xref: /freebsd/sys/dev/iscsi/icl_soft.c (revision 06c3fb2749bda94cb5201f81ffdb8fa6c3161b2e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2012 The FreeBSD Foundation
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  * Software implementation of iSCSI Common Layer kobj(9) interface.
34  */
35 
36 #include <sys/param.h>
37 #include <sys/bio.h>
38 #include <sys/capsicum.h>
39 #include <sys/condvar.h>
40 #include <sys/conf.h>
41 #include <sys/gsb_crc32.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 <vm/vm_page.h>
58 #include <netinet/in.h>
59 #include <netinet/tcp.h>
60 
61 #include <dev/iscsi/icl.h>
62 #include <dev/iscsi/iscsi_proto.h>
63 #include <icl_conn_if.h>
64 
65 #define ICL_CONN_STATE_BHS		1
66 #define ICL_CONN_STATE_AHS		2
67 #define ICL_CONN_STATE_HEADER_DIGEST	3
68 #define ICL_CONN_STATE_DATA		4
69 #define ICL_CONN_STATE_DATA_DIGEST	5
70 
71 struct icl_soft_conn {
72 	struct icl_conn	 ic;
73 
74 	/* soft specific stuff goes here. */
75 	STAILQ_HEAD(, icl_pdu) to_send;
76 	struct cv	 send_cv;
77 	struct cv	 receive_cv;
78 	struct icl_pdu	*receive_pdu;
79 	size_t		 receive_len;
80 	int		 receive_state;
81 	bool		 receive_running;
82 	bool		 check_send_space;
83 	bool		 send_running;
84 };
85 
86 struct icl_soft_pdu {
87 	struct icl_pdu	 ip;
88 
89 	/* soft specific stuff goes here. */
90 	u_int		 ref_cnt;
91 	icl_pdu_cb	 cb;
92 	int		 error;
93 };
94 
95 SYSCTL_NODE(_kern_icl, OID_AUTO, soft, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
96     "Software iSCSI");
97 static int coalesce = 1;
98 SYSCTL_INT(_kern_icl_soft, OID_AUTO, coalesce, CTLFLAG_RWTUN,
99     &coalesce, 0, "Try to coalesce PDUs before sending");
100 static int partial_receive_len = 256 * 1024;
101 SYSCTL_INT(_kern_icl_soft, OID_AUTO, partial_receive_len, CTLFLAG_RWTUN,
102     &partial_receive_len, 0, "Minimum read size for partially received "
103     "data segment");
104 static int max_data_segment_length = 256 * 1024;
105 SYSCTL_INT(_kern_icl_soft, OID_AUTO, max_data_segment_length, CTLFLAG_RWTUN,
106     &max_data_segment_length, 0, "Maximum data segment length");
107 static int first_burst_length = 1024 * 1024;
108 SYSCTL_INT(_kern_icl_soft, OID_AUTO, first_burst_length, CTLFLAG_RWTUN,
109     &first_burst_length, 0, "First burst length");
110 static int max_burst_length = 1024 * 1024;
111 SYSCTL_INT(_kern_icl_soft, OID_AUTO, max_burst_length, CTLFLAG_RWTUN,
112     &max_burst_length, 0, "Maximum burst length");
113 static int sendspace = 1536 * 1024;
114 SYSCTL_INT(_kern_icl_soft, OID_AUTO, sendspace, CTLFLAG_RWTUN,
115     &sendspace, 0, "Default send socket buffer size");
116 static int recvspace = 1536 * 1024;
117 SYSCTL_INT(_kern_icl_soft, OID_AUTO, recvspace, CTLFLAG_RWTUN,
118     &recvspace, 0, "Default receive socket buffer size");
119 
120 static MALLOC_DEFINE(M_ICL_SOFT, "icl_soft", "iSCSI software backend");
121 static uma_zone_t icl_soft_pdu_zone;
122 
123 static volatile u_int	icl_ncons;
