steal the socket from userland.
1113 	 */
1114 	error = fget(curthread, fd,
1115 	    cap_rights_init_one(&rights, CAP_SOCK_CLIENT), &fp);
1116 	if (error != 0)
1117 		return (error);
1118 	if (fp->f_type != DTYPE_SOCKET) {
1119 		fdrop(fp, curthread);
1120 		return (EINVAL);
1121 	}
1122 	so = fp->f_data;
1123 	if (so->so_type != SOCK_STREAM ||
1124 	    so->so_proto->pr_protocol != IPPROTO_TCP) {
1125 		fdrop(fp, curthread);
1126 		return (EINVAL);
1127 	}
1128 
1129 	ICL_CONN_LOCK(ic);
1130 	if (ic->ic_socket != NULL) {
1131 		ICL_CONN_UNLOCK(ic);
1132 		fdrop(fp, curthread);
1133 		return (EBUSY);
1134 	}
1135 	ic->ic_disconnecting = false;
1136 	ic->ic_socket = so;
1137 	fp->f_ops = &badfileops;
1138 	fp->f_data = NULL;
1139 	fdrop(fp, curthread);
1140 	ICL_CONN_UNLOCK(ic);
1141 
1142 	/* Find the adapter offloading this socket. */
1143 	fa.sc = NULL;
1144 	fa.so = so;
1145 	t4_iterate(find_offload_adapter, &fa);
1146 	if (fa.sc == NULL) {
1147 		error = EINVAL;
1148 		goto out;
1149 	}
1150 	icc->sc = fa.sc;
1151 
1152 	max_rx_pdu_len = ISCSI_BHS_SIZE + ic->ic_max_recv_data_segment_length;
1153 	max_tx_pdu_len = ISCSI_BHS_SIZE + ic->ic_max_send_data_segment_length;
1154 	if (ic->ic_header_crc32c) {
1155 		max_rx_pdu_len += ISCSI_HEADER_DIGEST_SIZE;
1156 		max_tx_pdu_len += ISCSI_HEADER_DIGEST_SIZE;
1157 	}
1158 	if (ic->ic_data_crc32c) {
1159 		max_rx_pdu_len += ISCSI_DATA_DIGEST_SIZE;
1160 		max_tx_pdu_len += ISCSI_DATA_DIGEST_SIZE;
1161 	}
1162 
1163 	inp = sotoinpcb(so);
1164 	INP_WLOCK(inp);
1165 	tp = intotcpcb(inp);
1166 	if (inp->inp_flags & INP_DROPPED) {
1167 		INP_WUNLOCK(inp);
1168 		error = ENOTCONN;
1169 		goto out;
1170 	}
1171 
1172 	/*
1173 	 * socket could not have been "unoffloaded" if here.
1174 	 */
1175 	MPASS(tp->t_flags & TF_TOE);
1176 	MPASS(tp->tod != NULL);
1177 	MPASS(tp->t_toe != NULL);
1178 	toep = tp->t_toe;
1179 	MPASS(toep->vi->adapter == icc->sc);
1180 
1181 	if (ulp_mode(toep) != ULP_MODE_NONE) {
1182 		INP_WUNLOCK(inp);
1183 		error = EINVAL;
1184 		goto out;
1185 	}
1186 
1187 	icc->toep = toep;
1188 
1189 	icc->ulp_submode = 0;
1190 	if (ic->ic_header_crc32c)
1191 		icc->ulp_submode |= ULP_CRC_HEADER;
1192 	if (ic->ic_data_crc32c)
1193 		icc->ulp_submode |= ULP_CRC_DATA;
1194 
1195 	if (icc->sc->tt.iso && chip_id(icc->sc) >= CHELSIO_T5 &&
1196 	    !is_memfree(icc->sc)) {
1197 		max_iso_payload = rounddown(CXGBEI_MAX_ISO_PAYLOAD,
1198 		    tp->t_maxseg);
1199 		max_iso_pdus = max_iso_payload / max_tx_pdu_len;
1200 		ic->ic_hw_isomax = max_iso_pdus *
1201 		    ic->ic_max_send_data_segment_length;
1202 	} else
1203 		max_iso_pdus = 1;
1204 
1205 	toep->params.ulp_mode = ULP_MODE_ISCSI;
