xref: /freebsd/sys/dev/iscsi/icl_soft.c (revision 65b71c48d226f3ef3eb1c4452dccc678a7d2ac96)
1 /*-
2  * Copyright (c) 2012 The FreeBSD Foundation
3  * All rights reserved.
4  *
5  * This software was developed by Edward Tomasz Napierala under sponsorship
6  * from the FreeBSD Foundation.
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 
31 /*
32  * Software implementation of iSCSI Common Layer kobj(9) interface.
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include <sys/param.h>
39 #include <sys/capsicum.h>
40 #include <sys/condvar.h>
41 #include <sys/conf.h>
42 #include <sys/file.h>
43 #include <sys/kernel.h>
44 #include <sys/kthread.h>
45 #include <sys/lock.h>
46 #include <sys/mbuf.h>
47 #include <sys/mutex.h>
48 #include <sys/module.h>
49 #include <sys/protosw.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/systm.h>
54 #include <sys/sx.h>
55 #include <sys/uio.h>
56 #include <vm/uma.h>
57 #include <netinet/in.h>
58 #include <netinet/tcp.h>
59 
60 #include <dev/iscsi/icl.h>
61 #include <dev/iscsi/iscsi_proto.h>
62 #include <icl_conn_if.h>
63 
64 static int coalesce = 1;
65 SYSCTL_INT(_kern_icl, OID_AUTO, coalesce, CTLFLAG_RWTUN,
66     &coalesce, 0, "Try to coalesce PDUs before sending");
67 static int partial_receive_len = 128 * 1024;
68 SYSCTL_INT(_kern_icl, OID_AUTO, partial_receive_len, CTLFLAG_RWTUN,
69     &partial_receive_len, 0, "Minimum read size for partially received "
70     "data segment");
71 static int sendspace = 1048576;
72 SYSCTL_INT(_kern_icl, OID_AUTO, sendspace, CTLFLAG_RWTUN,
73     &sendspace, 0, "Default send socket buffer size");
74 static int recvspace = 1048576;
75 SYSCTL_INT(_kern_icl, OID_AUTO, recvspace, CTLFLAG_RWTUN,
76     &recvspace, 0, "Default receive socket buffer size");
77 
78 static MALLOC_DEFINE(M_ICL_SOFT, "icl_soft", "iSCSI software backend");
79 static uma_zone_t icl_pdu_zone;
80 
81 static volatile u_int	icl_ncons;
82 
83 #define ICL_CONN_LOCK(X)		mtx_lock(X->ic_lock)
84 #define ICL_CONN_UNLOCK(X)		mtx_unlock(X->ic_lock)
85 #define ICL_CONN_LOCK_ASSERT(X)		mtx_assert(X->ic_lock, MA_OWNED)
86 #define ICL_CONN_LOCK_ASSERT_NOT(X)	mtx_assert(X->ic_lock, MA_NOTOWNED)
87 
88 STAILQ_HEAD(icl_pdu_stailq, icl_pdu);
89 
90 static icl_conn_new_pdu_t	icl_soft_conn_new_pdu;
91 static icl_conn_pdu_free_t	icl_soft_conn_pdu_free;
92 static icl_conn_pdu_data_segment_length_t
93 				    icl_soft_conn_pdu_data_segment_length;
94 static icl_conn_pdu_append_data_t	icl_soft_conn_pdu_append_data;
95 static icl_conn_pdu_get_data_t	icl_soft_conn_pdu_get_data;
96 static icl_conn_pdu_queue_t	icl_soft_conn_pdu_queue;
97 static icl_conn_handoff_t	icl_soft_conn_handoff;
98 static icl_conn_free_t		icl_soft_conn_free;
99 static icl_conn_close_t		icl_soft_conn_close;
100 static icl_conn_task_setup_t	icl_soft_conn_task_setup;
101 static icl_conn_task_done_t	icl_soft_conn_task_done;
102 static icl_conn_transfer_setup_t	icl_soft_conn_transfer_setup;
103 static icl_conn_transfer_done_t	icl_soft_conn_transfer_done;
104 
105 static kobj_method_t icl_soft_methods[] = {
106 	KOBJMETHOD(icl_conn_new_pdu, icl_soft_conn_new_pdu),
107 	KOBJMETHOD(icl_conn_pdu_free, icl_soft_conn_pdu_free),
108 	KOBJMETHOD(icl_conn_pdu_data_segment_length,
109 	    icl_soft_conn_pdu_data_segment_length),
110 	KOBJMETHOD(icl_conn_pdu_append_data, icl_soft_conn_pdu_append_data),
111 	KOBJMETHOD(icl_conn_pdu_get_data, icl_soft_conn_pdu_get_data),
112 	KOBJMETHOD(icl_conn_pdu_queue, icl_soft_conn_pdu_queue),
113 	KOBJMETHOD(icl_conn_handoff, icl_soft_conn_handoff),
114 	KOBJMETHOD(icl_conn_free, icl_soft_conn_free),
115 	KOBJMETHOD(icl_conn_close, icl_soft_conn_close),
116 	KOBJMETHOD(icl_conn_task_setup, icl_soft_conn_task_setup),
117 	KOBJMETHOD(icl_conn_task_done, icl_soft_conn_task_done),
118 	KOBJMETHOD(icl_conn_transfer_setup, icl_soft_conn_transfer_setup),
119 	KOBJMETHOD(icl_conn_transfer_done, icl_soft_conn_transfer_done),
120 	{ 0, 0 }
121 };
122 
123 DEFINE_CLASS(icl_soft, icl_soft_methods, sizeof(struct icl_conn));
124 
125 static void
126 icl_conn_fail(struct icl_conn *ic)
127 {
128 	if (ic->ic_socket == NULL)
129 		return;
130 
131 	/*
132 	 * XXX
133 	 */
134 	ic->ic_socket->so_error = EDOOFUS;
135 	(ic->ic_error)(ic);
136 }
137 
138 static struct mbuf *
139 icl_conn_receive(struct icl_conn *ic, size_t len)
140 {
141 	struct uio uio;
142 	struct socket *so;
143 	struct mbuf *m;
144 	int error, flags;
145 
146 	so = ic->ic_socket;
147 
148 	memset(&uio, 0, sizeof(uio));
149 	uio.uio_resid = len;
150 
151 	flags = MSG_DONTWAIT;
152 	error = soreceive(so, NULL, &uio, &m, NULL, &flags);
153 	if (error != 0) {
154 		ICL_DEBUG("soreceive error %d", error);
155 		return (NULL);
156 	}
157 	if (uio.uio_resid != 0) {
158 		m_freem(m);
159 		ICL_DEBUG("short read");
160 		return (NULL);
161 	}
162 
163 	return (m);
164 }
165 
166 static struct icl_pdu *
167 icl_pdu_new_empty(struct icl_conn *ic, int flags)
168 {
169 	struct icl_pdu *ip;
170 
171 #ifdef DIAGNOSTIC
172 	refcount_acquire(&ic->ic_outstanding_pdus);
173 #endif
174 	ip = uma_zalloc(icl_pdu_zone, flags | M_ZERO);
175 	if (ip == NULL) {
176 		ICL_WARN("failed to allocate %zd bytes", sizeof(*ip));
177 #ifdef DIAGNOSTIC
178 		refcount_release(&ic->ic_outstanding_pdus);
179 #endif
180 		return (NULL);
181 	}
182 
183 	ip->ip_conn = ic;
184 
185 	return (ip);
186 }
187 
188 static void
189 icl_pdu_free(struct icl_pdu *ip)
190 {
191 	struct icl_conn *ic;
192 
193 	ic = ip->ip_conn;
194 
195 	m_freem(ip->ip_bhs_mbuf);
196 	m_freem(ip->ip_ahs_mbuf);
197 	m_freem(ip->ip_data_mbuf);
198 	uma_zfree(icl_pdu_zone, ip);
199 #ifdef DIAGNOSTIC
200 	refcount_release(&ic->ic_outstanding_pdus);
201 #endif
202 }
203 
204 void
205 icl_soft_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip)
206 {
207 
208 	icl_pdu_free(ip);
209 }
210 
211 /*
212  * Allocate icl_pdu with empty BHS to fill up by the caller.