124 
125 STAILQ_HEAD(icl_pdu_stailq, icl_pdu);
126 
127 static icl_conn_new_pdu_t	icl_soft_conn_new_pdu;
128 static icl_conn_pdu_free_t	icl_soft_conn_pdu_free;
129 static icl_conn_pdu_data_segment_length_t
130 				    icl_soft_conn_pdu_data_segment_length;
131 static icl_conn_pdu_append_bio_t	icl_soft_conn_pdu_append_bio;
132 static icl_conn_pdu_append_data_t	icl_soft_conn_pdu_append_data;
133 static icl_conn_pdu_get_bio_t	icl_soft_conn_pdu_get_bio;
134 static icl_conn_pdu_get_data_t	icl_soft_conn_pdu_get_data;
135 static icl_conn_pdu_queue_t	icl_soft_conn_pdu_queue;
136 static icl_conn_pdu_queue_cb_t	icl_soft_conn_pdu_queue_cb;
137 static icl_conn_handoff_t	icl_soft_conn_handoff;
138 static icl_conn_free_t		icl_soft_conn_free;
139 static icl_conn_close_t		icl_soft_conn_close;
140 static icl_conn_task_setup_t	icl_soft_conn_task_setup;
141 static icl_conn_task_done_t	icl_soft_conn_task_done;
142 static icl_conn_transfer_setup_t	icl_soft_conn_transfer_setup;
143 static icl_conn_transfer_done_t	icl_soft_conn_transfer_done;
144 #ifdef ICL_KERNEL_PROXY
145 static icl_conn_connect_t	icl_soft_conn_connect;
146 #endif
147 
148 static kobj_method_t icl_soft_methods[] = {
149 	KOBJMETHOD(icl_conn_new_pdu, icl_soft_conn_new_pdu),
150 	KOBJMETHOD(icl_conn_pdu_free, icl_soft_conn_pdu_free),
151 	KOBJMETHOD(icl_conn_pdu_data_segment_length,
152 	    icl_soft_conn_pdu_data_segment_length),
153 	KOBJMETHOD(icl_conn_pdu_append_bio, icl_soft_conn_pdu_append_bio),
154 	KOBJMETHOD(icl_conn_pdu_append_data, icl_soft_conn_pdu_append_data),
155 	KOBJMETHOD(icl_conn_pdu_get_bio, icl_soft_conn_pdu_get_bio),
156 	KOBJMETHOD(icl_conn_pdu_get_data, icl_soft_conn_pdu_get_data),
157 	KOBJMETHOD(icl_conn_pdu_queue, icl_soft_conn_pdu_queue),
158 	KOBJMETHOD(icl_conn_pdu_queue_cb, icl_soft_conn_pdu_queue_cb),
159 	KOBJMETHOD(icl_conn_handoff, icl_soft_conn_handoff),
160 	KOBJMETHOD(icl_conn_free, icl_soft_conn_free),
161 	KOBJMETHOD(icl_conn_close, icl_soft_conn_close),
162 	KOBJMETHOD(icl_conn_task_setup, icl_soft_conn_task_setup),
163 	KOBJMETHOD(icl_conn_task_done, icl_soft_conn_task_done),
164 	KOBJMETHOD(icl_conn_transfer_setup, icl_soft_conn_transfer_setup),
165 	KOBJMETHOD(icl_conn_transfer_done, icl_soft_conn_transfer_done),
166 #ifdef ICL_KERNEL_PROXY
167 	KOBJMETHOD(icl_conn_connect, icl_soft_conn_connect),
168 #endif
169 	{ 0, 0 }
170 };
171 
172 DEFINE_CLASS(icl_soft, icl_soft_methods, sizeof(struct icl_soft_conn));
173 
174 static void
175 icl_conn_fail(struct icl_conn *ic)
176 {
177 	if (ic->ic_socket == NULL)
178 		return;
179 
180 	/*
181 	 * XXX
182 	 */
183 	ic->ic_socket->so_error = EDOOFUS;
184 	(ic->ic_error)(ic);
185 }
186 
187 static void
188 icl_soft_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip)
189 {
190 	struct icl_soft_pdu *isp = (struct icl_soft_pdu *)ip;
191 
192 	KASSERT(isp->ref_cnt == 0, ("freeing active PDU"));
193 	m_freem(ip->ip_bhs_mbuf);
194 	m_freem(ip->ip_ahs_mbuf);