1206 	toep->ulpcb = icc;
1207 
1208 	send_iscsi_flowc_wr(icc->sc, toep,
1209 	    roundup(max_iso_pdus * max_tx_pdu_len, tp->t_maxseg));
1210 	set_ulp_mode_iscsi(icc->sc, toep, icc->ulp_submode);
1211 	INP_WUNLOCK(inp);
1212 
1213 	error = kthread_add(icl_cxgbei_tx_main, icc, NULL, &icc->tx_thread, 0,
1214 	    0, "%stx (cxgbei)", ic->ic_name);
1215 	if (error != 0)
1216 		goto out;
1217 
1218 	error = kthread_add(icl_cxgbei_rx_main, icc, NULL, &icc->rx_thread, 0,
1219 	    0, "%srx (cxgbei)", ic->ic_name);
1220 	if (error != 0)
1221 		goto out;
1222 
1223 	error = icl_cxgbei_setsockopt(ic, so, max_tx_pdu_len, max_rx_pdu_len);
1224 out:
1225 	if (error != 0)
1226 		icl_cxgbei_conn_close(ic);
1227 	return (error);
1228 }
1229 
1230 void
1231 icl_cxgbei_conn_close(struct icl_conn *ic)
1232 {
1233 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
1234 	struct icl_pdu *ip;
1235 	struct socket *so;
1236 	struct sockbuf *sb;
1237 	struct inpcb *inp;
1238 	struct toepcb *toep = icc->toep;
1239 
1240 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
1241 	ICL_CONN_LOCK_ASSERT_NOT(ic);
1242 
1243 	ICL_CONN_LOCK(ic);
1244 	so = ic->ic_socket;
1245 	if (ic->ic_disconnecting || so == NULL) {
1246 		CTR4(KTR_CXGBE, "%s: icc %p (disconnecting = %d), so %p",
1247 		    __func__, icc, ic->ic_disconnecting, so);
1248 		ICL_CONN_UNLOCK(ic);
1249 		return;
1250 	}
1251 	ic->ic_disconnecting = true;
1252 
1253 #ifdef DIAGNOSTIC
1254 	KASSERT(ic->ic_outstanding_pdus == 0,
1255 	    ("destroying session with %d outstanding PDUs",
1256 	     ic->ic_outstanding_pdus));
1257 #endif
1258 
1259 	CTR3(KTR_CXGBE, "%s: tid %d, icc %p", __func__, toep ? toep->tid : -1,
1260 	    icc);
1261 
1262 	/*
1263 	 * Wait for the transmit thread to stop processing
1264 	 * this connection.
1265 	 */
1266 	if (icc->tx_thread != NULL) {
1267 		wakeup(&icc->tx_active);
1268 		mtx_sleep(icc->tx_thread, ic->ic_lock, 0, "conclo", 0);
1269 	}
1270 
1271 	/* Discard PDUs queued for TX. */
1272 	while (!STAILQ_EMPTY(&icc->sent_pdus)) {
1273 		ip = STAILQ_FIRST(&icc->sent_pdus);
1274 		STAILQ_REMOVE_HEAD(&icc->sent_pdus, ip_next);
1275 		icl_cxgbei_pdu_done(ip, ENOTCONN);
1276 	}
1277 	ICL_CONN_UNLOCK(ic);
1278 
1279 	inp = sotoinpcb(so);
1280 	sb = &so->so_rcv;
1281 
1282 	/*
1283 	 * Wait for the receive thread to stop processing this
1284 	 * connection.
1285 	 */
1286 	SOCKBUF_LOCK(sb);
1287 	if (icc->rx_thread != NULL) {
1288 		icc->rx_exiting = true;
1289 		wakeup(&icc->rx_active);
1290 		mtx_sleep(icc->rx_thread, SOCKBUF_MTX(sb), 0, "conclo", 0);
1291 	}
1292 
1293 	/*
1294 	 * Discard received PDUs not passed to the iSCSI layer.