213  */
214 struct icl_pdu *
215 icl_soft_conn_new_pdu(struct icl_conn *ic, int flags)
216 {
217 	struct icl_pdu *ip;
218 
219 	ip = icl_pdu_new_empty(ic, flags);
220 	if (ip == NULL)
221 		return (NULL);
222 
223 	ip->ip_bhs_mbuf = m_getm2(NULL, sizeof(struct iscsi_bhs),
224 	    flags, MT_DATA, M_PKTHDR);
225 	if (ip->ip_bhs_mbuf == NULL) {
226 		ICL_WARN("failed to allocate %zd bytes", sizeof(*ip));
227 		icl_pdu_free(ip);
228 		return (NULL);
229 	}
230 	ip->ip_bhs = mtod(ip->ip_bhs_mbuf, struct iscsi_bhs *);
231 	memset(ip->ip_bhs, 0, sizeof(struct iscsi_bhs));
232 	ip->ip_bhs_mbuf->m_len = sizeof(struct iscsi_bhs);
233 
234 	return (ip);
235 }
236 
237 static int
238 icl_pdu_ahs_length(const struct icl_pdu *request)
239 {
240 
241 	return (request->ip_bhs->bhs_total_ahs_len * 4);
242 }
243 
244 static size_t
245 icl_pdu_data_segment_length(const struct icl_pdu *request)
246 {
247 	uint32_t len = 0;
248 
249 	len += request->ip_bhs->bhs_data_segment_len[0];
250 	len <<= 8;
251 	len += request->ip_bhs->bhs_data_segment_len[1];
252 	len <<= 8;
253 	len += request->ip_bhs->bhs_data_segment_len[2];
254 
255 	return (len);
256 }
257 
258 size_t
259 icl_soft_conn_pdu_data_segment_length(struct icl_conn *ic,
260     const struct icl_pdu *request)
261 {
262 
263 	return (icl_pdu_data_segment_length(request));
264 }
265 
266 static void
267 icl_pdu_set_data_segment_length(struct icl_pdu *response, uint32_t len)
268 {
269 
270 	response->ip_bhs->bhs_data_segment_len[2] = len;
271 	response->ip_bhs->bhs_data_segment_len[1] = len >> 8;
272 	response->ip_bhs->bhs_data_segment_len[0] = len >> 16;
273 }
274 
275 static size_t
276 icl_pdu_padding(const struct icl_pdu *ip)
277 {
278 
279 	if ((ip->ip_data_len % 4) != 0)
280 		return (4 - (ip->ip_data_len % 4));
281 
282 	return (0);
283 }
284 
285 static size_t
286 icl_pdu_size(const struct icl_pdu *response)
287 {
288 	size_t len;
289 
290 	KASSERT(response->ip_ahs_len == 0, ("responding with AHS"));
291 
292 	len = sizeof(struct iscsi_bhs) + response->ip_data_len +
293 	    icl_pdu_padding(response);
294 	if (response->ip_conn->ic_header_crc32c)
295 		len += ISCSI_HEADER_DIGEST_SIZE;
296 	if (response->ip_data_len != 0 && response->ip_conn->ic_data_crc32c)
297 		len += ISCSI_DATA_DIGEST_SIZE;
298 
299 	return (len);
300 }
301 
302 static int
303 icl_pdu_receive_bhs(struct icl_pdu *request, size_t *availablep)
304 {
305 	struct mbuf *m;
306 
307 	m = icl_conn_receive(request->ip_conn, sizeof(struct iscsi_bhs));
308 	if (m == NULL) {
309 		ICL_DEBUG("failed to receive BHS");
310 		return (-1);
311 	}
312 
313 	request->ip_bhs_mbuf = m_pullup(m, sizeof(struct iscsi_bhs));
314 	if (request->ip_bhs_mbuf == NULL) {
315 		ICL_WARN("m_pullup failed");
316 		return (-1);
317 	}
318 	request->ip_bhs = mtod(request->ip_bhs_mbuf, struct iscsi_bhs *);
319 
320 	/*
321 	 * XXX: For architectures with strict alignment requirements
322 	 * 	we may need to allocate ip_bhs and copy the data into it.
323 	 * 	For some reason, though, not doing this doesn't seem
324 	 * 	to cause problems; tested on sparc64.
325 	 */
326 
327 	*availablep -= sizeof(struct iscsi_bhs);
328 	return (0);
329 }
330 
331 static int
332 icl_pdu_receive_ahs(struct icl_pdu *request, size_t *availablep)
333 {
334 
335 	request->ip_ahs_len = icl_pdu_ahs_length(request);
336 	if (request->ip_ahs_len == 0)
337 		return (0);
338 
339 	request->ip_ahs_mbuf = icl_conn_receive(request->ip_conn,
340 	    request->ip_ahs_len);
341 	if (request->ip_ahs_mbuf == NULL) {
342 		ICL_DEBUG("failed to receive AHS");
343 		return (-1);
344 	}
345 
346 	*availablep -= request->ip_ahs_len;
347 	return (0);
348 }
349 
350 static uint32_t
351 icl_mbuf_to_crc32c(const struct mbuf *m0)
352 {
353 	uint32_t digest = 0xffffffff;
354 	const struct mbuf *m;
355 
356 	for (m = m0; m != NULL; m = m->m_next)
357 		digest = calculate_crc32c(digest,
358 		    mtod(m, const void *), m->m_len);
359 
360 	digest = digest ^ 0xffffffff;
361 
362 	return (digest);
363 }
364 
365 static int
366 icl_pdu_check_header_digest(struct icl_pdu *request, size_t *availablep)
367 {
368 	struct mbuf *m;
369 	uint32_t received_digest, valid_digest;
370 
371 	if (request->ip_conn->ic_header_crc32c == false)
372 		return (0);
373 
374 	m = icl_conn_receive(request->ip_conn, ISCSI_HEADER_DIGEST_SIZE);
375 	if (m == NULL) {
376 		ICL_DEBUG("failed to receive header digest");
377 		return (-1);
378 	}
379 
380 	CTASSERT(sizeof(received_digest) == ISCSI_HEADER_DIGEST_SIZE);
381 	m_copydata(m, 0, ISCSI_HEADER_DIGEST_SIZE, (void *)&received_digest);
382 	m_freem(m);
383 
384 	*availablep -= ISCSI_HEADER_DIGEST_SIZE;
385 
386 	/*
387 	 * XXX: Handle AHS.