195 	m_freem(ip->ip_data_mbuf);
196 	uma_zfree(icl_soft_pdu_zone, isp);
197 #ifdef DIAGNOSTIC
198 	refcount_release(&ic->ic_outstanding_pdus);
199 #endif
200 }
201 
202 static void
203 icl_soft_pdu_call_cb(struct icl_pdu *ip)
204 {
205 	struct icl_soft_pdu *isp = (struct icl_soft_pdu *)ip;
206 
207 	if (isp->cb != NULL)
208 		isp->cb(ip, isp->error);
209 #ifdef DIAGNOSTIC
210 	refcount_release(&ip->ip_conn->ic_outstanding_pdus);
211 #endif
212 	uma_zfree(icl_soft_pdu_zone, isp);
213 }
214 
215 static void
216 icl_soft_pdu_done(struct icl_pdu *ip, int error)
217 {
218 	struct icl_soft_pdu *isp = (struct icl_soft_pdu *)ip;
219 
220 	if (error != 0)
221 		isp->error = error;
222 
223 	m_freem(ip->ip_bhs_mbuf);
224 	ip->ip_bhs_mbuf = NULL;
225 	m_freem(ip->ip_ahs_mbuf);
226 	ip->ip_ahs_mbuf = NULL;
227 	m_freem(ip->ip_data_mbuf);
228 	ip->ip_data_mbuf = NULL;
229 
230 	if (atomic_fetchadd_int(&isp->ref_cnt, -1) == 1)
231 		icl_soft_pdu_call_cb(ip);
232 }
233 
234 static void
235 icl_soft_mbuf_done(struct mbuf *mb)
236 {
237 	struct icl_soft_pdu *isp = (struct icl_soft_pdu *)mb->m_ext.ext_arg1;
238 
239 	icl_soft_pdu_call_cb(&isp->ip);
240 }
241 
242 /*
243  * Allocate icl_pdu with empty BHS to fill up by the caller.
244  */
245 struct icl_pdu *
246 icl_soft_conn_new_pdu(struct icl_conn *ic, int flags)
247 {
248 	struct icl_soft_pdu *isp;
249 	struct icl_pdu *ip;
250 
251 #ifdef DIAGNOSTIC
252 	refcount_acquire(&ic->ic_outstanding_pdus);
253 #endif
254 	isp = uma_zalloc(icl_soft_pdu_zone, flags | M_ZERO);
255 	if (isp == NULL) {
256 		ICL_WARN("failed to allocate soft PDU");
257 #ifdef DIAGNOSTIC
258 		refcount_release(&ic->ic_outstanding_pdus);
259 #endif
260 		return (NULL);
261 	}
262 	ip = &isp->ip;
263 	ip->ip_conn = ic;
264 
265 	CTASSERT(sizeof(struct iscsi_bhs) <= MHLEN);
266 	ip->ip_bhs_mbuf = m_gethdr(flags, MT_DATA);
267 	if (ip->ip_bhs_mbuf == NULL) {
268 		ICL_WARN("failed to allocate BHS mbuf");
269 		icl_soft_conn_pdu_free(ic, ip);
270 		return (NULL);
271 	}
272 	ip->ip_bhs = mtod(ip->ip_bhs_mbuf, struct iscsi_bhs *);
273 	memset(ip->ip_bhs, 0, sizeof(struct iscsi_bhs));
274 	ip->ip_bhs_mbuf->m_len = sizeof(struct iscsi_bhs);
275 
276 	return (ip);
277 }
278 
279 static int
280 icl_pdu_ahs_length(const struct icl_pdu *request)
281 {
282 
283 	return (request->ip_bhs->bhs_total_ahs_len * 4);
284 }
285 
286 static size_t
287 icl_pdu_data_segment_length(const struct icl_pdu *request)
288 {
289 	uint32_t len = 0;
290 
291 	len += request->ip_bhs->bhs_data_segment_len[0];
292 	len <<= 8;
293 	len += request->ip_bhs->bhs_data_segment_len[1];
294 	len <<= 8;
295 	len += request->ip_bhs->bhs_data_segment_len[2];
296 
297 	return (len);
298 }
299 
300 size_t
301 icl_soft_conn_pdu_data_segment_length(struct icl_conn *ic,
302     const struct icl_pdu *request)
303 {
304 
305 	return (icl_pdu_data_segment_length(request));
306 }
307 
308 static void
309 icl_pdu_set_data_segment_length(struct icl_pdu *response, uint32_t len)