1295 	 */
1296 	while (!STAILQ_EMPTY(&icc->rcvd_pdus)) {
1297 		ip = STAILQ_FIRST(&icc->rcvd_pdus);
1298 		STAILQ_REMOVE_HEAD(&icc->rcvd_pdus, ip_next);
1299 		icl_cxgbei_pdu_done(ip, ENOTCONN);
1300 	}
1301 	SOCKBUF_UNLOCK(sb);
1302 
1303 	INP_WLOCK(inp);
1304 	if (toep != NULL) {	/* NULL if connection was never offloaded. */
1305 		toep->ulpcb = NULL;
1306 
1307 		/* Discard mbufs queued for TX. */
1308 		mbufq_drain(&toep->ulp_pduq);
1309 
1310 		/*
1311 		 * Grab a reference to use when waiting for the final
1312 		 * CPL to be received.  If toep->inp is NULL, then
1313 		 * final_cpl_received() has already been called (e.g.
1314 		 * due to the peer sending a RST).
1315 		 */
1316 		if (toep->inp != NULL) {
1317 			toep = hold_toepcb(toep);
1318 			toep->flags |= TPF_WAITING_FOR_FINAL;
1319 		} else
1320 			toep = NULL;
1321 	}
1322 	INP_WUNLOCK(inp);
1323 
1324 	ICL_CONN_LOCK(ic);
1325 	ic->ic_socket = NULL;
1326 	ICL_CONN_UNLOCK(ic);
1327 
1328 	/*
1329 	 * XXXNP: we should send RST instead of FIN when PDUs held in various
1330 	 * queues were purged instead of delivered reliably but soabort isn't
1331 	 * really general purpose and wouldn't do the right thing here.
1332 	 */
1333 	soclose(so);
1334 
1335 	/*
1336 	 * Wait for the socket to fully close.  This ensures any
1337 	 * pending received data has been received (and in particular,
1338 	 * any data that would be received by DDP has been handled).
1339 	 * Callers assume that it is safe to free buffers for tasks
1340 	 * and transfers after this function returns.
1341 	 */
1342 	if (toep != NULL) {
1343 		struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep);
1344 
1345 		mtx_lock(lock);
1346 		while ((toep->flags & TPF_WAITING_FOR_FINAL) != 0)
1347 			mtx_sleep(toep, lock, PSOCK, "conclo2", 0);
1348 		mtx_unlock(lock);
1349 		free_toepcb(toep);
1350 	}
1351 }
1352 
1353 static void
1354 cxgbei_insert_cmp(struct icl_cxgbei_conn *icc, struct cxgbei_cmp *cmp,
1355     uint32_t tt)
1356 {
1357 #ifdef INVARIANTS
1358 	struct cxgbei_cmp *cmp2;
1359 #endif
1360 
1361 	cmp->tt = tt;
1362 
1363 	mtx_lock(&icc->cmp_lock);
1364 #ifdef INVARIANTS
1365 	LIST_FOREACH(cmp2, &icc->cmp_table[TT_HASH(icc, tt)], link) {
1366 		KASSERT(cmp2->tt != tt, ("%s: duplicate cmp", __func__));
1367 	}
1368 #endif
1369 	LIST_INSERT_HEAD(&icc->cmp_table[TT_HASH(icc, tt)], cmp, link);
1370 	mtx_unlock(&icc->cmp_lock);
1371 }
1372 
1373 struct cxgbei_cmp *
1374 cxgbei_find_cmp(struct icl_cxgbei_conn *icc, uint32_t tt)
1375 {
1376 	struct cxgbei_cmp *cmp;
1377 
1378 	mtx_lock(&icc->cmp_lock);
1379 	LIST_FOREACH(cmp, &icc->cmp_table[TT_HASH(icc, tt)], link) {
1380 		if (cmp->tt == tt)
1381 			break;
1382 	}
1383 	mtx_unlock(&icc->cmp_lock);
1384 	return (cmp);
1385 }
1386 
1387 static void
1388 cxgbei_rm_cmp(struct icl_cxgbei_conn *icc, struct cxgbei_cmp *cmp)
1389 {
1390 #ifdef INVARIANTS
1391 	struct cxgbei_cmp *cmp2;
1392 #endif
1393 
1394 	mtx_lock(&icc->cmp_lock);
1395 
1396 #ifdef INVARIANTS
1397 	LIST_FOREACH(cmp2, &icc->cmp_table[TT_HASH(icc, cmp->tt)], link) {
1398 		if (cmp2 == cmp)
1399 			goto found;
1400 	}