388 	 */
389 	valid_digest = icl_mbuf_to_crc32c(request->ip_bhs_mbuf);
390 	if (received_digest != valid_digest) {
391 		ICL_WARN("header digest check failed; got 0x%x, "
392 		    "should be 0x%x", received_digest, valid_digest);
393 		return (-1);
394 	}
395 
396 	return (0);
397 }
398 
399 /*
400  * Return the number of bytes that should be waiting in the receive socket
401  * before icl_pdu_receive_data_segment() gets called.
402  */
403 static size_t
404 icl_pdu_data_segment_receive_len(const struct icl_pdu *request)
405 {
406 	size_t len;
407 
408 	len = icl_pdu_data_segment_length(request);
409 	if (len == 0)
410 		return (0);
411 
412 	/*
413 	 * Account for the parts of data segment already read from
414 	 * the socket buffer.
415 	 */
416 	KASSERT(len > request->ip_data_len, ("len <= request->ip_data_len"));
417 	len -= request->ip_data_len;
418 
419 	/*
420 	 * Don't always wait for the full data segment to be delivered
421 	 * to the socket; this might badly affect performance due to
422 	 * TCP window scaling.
423 	 */
424 	if (len > partial_receive_len) {
425 #if 0
426 		ICL_DEBUG("need %zd bytes of data, limiting to %zd",
427 		    len, partial_receive_len));
428 #endif
429 		len = partial_receive_len;
430 
431 		return (len);
432 	}
433 
434 	/*
435 	 * Account for padding.  Note that due to the way code is written,
436 	 * the icl_pdu_receive_data_segment() must always receive padding
437 	 * along with the last part of data segment, because it would be
438 	 * impossible to tell whether we've already received the full data
439 	 * segment including padding, or without it.
440 	 */
441 	if ((len % 4) != 0)
442 		len += 4 - (len % 4);
443 
444 #if 0
445 	ICL_DEBUG("need %zd bytes of data", len));
446 #endif
447 
448 	return (len);
449 }
450 
451 static int
452 icl_pdu_receive_data_segment(struct icl_pdu *request,
453     size_t *availablep, bool *more_neededp)
454 {
455 	struct icl_conn *ic;
456 	size_t len, padding = 0;
457 	struct mbuf *m;
458 
459 	ic = request->ip_conn;
460 
461 	*more_neededp = false;
462 	ic->ic_receive_len = 0;
463 
464 	len = icl_pdu_data_segment_length(request);
465 	if (len == 0)
466 		return (0);
467 
468 	if ((len % 4) != 0)
469 		padding = 4 - (len % 4);
470 
471 	/*
472 	 * Account for already received parts of data segment.
473 	 */
474 	KASSERT(len > request->ip_data_len, ("len <= request->ip_data_len"));
475 	len -= request->ip_data_len;
476 
477 	if (len + padding > *availablep) {
478 		/*
479 		 * Not enough data in the socket buffer.  Receive as much
480 		 * as we can.  Don't receive padding, since, obviously, it's
481 		 * not the end of data segment yet.
482 		 */
483 #if 0
484 		ICL_DEBUG("limited from %zd to %zd",
485 		    len + padding, *availablep - padding));
486 #endif
487 		len = *availablep - padding;
488 		*more_neededp = true;
489 		padding = 0;
490 	}
491 
492 	/*
493 	 * Must not try to receive padding without at least one byte
494 	 * of actual data segment.
495 	 */
496 	if (len > 0) {
497 		m = icl_conn_receive(request->ip_conn, len + padding);
498 		if (m == NULL) {
499 			ICL_DEBUG("failed to receive data segment");
500 			return (-1);
501 		}
502 
503 		if (request->ip_data_mbuf == NULL)
504 			request->ip_data_mbuf = m;
505 		else
506 			m_cat(request->ip_data_mbuf, m);
507 
508 		request->ip_data_len += len;
509 		*availablep -= len + padding;
510 	} else
511 		ICL_DEBUG("len 0");
512 
513 	if (*more_neededp)
514 		ic->ic_receive_len =
515 		    icl_pdu_data_segment_receive_len(request);
516 
517 	return (0);
518 }
519 
520 static int
521 icl_pdu_check_data_digest(struct icl_pdu *request, size_t *availablep)
522 {
523 	struct mbuf *m;
524 	uint32_t received_digest, valid_digest;
525 
526 	if (request->ip_conn->ic_data_crc32c == false)
527 		return (0);
528 
529 	if (request->ip_data_len == 0)
530 		return (0);
531 
532 	m = icl_conn_receive(request->ip_conn, ISCSI_DATA_DIGEST_SIZE);
533 	if (m == NULL) {
534 		ICL_DEBUG("failed to receive data digest");
535 		return (-1);
536 	}
537 
538 	CTASSERT(sizeof(received_digest) == ISCSI_DATA_DIGEST_SIZE);
539 	m_copydata(m, 0, ISCSI_DATA_DIGEST_SIZE, (void *)&received_digest);
540 	m_freem(m);
541 
542 	*availablep -= ISCSI_DATA_DIGEST_SIZE;
543 
544 	/*
545 	 * Note that ip_data_mbuf also contains padding; since digest
546 	 * calculation is supposed to include that, we iterate over
547 	 * the entire ip_data_mbuf chain, not just ip_data_len bytes of it.
548 	 */
549 	valid_digest = icl_mbuf_to_crc32c(request->ip_data_mbuf);
550 	if (received_digest != valid_digest) {
551 		ICL_WARN("data digest check failed; got 0x%x, "
552 		    "should be 0x%x", received_digest, valid_digest);
553 		return (-1);
554 	}
555 
556 	return (0);
557 }
558 
559 /*
560  * Somewhat contrary to the name, this attempts to receive only one
561  * "part" of PDU at a time; call it repeatedly until it returns non-NULL.