310 {
311 
312 	response->ip_bhs->bhs_data_segment_len[2] = len;
313 	response->ip_bhs->bhs_data_segment_len[1] = len >> 8;
314 	response->ip_bhs->bhs_data_segment_len[0] = len >> 16;
315 }
316 
317 static size_t
318 icl_pdu_padding(const struct icl_pdu *ip)
319 {
320 
321 	if ((ip->ip_data_len % 4) != 0)
322 		return (4 - (ip->ip_data_len % 4));
323 
324 	return (0);
325 }
326 
327 static size_t
328 icl_pdu_size(const struct icl_pdu *response)
329 {
330 	size_t len;
331 
332 	KASSERT(response->ip_ahs_len == 0, ("responding with AHS"));
333 
334 	len = sizeof(struct iscsi_bhs) + response->ip_data_len +
335 	    icl_pdu_padding(response);
336 	if (response->ip_conn->ic_header_crc32c)
337 		len += ISCSI_HEADER_DIGEST_SIZE;
338 	if (response->ip_data_len != 0 && response->ip_conn->ic_data_crc32c)
339 		len += ISCSI_DATA_DIGEST_SIZE;
340 
341 	return (len);
342 }
343 
344 static void
345 icl_soft_receive_buf(struct mbuf **r, size_t *rs, void *buf, size_t s)
346 {
347 
348 	m_copydata(*r, 0, s, buf);
349 	m_adj(*r, s);
350 	while ((*r) != NULL && (*r)->m_len == 0)
351 		*r = m_free(*r);
352 	*rs -= s;
353 }
354 
355 static void
356 icl_pdu_receive_ahs(struct icl_pdu *request, struct mbuf **r, size_t *rs)
357 {
358 
359 	request->ip_ahs_len = icl_pdu_ahs_length(request);
360 	if (request->ip_ahs_len == 0)
361 		return;
362 
363 	request->ip_ahs_mbuf = *r;
364 	*r = m_split(request->ip_ahs_mbuf, request->ip_ahs_len, M_WAITOK);
365 	*rs -= request->ip_ahs_len;
366 }
367 
368 static int
369 mbuf_crc32c_helper(void *arg, void *data, u_int len)
370 {
371 	uint32_t *digestp = arg;
372 
373 	*digestp = calculate_crc32c(*digestp, data, len);
374 	return (0);
375 }
376 
377 static uint32_t
378 icl_mbuf_to_crc32c(struct mbuf *m0, size_t len)
379 {
380 	uint32_t digest = 0xffffffff;
381 
382 	m_apply(m0, 0, len, mbuf_crc32c_helper, &digest);
383 	digest = digest ^ 0xffffffff;
384 
385 	return (digest);
386 }
387 
388 static int
389 icl_pdu_check_header_digest(struct icl_pdu *request, struct mbuf **r, size_t *rs)
390 {
391 	uint32_t received_digest, valid_digest;
392 
393 	if (request->ip_conn->ic_header_crc32c == false)
394 		return (0);
395 
396 	CTASSERT(sizeof(received_digest) == ISCSI_HEADER_DIGEST_SIZE);
397 	icl_soft_receive_buf(r, rs, &received_digest, ISCSI_HEADER_DIGEST_SIZE);
398 
399 	/* Temporary attach AHS to BHS to calculate header digest. */
400 	request->ip_bhs_mbuf->m_next = request->ip_ahs_mbuf;
401 	valid_digest = icl_mbuf_to_crc32c(request->ip_bhs_mbuf, ISCSI_BHS_SIZE);
402 	request->ip_bhs_mbuf->m_next = NULL;
403 	if (received_digest != valid_digest) {
404 		ICL_WARN("header digest check failed; got 0x%x, "
405 		    "should be 0x%x", received_digest, valid_digest);
406 		return (-1);
407 	}
408 
409 	return (0);
410 }
411 
412 /*
413  * Return the number of bytes that should be waiting in the receive socket
414  * before icl_pdu_receive_data_segment() gets called.