1401 	panic("%s: could not find cmp", __func__);
1402 found:
1403 #endif
1404 	LIST_REMOVE(cmp, link);
1405 	mtx_unlock(&icc->cmp_lock);
1406 }
1407 
1408 int
1409 icl_cxgbei_conn_task_setup(struct icl_conn *ic, struct icl_pdu *ip,
1410     struct ccb_scsiio *csio, uint32_t *ittp, void **arg)
1411 {
1412 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
1413 	struct toepcb *toep = icc->toep;
1414 	struct adapter *sc = icc->sc;
1415 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
1416 	struct ppod_region *pr = &ci->pr;
1417 	struct cxgbei_ddp_state *ddp;
1418 	struct ppod_reservation *prsv;
1419 	struct inpcb *inp;
1420 	struct mbufq mq;
1421 	uint32_t itt;
1422 	int rc = 0;
1423 
1424 	ICL_CONN_LOCK_ASSERT(ic);
1425 
1426 	/* This is for the offload driver's state.  Must not be set already. */
1427 	MPASS(arg != NULL);
1428 	MPASS(*arg == NULL);
1429 
1430 	if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_IN ||
1431 	    csio->dxfer_len < ci->ddp_threshold || ic->ic_disconnecting ||
1432 	    ic->ic_socket == NULL) {
1433 no_ddp:
1434 		/*
1435 		 * No DDP for this I/O.	 Allocate an ITT (based on the one
1436 		 * passed in) that cannot be a valid hardware DDP tag in the
1437 		 * iSCSI region.
1438 		 */
1439 		itt = *ittp & M_PPOD_TAG;
1440 		itt = V_PPOD_TAG(itt) | pr->pr_invalid_bit;
1441 		*ittp = htobe32(itt);
1442 		MPASS(*arg == NULL);	/* State is maintained for DDP only. */
1443 		if (rc != 0)
1444 			counter_u64_add(
1445 			    toep->ofld_rxq->rx_iscsi_ddp_setup_error, 1);
1446 		return (0);
1447 	}
1448 
1449 	/*
1450 	 * Reserve resources for DDP, update the itt that should be used in the
1451 	 * PDU, and save DDP specific state for this I/O in *arg.
1452 	 */
1453 	ddp = malloc(sizeof(*ddp), M_CXGBEI, M_NOWAIT | M_ZERO);
1454 	if (ddp == NULL) {
1455 		rc = ENOMEM;
1456 		goto no_ddp;
1457 	}
1458 	prsv = &ddp->prsv;
1459 
1460 	mbufq_init(&mq, INT_MAX);
1461 	switch (csio->ccb_h.flags & CAM_DATA_MASK) {
1462 	case CAM_DATA_BIO:
1463 		rc = t4_alloc_page_pods_for_bio(pr,
1464 		    (struct bio *)csio->data_ptr, prsv);
1465 		if (rc != 0) {
1466 			free(ddp, M_CXGBEI);
1467 			goto no_ddp;
1468 		}
1469 
1470 		rc = t4_write_page_pods_for_bio(sc, toep, prsv,
1471 		    (struct bio *)csio->data_ptr, &mq);
1472 		if (__predict_false(rc != 0)) {
1473 			mbufq_drain(&mq);
1474 			t4_free_page_pods(prsv);
1475 			free(ddp, M_CXGBEI);
1476 			goto no_ddp;
1477 		}
1478 		break;
1479 	case CAM_DATA_VADDR:
1480 		rc = t4_alloc_page_pods_for_buf(pr, (vm_offset_t)csio->data_ptr,
1481 		    csio->dxfer_len, prsv);
1482 		if (rc != 0) {
1483 			free(ddp, M_CXGBEI);
1484 			goto no_ddp;
1485 		}
1486 
1487 		rc = t4_write_page_pods_for_buf(sc, toep, prsv,
1488 		    (vm_offset_t)csio->data_ptr, csio->dxfer_len, &mq);
1489 		if (__predict_false(rc != 0)) {
1490 			mbufq_drain(&mq);
1491 			t4_free_page_pods(prsv);
1492 			free(ddp, M_CXGBEI);
1493 			goto no_ddp;
1494 		}
1495 		break;
1496 	default:
1497 		free(ddp, M_CXGBEI);
1498 		rc = EINVAL;
1499 		goto no_ddp;
1500 	}
1501 
1502 	/*
1503 	 * Do not get inp from toep->inp as the toepcb might have
1504 	 * detached already.