562  */
563 static struct icl_pdu *
564 icl_conn_receive_pdu(struct icl_conn *ic, size_t *availablep)
565 {
566 	struct icl_pdu *request;
567 	struct socket *so;
568 	size_t len;
569 	int error;
570 	bool more_needed;
571 
572 	so = ic->ic_socket;
573 
574 	if (ic->ic_receive_state == ICL_CONN_STATE_BHS) {
575 		KASSERT(ic->ic_receive_pdu == NULL,
576 		    ("ic->ic_receive_pdu != NULL"));
577 		request = icl_pdu_new_empty(ic, M_NOWAIT);
578 		if (request == NULL) {
579 			ICL_DEBUG("failed to allocate PDU; "
580 			    "dropping connection");
581 			icl_conn_fail(ic);
582 			return (NULL);
583 		}
584 		ic->ic_receive_pdu = request;
585 	} else {
586 		KASSERT(ic->ic_receive_pdu != NULL,
587 		    ("ic->ic_receive_pdu == NULL"));
588 		request = ic->ic_receive_pdu;
589 	}
590 
591 	if (*availablep < ic->ic_receive_len) {
592 #if 0
593 		ICL_DEBUG("not enough data; need %zd, "
594 		    "have %zd", ic->ic_receive_len, *availablep);
595 #endif
596 		return (NULL);
597 	}
598 
599 	switch (ic->ic_receive_state) {
600 	case ICL_CONN_STATE_BHS:
601 		//ICL_DEBUG("receiving BHS");
602 		error = icl_pdu_receive_bhs(request, availablep);
603 		if (error != 0) {
604 			ICL_DEBUG("failed to receive BHS; "
605 			    "dropping connection");
606 			break;
607 		}
608 
609 		/*
610 		 * We don't enforce any limit for AHS length;
611 		 * its length is stored in 8 bit field.
612 		 */
613 
614 		len = icl_pdu_data_segment_length(request);
615 		if (len > ic->ic_max_data_segment_length) {
616 			ICL_WARN("received data segment "
617 			    "length %zd is larger than negotiated "
618 			    "MaxDataSegmentLength %zd; "
619 			    "dropping connection",
620 			    len, ic->ic_max_data_segment_length);
621 			error = EINVAL;
622 			break;
623 		}
624 
625 		ic->ic_receive_state = ICL_CONN_STATE_AHS;
626 		ic->ic_receive_len = icl_pdu_ahs_length(request);
627 		break;
628 
629 	case ICL_CONN_STATE_AHS:
630 		//ICL_DEBUG("receiving AHS");
631 		error = icl_pdu_receive_ahs(request, availablep);
632 		if (error != 0) {
633 			ICL_DEBUG("failed to receive AHS; "
634 			    "dropping connection");
635 			break;
636 		}
637 		ic->ic_receive_state = ICL_CONN_STATE_HEADER_DIGEST;
638 		if (ic->ic_header_crc32c == false)
639 			ic->ic_receive_len = 0;
640 		else
641 			ic->ic_receive_len = ISCSI_HEADER_DIGEST_SIZE;
642 		break;
643 
644 	case ICL_CONN_STATE_HEADER_DIGEST:
645 		//ICL_DEBUG("receiving header digest");
646 		error = icl_pdu_check_header_digest(request, availablep);
647 		if (error != 0) {
648 			ICL_DEBUG("header digest failed; "
649 			    "dropping connection");
650 			break;
651 		}
652 
653 		ic->ic_receive_state = ICL_CONN_STATE_DATA;
654 		ic->ic_receive_len =
655 		    icl_pdu_data_segment_receive_len(request);
656 		break;
657 
658 	case ICL_CONN_STATE_DATA:
659 		//ICL_DEBUG("receiving data segment");
660 		error = icl_pdu_receive_data_segment(request, availablep,
661 		    &more_needed);
662 		if (error != 0) {
663 			ICL_DEBUG("failed to receive data segment;"
664 			    "dropping connection");
665 			break;
666 		}
667 
668 		if (more_needed)
669 			break;
670 
671 		ic->ic_receive_state = ICL_CONN_STATE_DATA_DIGEST;
672 		if (request->ip_data_len == 0 || ic->ic_data_crc32c == false)
673 			ic->ic_receive_len = 0;
674 		else
675 			ic->ic_receive_len = ISCSI_DATA_DIGEST_SIZE;
676 		break;
677 
678 	case ICL_CONN_STATE_DATA_DIGEST:
679 		//ICL_DEBUG("receiving data digest");
680 		error = icl_pdu_check_data_digest(request, availablep);
681 		if (error != 0) {
682 			ICL_DEBUG("data digest failed; "
683 			    "dropping connection");
684 			break;
685 		}
686 
687 		/*
688 		 * We've received complete PDU; reset the receive state machine
689 		 * and return the PDU.
690 		 */
691 		ic->ic_receive_state = ICL_CONN_STATE_BHS;
692 		ic->ic_receive_len = sizeof(struct iscsi_bhs);
693 		ic->ic_receive_pdu = NULL;
694 		return (request);
695 
696 	default:
697 		panic("invalid ic_receive_state %d\n", ic->ic_receive_state);
698 	}
699 
700 	if (error != 0) {
701 		/*
702 		 * Don't free the PDU; it's pointed to by ic->ic_receive_pdu
703 		 * and will get freed in icl_soft_conn_close().
704 		 */
705 		icl_conn_fail(ic);
706 	}
707 
708 	return (NULL);
709 }
710 
711 static void
712 icl_conn_receive_pdus(struct icl_conn *ic, size_t available)
713 {
714 	struct icl_pdu *response;
715 	struct socket *so;
716 
717 	so = ic->ic_socket;
718 
719 	/*
720 	 * This can never happen; we're careful to only mess with ic->ic_socket
721 	 * pointer when the send/receive threads are not running.
722 	 */
723 	KASSERT(so != NULL, ("NULL socket"));
724 
725 	for (;;) {
726 		if (ic->ic_disconnecting)
727 			return;
728 
729 		if (so->so_error != 0) {
730 			ICL_DEBUG("connection error %d; "
731 			    "dropping connection", so->so_error);
732 			icl_conn_fail(ic);
733 			return;
734 		}
735 
736 		/*
737 		 * Loop until we have a complete PDU or there is not enough
738 		 * data in the socket buffer.
739 		 */
740 		if (available < ic->ic_receive_len) {
741 #if 0
742 			ICL_DEBUG("not enough data; have %zd, "
743 			    "need %zd", available,
744 			    ic->ic_receive_len);
745 #endif
746 			return;
747 		}
748 
749 		response = icl_conn_receive_pdu(ic, &available);
750 		if (response == NULL)
751 			continue;
752 
753 		if (response->ip_ahs_len > 0) {
754 			ICL_WARN("received PDU with unsupported "
755 			    "AHS; opcode 0x%x; dropping connection",
756 			    response->ip_bhs->bhs_opcode);
757 			icl_pdu_free(response);
758 			icl_conn_fail(ic);
759 			return;
760 		}
761 
762 		(ic->ic_receive)(response);
763 	}
764 }
765 
766 static void
767 icl_receive_thread(void *arg)
768 {
769 	struct icl_conn *ic;
770 	size_t available;
771 	struct socket *so;
772 
773 	ic = arg;
774 	so = ic->ic_socket;
775 
776 	for (;;) {
777 		if (ic->ic_disconnecting) {
778 			//ICL_DEBUG("terminating");
779 			break;
780 		}
781 
782 		/*
783 		 * Set the low watermark, to be checked by
784 		 * soreadable() in icl_soupcall_receive()
785 		 * to avoid unneccessary wakeups until there
786 		 * is enough data received to read the PDU.