415  */
416 static size_t
417 icl_pdu_data_segment_receive_len(const struct icl_pdu *request)
418 {
419 	size_t len;
420 
421 	len = icl_pdu_data_segment_length(request);
422 	if (len == 0)
423 		return (0);
424 
425 	/*
426 	 * Account for the parts of data segment already read from
427 	 * the socket buffer.
428 	 */
429 	KASSERT(len > request->ip_data_len, ("len <= request->ip_data_len"));
430 	len -= request->ip_data_len;
431 
432 	/*
433 	 * Don't always wait for the full data segment to be delivered
434 	 * to the socket; this might badly affect performance due to
435 	 * TCP window scaling.
436 	 */
437 	if (len > partial_receive_len) {
438 #if 0
439 		ICL_DEBUG("need %zd bytes of data, limiting to %zd",
440 		    len, partial_receive_len));
441 #endif
442 		len = partial_receive_len;
443 
444 		return (len);
445 	}
446 
447 	/*
448 	 * Account for padding.  Note that due to the way code is written,
449 	 * the icl_pdu_receive_data_segment() must always receive padding
450 	 * along with the last part of data segment, because it would be
451 	 * impossible to tell whether we've already received the full data
452 	 * segment including padding, or without it.
453 	 */
454 	if ((len % 4) != 0)
455 		len += 4 - (len % 4);
456 
457 #if 0
458 	ICL_DEBUG("need %zd bytes of data", len));
459 #endif
460 
461 	return (len);
462 }
463 
464 static int
465 icl_pdu_receive_data_segment(struct icl_pdu *request, struct mbuf **r,
466     size_t *rs, bool *more_neededp)
467 {
468 	struct icl_soft_conn *isc;
469 	size_t len, padding = 0;
470 	struct mbuf *m;
471 
472 	isc = (struct icl_soft_conn *)request->ip_conn;
473 
474 	*more_neededp = false;
475 	isc->receive_len = 0;
476 
477 	len = icl_pdu_data_segment_length(request);
478 	if (len == 0)
479 		return (0);
480 
481 	if ((len % 4) != 0)
482 		padding = 4 - (len % 4);
483 
484 	/*
485 	 * Account for already received parts of data segment.
486 	 */
487 	KASSERT(len > request->ip_data_len, ("len <= request->ip_data_len"));
488 	len -= request->ip_data_len;
489 
490 	if (len + padding > *rs) {
491 		/*
492 		 * Not enough data in the socket buffer.  Receive as much
493 		 * as we can.  Don't receive padding, since, obviously, it's
494 		 * not the end of data segment yet.
495 		 */
496 #if 0
497 		ICL_DEBUG("limited from %zd to %zd",
498 		    len + padding, *rs - padding));
499 #endif
500 		len = *rs - padding;
501 		*more_neededp = true;
502 		padding = 0;
503 	}
504 
505 	/*
506 	 * Must not try to receive padding without at least one byte
507 	 * of actual data segment.