1505 	 */
1506 	inp = sotoinpcb(ic->ic_socket);
1507 	INP_WLOCK(inp);
1508 	if ((inp->inp_flags & INP_DROPPED) != 0) {
1509 		INP_WUNLOCK(inp);
1510 		mbufq_drain(&mq);
1511 		t4_free_page_pods(prsv);
1512 		free(ddp, M_CXGBEI);
1513 		goto no_ddp;
1514 	}
1515 	mbufq_concat(&toep->ulp_pduq, &mq);
1516 	INP_WUNLOCK(inp);
1517 
1518 	ddp->cmp.last_datasn = -1;
1519 	cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag);
1520 	*ittp = htobe32(prsv->prsv_tag);
1521 	*arg = prsv;
1522 	counter_u64_add(toep->ofld_rxq->rx_iscsi_ddp_setup_ok, 1);
1523 	return (0);
1524 }
1525 
1526 void
1527 icl_cxgbei_conn_task_done(struct icl_conn *ic, void *arg)
1528 {
1529 
1530 	if (arg != NULL) {
1531 		struct cxgbei_ddp_state *ddp = arg;
1532 
1533 		cxgbei_rm_cmp(ic_to_icc(ic), &ddp->cmp);
1534 		t4_free_page_pods(&ddp->prsv);
1535 		free(ddp, M_CXGBEI);
1536 	}
1537 }
1538 
1539 static inline bool
1540 ddp_sgl_check(struct ctl_sg_entry *sg, int entries, int xferlen)
1541 {
1542 #ifdef INVARIANTS
1543 	int total_len = 0;
1544 #endif
1545 
1546 	MPASS(entries > 0);
1547 	if (((vm_offset_t)sg[--entries].addr & 3U) != 0)
1548 		return (false);
1549 
1550 #ifdef INVARIANTS
1551 	total_len += sg[entries].len;
1552 #endif
1553 
1554 	while (--entries >= 0) {
1555 		if (((vm_offset_t)sg[entries].addr & PAGE_MASK) != 0 ||
1556 		    (sg[entries].len % PAGE_SIZE) != 0)
1557 			return (false);
1558 #ifdef INVARIANTS
1559 		total_len += sg[entries].len;
1560 #endif
1561 	}
1562 
1563 	MPASS(total_len == xferlen);
1564 	return (true);
1565 }
1566 
1567 #define io_to_ddp_state(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND2].ptr)
1568 
1569 int
1570 icl_cxgbei_conn_transfer_setup(struct icl_conn *ic, struct icl_pdu *ip,
1571     union ctl_io *io, uint32_t *tttp, void **arg)
1572 {
1573 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
1574 	struct toepcb *toep = icc->toep;
1575 	struct ctl_scsiio *ctsio = &io->scsiio;
1576 	struct adapter *sc = icc->sc;
1577 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
1578 	struct ppod_region *pr = &ci->pr;
1579 	struct cxgbei_ddp_state *ddp;
1580 	struct ppod_reservation *prsv;
1581 	struct ctl_sg_entry *sgl, sg_entry;
1582 	struct inpcb *inp;
1583 	struct mbufq mq;
1584 	int sg_entries = ctsio->kern_sg_entries;
1585 	uint32_t ttt;
1586 	int xferlen, rc = 0, alias;
1587 
1588 	/* This is for the offload driver's state.  Must not be set already. */
1589 	MPASS(arg != NULL);
1590 	MPASS(*arg == NULL);
1591 
1592 	if (ctsio->ext_data_filled == 0) {
1593 		int first_burst;
1594 #ifdef INVARIANTS
1595 		struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
1596 
1597 		MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
1598 		MPASS(ic == ip->ip_conn);
1599 		MPASS(ip->ip_bhs_mbuf != NULL);
1600 #endif
1601 		first_burst = icl_pdu_data_segment_length(ip);
1602 
1603 		/*
1604 		 * Note that ICL calls conn_transfer_setup even if the first
1605 		 * burst had everything and there's nothing left to transfer.