787 		 */
788 		SOCKBUF_LOCK(&so->so_rcv);
789 		available = sbavail(&so->so_rcv);
790 		if (available < ic->ic_receive_len) {
791 			so->so_rcv.sb_lowat = ic->ic_receive_len;
792 			cv_wait(&ic->ic_receive_cv, &so->so_rcv.sb_mtx);
793 		} else
794 			so->so_rcv.sb_lowat = so->so_rcv.sb_hiwat + 1;
795 		SOCKBUF_UNLOCK(&so->so_rcv);
796 
797 		icl_conn_receive_pdus(ic, available);
798 	}
799 
800 	ICL_CONN_LOCK(ic);
801 	ic->ic_receive_running = false;
802 	cv_signal(&ic->ic_send_cv);
803 	ICL_CONN_UNLOCK(ic);
804 	kthread_exit();
805 }
806 
807 static int
808 icl_soupcall_receive(struct socket *so, void *arg, int waitflag)
809 {
810 	struct icl_conn *ic;
811 
812 	if (!soreadable(so))
813 		return (SU_OK);
814 
815 	ic = arg;
816 	cv_signal(&ic->ic_receive_cv);
817 	return (SU_OK);
818 }
819 
820 static int
821 icl_pdu_finalize(struct icl_pdu *request)
822 {
823 	size_t padding, pdu_len;
824 	uint32_t digest, zero = 0;
825 	int ok;
826 	struct icl_conn *ic;
827 
828 	ic = request->ip_conn;
829 
830 	icl_pdu_set_data_segment_length(request, request->ip_data_len);
831 
832 	pdu_len = icl_pdu_size(request);
833 
834 	if (ic->ic_header_crc32c) {
835 		digest = icl_mbuf_to_crc32c(request->ip_bhs_mbuf);
836 		ok = m_append(request->ip_bhs_mbuf, sizeof(digest),
837 		    (void *)&digest);
838 		if (ok != 1) {
839 			ICL_WARN("failed to append header digest");
840 			return (1);
841 		}
842 	}
843 
844 	if (request->ip_data_len != 0) {
845 		padding = icl_pdu_padding(request);
846 		if (padding > 0) {
847 			ok = m_append(request->ip_data_mbuf, padding,
848 			    (void *)&zero);
849 			if (ok != 1) {
850 				ICL_WARN("failed to append padding");
851 				return (1);
852 			}
853 		}
854 
855 		if (ic->ic_data_crc32c) {
856 			digest = icl_mbuf_to_crc32c(request->ip_data_mbuf);
857 
858 			ok = m_append(request->ip_data_mbuf, sizeof(digest),
859 			    (void *)&digest);
860 			if (ok != 1) {
861 				ICL_WARN("failed to append data digest");
862 				return (1);
863 			}
864 		}
865 
866 		m_cat(request->ip_bhs_mbuf, request->ip_data_mbuf);
867 		request->ip_data_mbuf = NULL;
868 	}
869 
870 	request->ip_bhs_mbuf->m_pkthdr.len = pdu_len;
871 
872 	return (0);
873 }
874 
875 static void
876 icl_conn_send_pdus(struct icl_conn *ic, struct icl_pdu_stailq *queue)
877 {
878 	struct icl_pdu *request, *request2;
879 	struct socket *so;
880 	size_t available, size, size2;
881 	int coalesced, error;
882 
883 	ICL_CONN_LOCK_ASSERT_NOT(ic);
884 
885 	so = ic->ic_socket;
886 
887 	SOCKBUF_LOCK(&so->so_snd);
888 	/*
889 	 * Check how much space do we have for transmit.  We can't just
890 	 * call sosend() and retry when we get EWOULDBLOCK or EMSGSIZE,
891 	 * as it always frees the mbuf chain passed to it, even in case
892 	 * of error.
893 	 */
894 	available = sbspace(&so->so_snd);
895 
896 	/*
897 	 * Notify the socket upcall that we don't need wakeups
898 	 * for the time being.
899 	 */
900 	so->so_snd.sb_lowat = so->so_snd.sb_hiwat + 1;
901 	SOCKBUF_UNLOCK(&so->so_snd);
902 
903 	while (!STAILQ_EMPTY(queue)) {
904 		request = STAILQ_FIRST(queue);
905 		size = icl_pdu_size(request);
906 		if (available < size) {
907 
908 			/*
909 			 * Set the low watermark, to be checked by
910 			 * sowriteable() in icl_soupcall_send()
911 			 * to avoid unneccessary wakeups until there
912 			 * is enough space for the PDU to fit.
913 			 */
914 			SOCKBUF_LOCK(&so->so_snd);
915 			available = sbspace(&so->so_snd);
916 			if (available < size) {
917 #if 1
918 				ICL_DEBUG("no space to send; "
919 				    "have %zd, need %zd",
920 				    available, size);
921 #endif
922 				so->so_snd.sb_lowat = size;
923 				SOCKBUF_UNLOCK(&so->so_snd);
924 				return;
925 			}
926 			SOCKBUF_UNLOCK(&so->so_snd);
927 		}
928 		STAILQ_REMOVE_HEAD(queue, ip_next);
929 		error = icl_pdu_finalize(request);
930 		if (error != 0) {
931 			ICL_DEBUG("failed to finalize PDU; "
932 			    "dropping connection");
933 			icl_conn_fail(ic);
934 			icl_pdu_free(request);
935 			return;
936 		}
937 		if (coalesce) {
938 			coalesced = 1;
939 			for (;;) {
940 				request2 = STAILQ_FIRST(queue);
941 				if (request2 == NULL)
942 					break;
943 				size2 = icl_pdu_size(request2);
944 				if (available < size + size2)
945 					break;
946 				STAILQ_REMOVE_HEAD(queue, ip_next);
947 				error = icl_pdu_finalize(request2);
948 				if (error != 0) {
949 					ICL_DEBUG("failed to finalize PDU; "
950 					    "dropping connection");
951 					icl_conn_fail(ic);
952 					icl_pdu_free(request);
953 					icl_pdu_free(request2);
954 					return;
955 				}
956 				m_cat(request->ip_bhs_mbuf, request2->ip_bhs_mbuf);
957 				request2->ip_bhs_mbuf = NULL;
958 				request->ip_bhs_mbuf->m_pkthdr.len += size2;
959 				size += size2;
960 				STAILQ_REMOVE_AFTER(queue, request, ip_next);
961 				icl_pdu_free(request2);
962 				coalesced++;
963 			}
964 #if 0
965 			if (coalesced > 1) {
966 				ICL_DEBUG("coalesced %d PDUs into %zd bytes",
967 				    coalesced, size);
968 			}
969 #endif
970 		}
971 		available -= size;
972 		error = sosend(so, NULL, NULL, request->ip_bhs_mbuf,
973 		    NULL, MSG_DONTWAIT, curthread);
974 		request->ip_bhs_mbuf = NULL; /* Sosend consumes the mbuf. */
975 		if (error != 0) {
976 			ICL_DEBUG("failed to send PDU, error %d; "
977 			    "dropping connection", error);
978 			icl_conn_fail(ic);
979 			icl_pdu_free(request);
980 			return;
981 		}
982 		icl_pdu_free(request);
983 	}
984 }
985 
986 static void
987 icl_send_thread(void *arg)
988 {
989 	struct icl_conn *ic;
990 	struct icl_pdu_stailq queue;
991 
992 	ic = arg;
993 
994 	STAILQ_INIT(&queue);
995 
996 	ICL_CONN_LOCK(ic);
997 	for (;;) {
998 		for (;;) {
999 			/*
1000 			 * If the local queue is empty, populate it from
1001 			 * the main one.  This way the icl_conn_send_pdus()
1002 			 * can go through all the queued PDUs without holding
1003 			 * any locks.