508 	 */
509 	if (len > 0) {
510 		m = *r;
511 		*r = m_split(m, len + padding, M_WAITOK);
512 		*rs -= len + padding;
513 
514 		if (request->ip_data_mbuf == NULL)
515 			request->ip_data_mbuf = m;
516 		else
517 			m_cat(request->ip_data_mbuf, m);
518 
519 		request->ip_data_len += len;
520 	} else
521 		ICL_DEBUG("len 0");
522 
523 	if (*more_neededp)
524 		isc->receive_len = icl_pdu_data_segment_receive_len(request);
525 
526 	return (0);
527 }
528 
529 static int
530 icl_pdu_check_data_digest(struct icl_pdu *request, struct mbuf **r, size_t *rs)
531 {
532 	uint32_t received_digest, valid_digest;
533 
534 	if (request->ip_conn->ic_data_crc32c == false)
535 		return (0);
536 
537 	if (request->ip_data_len == 0)
538 		return (0);
539 
540 	CTASSERT(sizeof(received_digest) == ISCSI_DATA_DIGEST_SIZE);
541 	icl_soft_receive_buf(r, rs, &received_digest, ISCSI_DATA_DIGEST_SIZE);
542 
543 	/*
544 	 * Note that ip_data_mbuf also contains padding; since digest
545 	 * calculation is supposed to include that, we iterate over
546 	 * the entire ip_data_mbuf chain, not just ip_data_len bytes of it.
547 	 */
548 	valid_digest = icl_mbuf_to_crc32c(request->ip_data_mbuf,
549 	    roundup2(request->ip_data_len, 4));
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_soft_conn *isc, struct mbuf **r, size_t *rs)
565 {
566 	struct icl_conn *ic = &isc->ic;
567 	struct icl_pdu *request;
568 	size_t len;
569 	int error = 0;
570 	bool more_needed;
571 
572 	if (isc->receive_state == ICL_CONN_STATE_BHS) {
573 		KASSERT(isc->receive_pdu == NULL,
574 		    ("isc->receive_pdu != NULL"));
575 		request = icl_soft_conn_new_pdu(ic, M_NOWAIT);
576 		if (request == NULL) {
577 			ICL_DEBUG("failed to allocate PDU; "
578 			    "dropping connection");
579 			icl_conn_fail(ic);
580 			return (NULL);
581 		}
582 		isc->receive_pdu = request;
583 	} else {
584 		KASSERT(isc->receive_pdu != NULL,
585 		    ("isc->receive_pdu == NULL"));
586 		request = isc->receive_pdu;
587 	}
588 
589 	switch (isc->receive_state) {
590 	case ICL_CONN_STATE_BHS:
591 		//ICL_DEBUG("receiving BHS");
592 		icl_soft_receive_buf(r, rs, request->ip_bhs,
593 		    sizeof(struct iscsi_bhs));
594 
595 		/*
596 		 * We don't enforce any limit for AHS length;
597 		 * its length is stored in 8 bit field.
598 		 */
599 
600 		len = icl_pdu_data_segment_length(request);
601 		if (len > ic->ic_max_recv_data_segment_length) {
602 			ICL_WARN("received data segment "
603 			    "length %zd is larger than negotiated; "
604 			    "dropping connection", len);
605 			error = EINVAL;
606 			break;
607 		}
608 
609 		isc->receive_state = ICL_CONN_STATE_AHS;
610 		isc->receive_len = icl_pdu_ahs_length(request);
611 		break;
612 
613 	case ICL_CONN_STATE_AHS:
614 		//ICL_DEBUG("receiving AHS");
615 		icl_pdu_receive_ahs(request, r, rs);
616 		isc->receive_state = ICL_CONN_STATE_HEADER_DIGEST;
617 		if (ic->ic_header_crc32c == false)
618 			isc->receive_len = 0;
619 		else
620 			isc->receive_len = ISCSI_HEADER_DIGEST_SIZE;
621 		break;
622 
623 	case ICL_CONN_STATE_HEADER_DIGEST:
624 		//ICL_DEBUG("receiving header digest");
625 		error = icl_pdu_check_header_digest(request, r, rs);
626 		if (error != 0) {
627 			ICL_DEBUG("header digest failed; "