1606 		 *
1607 		 * NB: The CTL frontend might have provided a buffer
1608 		 * whose length (kern_data_len) is smaller than the
1609 		 * FirstBurstLength of unsolicited data.  Treat those
1610 		 * as an empty transfer.
1611 		 */
1612 		xferlen = ctsio->kern_data_len;
1613 		if (xferlen < first_burst ||
1614 		    xferlen - first_burst < ci->ddp_threshold) {
1615 no_ddp:
1616 			/*
1617 			 * No DDP for this transfer.  Allocate a TTT (based on
1618 			 * the one passed in) that cannot be a valid hardware
1619 			 * DDP tag in the iSCSI region.
1620 			 */
1621 			ttt = *tttp & M_PPOD_TAG;
1622 			ttt = V_PPOD_TAG(ttt) | pr->pr_invalid_bit;
1623 			*tttp = htobe32(ttt);
1624 			MPASS(io_to_ddp_state(io) == NULL);
1625 			if (rc != 0)
1626 				counter_u64_add(
1627 				    toep->ofld_rxq->rx_iscsi_ddp_setup_error, 1);
1628 			return (0);
1629 		}
1630 
1631 		if (sg_entries == 0) {
1632 			sgl = &sg_entry;
1633 			sgl->len = xferlen;
1634 			sgl->addr = (void *)ctsio->kern_data_ptr;
1635 			sg_entries = 1;
1636 		} else
1637 			sgl = (void *)ctsio->kern_data_ptr;
1638 
1639 		if (!ddp_sgl_check(sgl, sg_entries, xferlen))
1640 			goto no_ddp;
1641 
1642 		/*
1643 		 * Reserve resources for DDP, update the ttt that should be used
1644 		 * in the PDU, and save DDP specific state for this I/O.
1645 		 */
1646 		MPASS(io_to_ddp_state(io) == NULL);
1647 		ddp = malloc(sizeof(*ddp), M_CXGBEI, M_NOWAIT | M_ZERO);
1648 		if (ddp == NULL) {
1649 			rc = ENOMEM;
1650 			goto no_ddp;
1651 		}
1652 		prsv = &ddp->prsv;
1653 
1654 		rc = t4_alloc_page_pods_for_sgl(pr, sgl, sg_entries, prsv);
1655 		if (rc != 0) {
1656 			free(ddp, M_CXGBEI);
1657 			goto no_ddp;
1658 		}
1659 
1660 		mbufq_init(&mq, INT_MAX);
1661 		rc = t4_write_page_pods_for_sgl(sc, toep, prsv, sgl, sg_entries,
1662 		    xferlen, &mq);
1663 		if (__predict_false(rc != 0)) {
1664 			mbufq_drain(&mq);
1665 			t4_free_page_pods(prsv);
1666 			free(ddp, M_CXGBEI);
1667 			goto no_ddp;
1668 		}
1669 
1670 		/*
1671 		 * Do not get inp from toep->inp as the toepcb might
1672 		 * have detached already.