1004 			 */
1005 			if (STAILQ_EMPTY(&queue))
1006 				STAILQ_SWAP(&ic->ic_to_send, &queue, icl_pdu);
1007 
1008 			ic->ic_check_send_space = false;
1009 			ICL_CONN_UNLOCK(ic);
1010 			icl_conn_send_pdus(ic, &queue);
1011 			ICL_CONN_LOCK(ic);
1012 
1013 			/*
1014 			 * The icl_soupcall_send() was called since the last
1015 			 * call to sbspace(); go around;
1016 			 */
1017 			if (ic->ic_check_send_space)
1018 				continue;
1019 
1020 			/*
1021 			 * Local queue is empty, but we still have PDUs
1022 			 * in the main one; go around.
1023 			 */
1024 			if (STAILQ_EMPTY(&queue) &&
1025 			    !STAILQ_EMPTY(&ic->ic_to_send))
1026 				continue;
1027 
1028 			/*
1029 			 * There might be some stuff in the local queue,
1030 			 * which didn't get sent due to not having enough send
1031 			 * space.  Wait for socket upcall.
1032 			 */
1033 			break;
1034 		}
1035 
1036 		if (ic->ic_disconnecting) {
1037 			//ICL_DEBUG("terminating");
1038 			break;
1039 		}
1040 
1041 		cv_wait(&ic->ic_send_cv, ic->ic_lock);
1042 	}
1043 
1044 	/*
1045 	 * We're exiting; move PDUs back to the main queue, so they can
1046 	 * get freed properly.  At this point ordering doesn't matter.
1047 	 */
1048 	STAILQ_CONCAT(&ic->ic_to_send, &queue);
1049 
1050 	ic->ic_send_running = false;
1051 	cv_signal(&ic->ic_send_cv);
1052 	ICL_CONN_UNLOCK(ic);
1053 	kthread_exit();
1054 }
1055 
1056 static int
1057 icl_soupcall_send(struct socket *so, void *arg, int waitflag)
1058 {
1059 	struct icl_conn *ic;
1060 
1061 	if (!sowriteable(so))
1062 		return (SU_OK);
1063 
1064 	ic = arg;
1065 
1066 	ICL_CONN_LOCK(ic);
1067 	ic->ic_check_send_space = true;
1068 	ICL_CONN_UNLOCK(ic);
1069 
1070 	cv_signal(&ic->ic_send_cv);
1071 
1072 	return (SU_OK);
1073 }
1074 
1075 static int
1076 icl_pdu_append_data(struct icl_pdu *request, const void *addr, size_t len,
1077     int flags)
1078 {
1079 	struct mbuf *mb, *newmb;
1080 	size_t copylen, off = 0;
1081 
1082 	KASSERT(len > 0, ("len == 0"));
1083 
1084 	newmb = m_getm2(NULL, len, flags, MT_DATA, M_PKTHDR);
1085 	if (newmb == NULL) {
1086 		ICL_WARN("failed to allocate mbuf for %zd bytes", len);
1087 		return (ENOMEM);
1088 	}
1089 
1090 	for (mb = newmb; mb != NULL; mb = mb->m_next) {
1091 		copylen = min(M_TRAILINGSPACE(mb), len - off);
1092 		memcpy(mtod(mb, char *), (const char *)addr + off, copylen);
1093 		mb->m_len = copylen;
1094 		off += copylen;
1095 	}
1096 	KASSERT(off == len, ("%s: off != len", __func__));
1097 
1098 	if (request->ip_data_mbuf == NULL) {
1099 		request->ip_data_mbuf = newmb;
1100 		request->ip_data_len = len;
1101 	} else {
1102 		m_cat(request->ip_data_mbuf, newmb);
1103 		request->ip_data_len += len;
1104 	}
1105 
1106 	return (0);
1107 }
1108 
1109 int
1110 icl_soft_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *request,
1111     const void *addr, size_t len, int flags)
1112 {
1113 
1114 	return (icl_pdu_append_data(request, addr, len, flags));
1115 }
1116 
1117 static void
1118 icl_pdu_get_data(struct icl_pdu *ip, size_t off, void *addr, size_t len)
1119 {
1120 
1121 	m_copydata(ip->ip_data_mbuf, off, len, addr);
1122 }
1123 
1124 void
1125 icl_soft_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip,
1126     size_t off, void *addr, size_t len)
1127 {
1128 
1129 	return (icl_pdu_get_data(ip, off, addr, len));
1130 }
1131 
1132 static void
1133 icl_pdu_queue(struct icl_pdu *ip)
1134 {
1135 	struct icl_conn *ic;
1136 
1137 	ic = ip->ip_conn;
1138 
1139 	ICL_CONN_LOCK_ASSERT(ic);
1140 
1141 	if (ic->ic_disconnecting || ic->ic_socket == NULL) {
1142 		ICL_DEBUG("icl_pdu_queue on closed connection");
1143 		icl_pdu_free(ip);
1144 		return;
1145 	}
1146 
1147 	if (!STAILQ_EMPTY(&ic->ic_to_send)) {
1148 		STAILQ_INSERT_TAIL(&ic->ic_to_send, ip, ip_next);
1149 		/*
1150 		 * If the queue is not empty, someone else had already
1151 		 * signaled the send thread; no need to do that again,
1152 		 * just return.