628 			    "dropping connection");
629 			break;
630 		}
631 
632 		isc->receive_state = ICL_CONN_STATE_DATA;
633 		isc->receive_len = icl_pdu_data_segment_receive_len(request);
634 		break;
635 
636 	case ICL_CONN_STATE_DATA:
637 		//ICL_DEBUG("receiving data segment");
638 		error = icl_pdu_receive_data_segment(request, r, rs,
639 		    &more_needed);
640 		if (error != 0) {
641 			ICL_DEBUG("failed to receive data segment;"
642 			    "dropping connection");
643 			break;
644 		}
645 
646 		if (more_needed)
647 			break;
648 
649 		isc->receive_state = ICL_CONN_STATE_DATA_DIGEST;
650 		if (request->ip_data_len == 0 || ic->ic_data_crc32c == false)
651 			isc->receive_len = 0;
652 		else
653 			isc->receive_len = ISCSI_DATA_DIGEST_SIZE;
654 		break;
655 
656 	case ICL_CONN_STATE_DATA_DIGEST:
657 		//ICL_DEBUG("receiving data digest");
658 		error = icl_pdu_check_data_digest(request, r, rs);
659 		if (error != 0) {
660 			ICL_DEBUG("data digest failed; "
661 			    "dropping connection");
662 			break;
663 		}
664 
665 		/*
666 		 * We've received complete PDU; reset the receive state machine
667 		 * and return the PDU.
668 		 */
669 		isc->receive_state = ICL_CONN_STATE_BHS;
670 		isc->receive_len = sizeof(struct iscsi_bhs);
671 		isc->receive_pdu = NULL;
672 		return (request);
673 
674 	default:
675 		panic("invalid receive_state %d\n", isc->receive_state);
676 	}
677 
678 	if (error != 0) {
679 		/*
680 		 * Don't free the PDU; it's pointed to by isc->receive_pdu
681 		 * and will get freed in icl_soft_conn_close().
682 		 */
683 		icl_conn_fail(ic);
684 	}
685 
686 	return (NULL);
687 }
688 
689 static void
690 icl_conn_receive_pdus(struct icl_soft_conn *isc, struct mbuf **r, size_t *rs)
691 {
692 	struct icl_conn *ic = &isc->ic;
693 	struct icl_pdu *response;
694 
695 	for (;;) {
696 		if (ic->ic_disconnecting)
697 			return;
698 
699 		/*
700 		 * Loop until we have a complete PDU or there is not enough
701 		 * data in the socket buffer.
702 		 */
703 		if (*rs < isc->receive_len) {
704 #if 0
705 			ICL_DEBUG("not enough data; have %zd, need %zd",
706 			    *rs, isc->receive_len);
707 #endif
708 			return;
709 		}
710 
711 		response = icl_conn_receive_pdu(isc, r, rs);
712 		if (response == NULL)
713 			continue;
714 
715 		if (response->ip_ahs_len > 0) {
716 			ICL_WARN("received PDU with unsupported "
717 			    "AHS; opcode 0x%x; dropping connection",
718 			    response->ip_bhs->bhs_opcode);
719 			icl_soft_conn_pdu_free(ic, response);
720 			icl_conn_fail(ic);
721 			return;
722 		}
723 
724 		(ic->ic_receive)(response);
725 	}
726 }
727 
728 static void
729 icl_receive_thread(void *arg)
730 {
731 	struct icl_soft_conn *isc = arg;
732 	struct icl_conn *ic = &isc->ic;
733 	size_t available, read = 0;
734 	struct socket *so;
735 	struct mbuf *m, *r = NULL;
736 	struct uio uio;
737 	int error, flags;
738 
739 	so = ic->ic_socket;
740 
741 	for (;;) {
742 		SOCKBUF_LOCK(&so->so_rcv);
743 		if (ic->ic_disconnecting) {
744 			SOCKBUF_UNLOCK(&so->so_rcv);
745 			break;
746 		}
747 
748 		/*
749 		 * Set the low watermark, to be checked by
750 		 * soreadable() in icl_soupcall_receive()
751 		 * to avoid unnecessary wakeups until there
752 		 * is enough data received to read the PDU.