1673 		 */
1674 		ICL_CONN_LOCK(ic);
1675 		if (ic->ic_disconnecting || ic->ic_socket == NULL) {
1676 			ICL_CONN_UNLOCK(ic);
1677 			mbufq_drain(&mq);
1678 			t4_free_page_pods(prsv);
1679 			free(ddp, M_CXGBEI);
1680 			return (ECONNRESET);
1681 		}
1682 		inp = sotoinpcb(ic->ic_socket);
1683 		INP_WLOCK(inp);
1684 		ICL_CONN_UNLOCK(ic);
1685 		if ((inp->inp_flags & INP_DROPPED) != 0) {
1686 			INP_WUNLOCK(inp);
1687 			mbufq_drain(&mq);
1688 			t4_free_page_pods(prsv);
1689 			free(ddp, M_CXGBEI);
1690 			return (ECONNRESET);
1691 		}
1692 		mbufq_concat(&toep->ulp_pduq, &mq);
1693 		INP_WUNLOCK(inp);
1694 
1695 		ddp->cmp.next_buffer_offset = ctsio->kern_rel_offset +
1696 		    first_burst;
1697 		ddp->cmp.last_datasn = -1;
1698 		cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag);
1699 		*tttp = htobe32(prsv->prsv_tag);
1700 		io_to_ddp_state(io) = ddp;
1701 		*arg = ctsio;
1702 		counter_u64_add(toep->ofld_rxq->rx_iscsi_ddp_setup_ok, 1);
1703 		return (0);
1704 	}
1705 
1706 	/*
1707 	 * In the middle of an I/O.  A non-NULL page pod reservation indicates
1708 	 * that a DDP buffer is being used for the I/O.
1709 	 */
1710 	ddp = io_to_ddp_state(ctsio);
1711 	if (ddp == NULL)
1712 		goto no_ddp;
1713 	prsv = &ddp->prsv;
1714 
1715 	alias = (prsv->prsv_tag & pr->pr_alias_mask) >> pr->pr_alias_shift;
1716 	alias++;
1717 	prsv->prsv_tag &= ~pr->pr_alias_mask;
1718 	prsv->prsv_tag |= alias << pr->pr_alias_shift & pr->pr_alias_mask;
1719 
1720 	ddp->cmp.last_datasn = -1;
1721 	cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag);
1722 	*tttp = htobe32(prsv->prsv_tag);
1723 	*arg = ctsio;
1724 
1725 	return (0);
1726 }
1727 
1728 void
1729 icl_cxgbei_conn_transfer_done(struct icl_conn *ic, void *arg)
1730 {
1731 	struct ctl_scsiio *ctsio = arg;
1732 
1733 	if (ctsio != NULL) {
1734 		struct cxgbei_ddp_state *ddp;
1735 
1736 		ddp = io_to_ddp_state(ctsio);
1737 		MPASS(ddp != NULL);
1738 
1739 		cxgbei_rm_cmp(ic_to_icc(ic), &ddp->cmp);
1740 		if (ctsio->kern_data_len == ctsio->ext_data_filled ||
1741 		    ic->ic_disconnecting) {
1742 			t4_free_page_pods(&ddp->prsv);
1743 			free(ddp, M_CXGBEI);
1744 			io_to_ddp_state(ctsio) = NULL;
1745 		}
1746 	}
1747 }
1748 
1749 #ifdef COMPAT_FREEBSD13
1750 static void
1751 cxgbei_limits(struct adapter *sc, void *arg)
1752 {
1753 	struct icl_drv_limits *idl = arg;
1754 	struct cxgbei_data *ci;
1755 	int max_dsl;
1756 
1757 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4lims") != 0)
1758 		return;
1759 
1760 	if (uld_active(sc, ULD_ISCSI)) {
1761 		ci = sc->iscsi_ulp_softc;
1762 		MPASS(ci != NULL);
1763 
1764 
1765 		max_dsl = ci->max_rx_data_len;
1766 		if (idl->idl_max_recv_data_segment_length > max_dsl)
1767 			idl->idl_max_recv_data_segment_length = max_dsl;
1768 