1153 		 */
1154 		return;
1155 	}
1156 
1157 	STAILQ_INSERT_TAIL(&ic->ic_to_send, ip, ip_next);
1158 	cv_signal(&ic->ic_send_cv);
1159 }
1160 
1161 void
1162 icl_soft_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip)
1163 {
1164 
1165 	icl_pdu_queue(ip);
1166 }
1167 
1168 static struct icl_conn *
1169 icl_soft_new_conn(const char *name, struct mtx *lock)
1170 {
1171 	struct icl_conn *ic;
1172 
1173 	refcount_acquire(&icl_ncons);
1174 
1175 	ic = (struct icl_conn *)kobj_create(&icl_soft_class, M_ICL_SOFT, M_WAITOK | M_ZERO);
1176 
1177 	STAILQ_INIT(&ic->ic_to_send);
1178 	ic->ic_lock = lock;
1179 	cv_init(&ic->ic_send_cv, "icl_tx");
1180 	cv_init(&ic->ic_receive_cv, "icl_rx");
1181 #ifdef DIAGNOSTIC
1182 	refcount_init(&ic->ic_outstanding_pdus, 0);
1183 #endif
1184 	ic->ic_max_data_segment_length = ICL_MAX_DATA_SEGMENT_LENGTH;
1185 	ic->ic_name = name;
1186 	ic->ic_offload = "None";
1187 
1188 	return (ic);
1189 }
1190 
1191 void
1192 icl_soft_conn_free(struct icl_conn *ic)
1193 {
1194 
1195 	cv_destroy(&ic->ic_send_cv);
1196 	cv_destroy(&ic->ic_receive_cv);
1197 	kobj_delete((struct kobj *)ic, M_ICL_SOFT);
1198 	refcount_release(&icl_ncons);
1199 }
1200 
1201 static int
1202 icl_conn_start(struct icl_conn *ic)
1203 {
1204 	size_t minspace;
1205 	struct sockopt opt;
1206 	int error, one = 1;
1207 
1208 	ICL_CONN_LOCK(ic);
1209 
1210 	/*
1211 	 * XXX: Ugly hack.
1212 	 */
1213 	if (ic->ic_socket == NULL) {
1214 		ICL_CONN_UNLOCK(ic);
1215 		return (EINVAL);
1216 	}
1217 
1218 	ic->ic_receive_state = ICL_CONN_STATE_BHS;
1219 	ic->ic_receive_len = sizeof(struct iscsi_bhs);
1220 	ic->ic_disconnecting = false;
1221 
1222 	ICL_CONN_UNLOCK(ic);
1223 
1224 	/*
1225 	 * For sendspace, this is required because the current code cannot
1226 	 * send a PDU in pieces; thus, the minimum buffer size is equal
1227 	 * to the maximum PDU size.  "+4" is to account for possible padding.
1228 	 *
1229 	 * What we should actually do here is to use autoscaling, but set
1230 	 * some minimal buffer size to "minspace".  I don't know a way to do
1231 	 * that, though.
1232 	 */
1233 	minspace = sizeof(struct iscsi_bhs) + ic->ic_max_data_segment_length +
1234 	    ISCSI_HEADER_DIGEST_SIZE + ISCSI_DATA_DIGEST_SIZE + 4;
1235 	if (sendspace < minspace) {
1236 		ICL_WARN("kern.icl.sendspace too low; must be at least %zd",
1237 		    minspace);
1238 		sendspace = minspace;
1239 	}
1240 	if (recvspace < minspace) {
1241 		ICL_WARN("kern.icl.recvspace too low; must be at least %zd",
1242 		    minspace);
1243 		recvspace = minspace;
1244 	}
1245 
1246 	error = soreserve(ic->ic_socket, sendspace, recvspace);
1247 	if (error != 0) {
1248 		ICL_WARN("soreserve failed with error %d", error);
1249 		icl_soft_conn_close(ic);
1250 		return (error);
1251 	}
1252 	ic->ic_socket->so_snd.sb_flags |= SB_AUTOSIZE;
1253 	ic->ic_socket->so_rcv.sb_flags |= SB_AUTOSIZE;
1254 
1255 	/*
1256 	 * Disable Nagle.
1257 	 */
1258 	bzero(&opt, sizeof(opt));
1259 	opt.sopt_dir = SOPT_SET;
1260 	opt.sopt_level = IPPROTO_TCP;
1261 	opt.sopt_name = TCP_NODELAY;
1262 	opt.sopt_val = &one;
1263 	opt.sopt_valsize = sizeof(one);
1264 	error = sosetopt(ic->ic_socket, &opt);
1265 	if (error != 0) {
1266 		ICL_WARN("disabling TCP_NODELAY failed with error %d", error);
1267 		icl_soft_conn_close(ic);
1268 		return (error);
1269 	}
1270 
1271 	/*
1272 	 * Register socket upcall, to get notified about incoming PDUs
1273 	 * and free space to send outgoing ones.
1274 	 */
1275 	SOCKBUF_LOCK(&ic->ic_socket->so_snd);
1276 	soupcall_set(ic->ic_socket, SO_SND, icl_soupcall_send, ic);
1277 	SOCKBUF_UNLOCK(&ic->ic_socket->so_snd);
1278 	SOCKBUF_LOCK(&ic->ic_socket->so_rcv);
1279 	soupcall_set(ic->ic_socket, SO_RCV, icl_soupcall_receive, ic);
1280 	SOCKBUF_UNLOCK(&ic->ic_socket->so_rcv);
1281 
1282 	/*
1283 	 * Start threads.
1284 	 */
1285 	ICL_CONN_LOCK(ic);
1286 	ic->ic_send_running = ic->ic_receive_running = true;
1287 	ICL_CONN_UNLOCK(ic);
1288 	error = kthread_add(icl_send_thread, ic, NULL, NULL, 0, 0, "%stx",
1289 	    ic->ic_name);
1290 	if (error != 0) {
1291 		ICL_WARN("kthread_add(9) failed with error %d", error);
1292 		ICL_CONN_LOCK(ic);
1293 		ic->ic_send_running = ic->ic_receive_running = false;
1294 		cv_signal(&ic->ic_send_cv);
1295 		ICL_CONN_UNLOCK(ic);
1296 		icl_soft_conn_close(ic);
1297 		return (error);
1298 	}
1299 	error = kthread_add(icl_receive_thread, ic, NULL, NULL, 0, 0, "%srx",
1300 	    ic->ic_name);
1301 	if (error != 0) {
1302 		ICL_WARN("kthread_add(9) failed with error %d", error);
1303 		ICL_CONN_LOCK(ic);
1304 		ic->ic_receive_running = false;
1305 		cv_signal(&ic->ic_send_cv);
1306 		ICL_CONN_UNLOCK(ic);
1307 		icl_soft_conn_close(ic);
1308 		return (error);
1309 	}
1310 
1311 	return (0);
1312 }
1313 
1314 int
1315 icl_soft_conn_handoff(struct icl_conn *ic, int fd)
1316 {
1317 	struct file *fp;
1318 	struct socket *so;
1319 	cap_rights_t rights;
1320 	int error;
1321 
1322 	ICL_CONN_LOCK_ASSERT_NOT(ic);
1323 
1324 	/*
1325 	 * Steal the socket from userland.
1326 	 */
1327 	error = fget(curthread, fd,
1328 	    cap_rights_init(&rights, CAP_SOCK_CLIENT), &fp);
1329 	if (error != 0)
1330 		return (error);
1331 	if (fp->f_type != DTYPE_SOCKET) {
1332 		fdrop(fp, curthread);
1333 		return (EINVAL);
1334 	}
1335 	so = fp->f_data;
1336 	if (so->so_type != SOCK_STREAM) {
1337 		fdrop(fp, curthread);
1338 		return (EINVAL);
1339 	}
1340 
1341 	ICL_CONN_LOCK(ic);
1342 
1343 	if (ic->ic_socket != NULL) {
1344 		ICL_CONN_UNLOCK(ic);
1345 		fdrop(fp, curthread);
1346 		return (EBUSY);
1347 	}
1348 
1349 	ic->ic_socket = fp->f_data;
1350 	fp->f_ops = &badfileops;
1351 	fp->f_data = NULL;
1352 	fdrop(fp, curthread);
1353 	ICL_CONN_UNLOCK(ic);
1354 
1355 	error = icl_conn_start(ic);
1356 
1357 	return (error);
1358 }
1359 
1360 void
1361 icl_soft_conn_close(struct icl_conn *ic)
1362 {
1363 	struct icl_pdu *pdu;
1364 	struct socket *so;
1365 
1366 	ICL_CONN_LOCK(ic);
1367 
1368 	/*
1369 	 * Wake up the threads, so they can properly terminate.
1370 	 */
1371 	ic->ic_disconnecting = true;
1372 	while (ic->ic_receive_running || ic->ic_send_running) {
1373 		cv_signal(&ic->ic_receive_cv);
1374 		cv_signal(&ic->ic_send_cv);
1375 		cv_wait(&ic->ic_send_cv, ic->ic_lock);
1376 	}
1377 
1378 	/* Some other thread could close the connection same time. */
1379 	so = ic->ic_socket;
1380 	if (so == NULL) {
1381 		ICL_CONN_UNLOCK(ic);
1382 		return;
1383 	}
1384 	ic->ic_socket = NULL;
1385 
1386 	/*
1387 	 * Deregister socket upcalls.
1388 	 */
1389 	ICL_CONN_UNLOCK(ic);
1390 	SOCKBUF_LOCK(&so->so_snd);
1391 	if (so->so_snd.sb_upcall != NULL)
1392 		soupcall_clear(so, SO_SND);
1393 	SOCKBUF_UNLOCK(&so->so_snd);
1394 	SOCKBUF_LOCK(&so->so_rcv);
1395 	if (so->so_rcv.sb_upcall != NULL)
1396 		soupcall_clear(so, SO_RCV);
1397 	SOCKBUF_UNLOCK(&so->so_rcv);
1398 	soclose(so);
1399 	ICL_CONN_LOCK(ic);
1400 
1401 	if (ic->ic_receive_pdu != NULL) {
1402 		//ICL_DEBUG("freeing partially received PDU");
1403 		icl_pdu_free(ic->ic_receive_pdu);
1404 		ic->ic_receive_pdu = NULL;
1405 	}
1406 
1407 	/*
1408 	 * Remove any outstanding PDUs from the send queue.
1409 	 */
1410 	while (!STAILQ_EMPTY(&ic->ic_to_send)) {
1411 		pdu = STAILQ_FIRST(&ic->ic_to_send);
1412 		STAILQ_REMOVE_HEAD(&ic->ic_to_send, ip_next);
1413 		icl_pdu_free(pdu);
1414 	}
1415 
1416 	KASSERT(STAILQ_EMPTY(&ic->ic_to_send),
1417 	    ("destroying session with non-empty send queue"));
1418 #ifdef DIAGNOSTIC
1419 	KASSERT(ic->ic_outstanding_pdus == 0,
1420 	    ("destroying session with %d outstanding PDUs",
1421 	     ic->ic_outstanding_pdus));
1422 #endif
1423 	ICL_CONN_UNLOCK(ic);
1424 }
1425 
1426 int
1427 icl_soft_conn_task_setup(struct icl_conn *ic, struct ccb_scsiio *csio,
1428     uint32_t *task_tagp, void **prvp)
1429 {
1430 
1431 	return (0);
1432 }
1433 
1434 void
1435 icl_soft_conn_task_done(struct icl_conn *ic, void *prv)
1436 {
1437 }
1438 
1439 int
1440 icl_soft_conn_transfer_setup(struct icl_conn *ic, union ctl_io *io,
1441     uint32_t *transfer_tag, void **prvp)
1442 {
1443 
1444 	return (0);
1445 }
1446 
1447 void
1448 icl_soft_conn_transfer_done(struct icl_conn *ic, void *prv)
1449 {
1450 }
1451 
1452 static int
1453 icl_soft_limits(size_t *limitp)
1454 {
1455 
1456 	*limitp = 128 * 1024;
1457 
1458 	return (0);
1459 }
1460 
1461 #ifdef ICL_KERNEL_PROXY
1462 int
1463 icl_conn_handoff_sock(struct icl_conn *ic, struct socket *so)
1464 {
1465 	int error;
1466 
1467 	ICL_CONN_LOCK_ASSERT_NOT(ic);
1468 
1469 	if (so->so_type != SOCK_STREAM)
1470 		return (EINVAL);
1471 
1472 	ICL_CONN_LOCK(ic);
1473 	if (ic->ic_socket != NULL) {
1474 		ICL_CONN_UNLOCK(ic);
1475 		return (EBUSY);
1476 	}
1477 	ic->ic_socket = so;
1478 	ICL_CONN_UNLOCK(ic);
1479 
1480 	error = icl_conn_start(ic);
1481 
1482 	return (error);
1483 }
1484 #endif /* ICL_KERNEL_PROXY */
1485 
1486 static int
1487 icl_soft_load(void)
1488 {
1489 	int error;
1490 
1491 	icl_pdu_zone = uma_zcreate("icl_pdu",
1492 	    sizeof(struct icl_pdu), NULL, NULL, NULL, NULL,
1493 	    UMA_ALIGN_PTR, 0);
1494 	refcount_init(&icl_ncons, 0);
1495 
1496 	/*
1497 	 * The reason we call this "none" is that to the user,
1498 	 * it's known as "offload driver"; "offload driver: soft"
1499 	 * doesn't make much sense.
1500 	 */
1501 	error = icl_register("none", 0, icl_soft_limits, icl_soft_new_conn);
1502 	KASSERT(error == 0, ("failed to register"));
1503 
1504 	return (error);
1505 }
1506 
1507 static int
1508 icl_soft_unload(void)
1509 {
1510 
1511 	if (icl_ncons != 0)
1512 		return (EBUSY);
1513 
1514 	icl_unregister("none");
1515 
1516 	uma_zdestroy(icl_pdu_zone);
1517 
1518 	return (0);
1519 }
1520 
1521 static int
1522 icl_soft_modevent(module_t mod, int what, void *arg)
1523 {
1524 
1525 	switch (what) {
1526 	case MOD_LOAD:
1527 		return (icl_soft_load());
1528 	case MOD_UNLOAD:
1529 		return (icl_soft_unload());
1530 	default:
1531 		return (EINVAL);
1532 	}
1533 }
1534 
1535 moduledata_t icl_soft_data = {
1536 	"icl_soft",
1537 	icl_soft_modevent,
1538 	0
1539 };
1540 
1541 DECLARE_MODULE(icl_soft, icl_soft_data, SI_SUB_DRIVERS, SI_ORDER_MIDDLE);
1542 MODULE_DEPEND(icl_soft, icl, 1, 1, 1);
1543 MODULE_VERSION(icl_soft, 1);
1544