1769 		max_dsl = ci->max_tx_data_len;
1770 		if (idl->idl_max_send_data_segment_length > max_dsl)
1771 			idl->idl_max_send_data_segment_length = max_dsl;
1772 	}
1773 
1774 	end_synchronized_op(sc, LOCK_HELD);
1775 }
1776 #endif
1777 
1778 static int
1779 cxgbei_limits_fd(struct icl_drv_limits *idl, int fd)
1780 {
1781 	struct find_ofld_adapter_rr fa;
1782 	struct file *fp;
1783 	struct socket *so;
1784 	struct adapter *sc;
1785 	struct cxgbei_data *ci;
1786 	cap_rights_t rights;
1787 	int error;
1788 
1789 	error = fget(curthread, fd,
1790 	    cap_rights_init_one(&rights, CAP_SOCK_CLIENT), &fp);
1791 	if (error != 0)
1792 		return (error);
1793 	if (fp->f_type != DTYPE_SOCKET) {
1794 		fdrop(fp, curthread);
1795 		return (EINVAL);
1796 	}
1797 	so = fp->f_data;
1798 	if (so->so_type != SOCK_STREAM ||
1799 	    so->so_proto->pr_protocol != IPPROTO_TCP) {
1800 		fdrop(fp, curthread);
1801 		return (EINVAL);
1802 	}
1803 
1804 	/* Find the adapter offloading this socket. */
1805 	fa.sc = NULL;
1806 	fa.so = so;
1807 	t4_iterate(find_offload_adapter, &fa);
1808 	if (fa.sc == NULL) {
1809 		fdrop(fp, curthread);
1810 		return (ENXIO);
1811 	}
1812 	fdrop(fp, curthread);
1813 
1814 	sc = fa.sc;
1815 	error = begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4lims");
1816 	if (error != 0)
1817 		return (error);
1818 
1819 	if (uld_active(sc, ULD_ISCSI)) {
1820 		ci = sc->iscsi_ulp_softc;
1821 		MPASS(ci != NULL);
1822 
1823 		idl->idl_max_recv_data_segment_length = ci->max_rx_data_len;
1824 		idl->idl_max_send_data_segment_length = ci->max_tx_data_len;
1825 	} else
1826 		error = ENXIO;
1827 
1828 	end_synchronized_op(sc, LOCK_HELD);
1829 
1830 	return (error);
1831 }
1832 
1833 static int
1834 icl_cxgbei_limits(struct icl_drv_limits *idl, int socket)
1835 {
1836 
1837 	/* Maximum allowed by the RFC.	cxgbei_limits will clip them. */
1838 	idl->idl_max_recv_data_segment_length = (1 << 24) - 1;
1839 	idl->idl_max_send_data_segment_length = (1 << 24) - 1;
1840 
1841 	/* These are somewhat arbitrary. */
1842 	idl->idl_max_burst_length = max_burst_length;
1843 	idl->idl_first_burst_length = first_burst_length;
1844 
1845 #ifdef COMPAT_FREEBSD13
1846 	if (socket == 0) {
1847 		t4_iterate(cxgbei_limits, idl);
1848 		return (0);
1849 	}
1850 #endif
1851 
1852 	return (cxgbei_limits_fd(idl, socket));
1853 }
1854 
1855 int
1856 icl_cxgbei_mod_load(void)
1857 {
1858 	int rc;
1859 
1860 	refcount_init(&icl_cxgbei_ncons, 0);
1861 
1862 	rc = icl_register("cxgbei", false, -100, icl_cxgbei_limits,
1863 	    icl_cxgbei_new_conn);
1864 
1865 	return (rc);
1866 }
1867 
1868 int
1869 icl_cxgbei_mod_unload(void)
1870 {
1871 
1872 	if (icl_cxgbei_ncons != 0)
1873 		return (EBUSY);
1874 
1875 	icl_unregister("cxgbei", false);
1876 
1877 	return (0);
1878 }
1879 #endif
1880