753 		 */
754 		available = sbavail(&so->so_rcv);
755 		if (read + available < isc->receive_len) {
756 			so->so_rcv.sb_lowat = isc->receive_len - read;
757 			cv_wait(&isc->receive_cv, SOCKBUF_MTX(&so->so_rcv));
758 			so->so_rcv.sb_lowat = so->so_rcv.sb_hiwat + 1;
759 			available = sbavail(&so->so_rcv);
760 		}
761 		SOCKBUF_UNLOCK(&so->so_rcv);
762 
763 		if (available == 0) {
764 			if (so->so_error != 0) {
765 				ICL_DEBUG("connection error %d; "
766 				    "dropping connection", so->so_error);
767 				icl_conn_fail(ic);
768 				break;
769 			}
770 			continue;
771 		}
772 
773 		memset(&uio, 0, sizeof(uio));
774 		uio.uio_resid = available;
775 		flags = MSG_DONTWAIT;
776 		error = soreceive(so, NULL, &uio, &m, NULL, &flags);
777 		if (error != 0) {
778 			ICL_DEBUG("soreceive error %d", error);
779 			break;
780 		}
781 		if (uio.uio_resid != 0) {
782 			m_freem(m);
783 			ICL_DEBUG("short read");
784 			break;
785 		}
786 		if (r)
787 			m_cat(r, m);
788 		else
789 			r = m;
790 		read += available;
791 
792 		icl_conn_receive_pdus(isc, &r, &read);
793 	}
794 
795 	if (r)
796 		m_freem(r);
797 
798 	ICL_CONN_LOCK(ic);
799 	isc->receive_running = false;
800 	cv_signal(&isc->send_cv);
801 	ICL_CONN_UNLOCK(ic);
802 	kthread_exit();
803 }
804 
805 static int
806 icl_soupcall_receive(struct socket *so, void *arg, int waitflag)
807 {
808 	struct icl_soft_conn *isc;
809 
810 	if (!soreadable(so))
811 		return (SU_OK);
812 
813 	isc = arg;
814 	cv_signal(&isc->receive_cv);
815 	return (SU_OK);
816 }
817 
818 static int
819 icl_pdu_finalize(struct icl_pdu *request)
820 {
821 	size_t padding, pdu_len;
822 	uint32_t digest, zero = 0;
823 	int ok;
824 	struct icl_conn *ic;
825 
826 	ic = request->ip_conn;
827 
828 	icl_pdu_set_data_segment_length(request, request->ip_data_len);
829 
830 	pdu_len = icl_pdu_size(request);
831 
832 	if (ic->ic_header_crc32c) {
833 		digest = icl_mbuf_to_crc32c(request->ip_bhs_mbuf,
834 		    ISCSI_BHS_SIZE);
835 		ok = m_append(request->ip_bhs_mbuf, sizeof(digest),
836 		    (void *)&digest);
837 		if (ok != 1) {
838 			ICL_WARN("failed to append header digest");
839 			return (1);
840 		}
841 	}
842 
843 	if (request->ip_data_len != 0) {
844 		padding = icl_pdu_padding(request);
845 		if (padding > 0) {
846 			ok = m_append(request->ip_data_mbuf, padding,
847 			    (void *)&zero);
848 			if (ok != 1) {
849 				ICL_WARN("failed to append padding");
850 				return (1);
851 			}
852 		}
853 
854 		if (ic->ic_data_crc32c) {
855 			digest = icl_mbuf_to_crc32c(request->ip_data_mbuf,
856 			    roundup2(request->ip_data_len, 4));
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_soft_conn *isc, struct icl_pdu_stailq *queue)
877 {
878 	struct icl_conn *ic = &isc->ic;
879 	struct icl_pdu *request, *request2;
880 	struct mbuf *m;
881 	struct socket *so;
882 	long available, size, size2;
883 #ifdef DEBUG_COALESCED
884 	int coalesced;
885 #endif
886 	int error;
887 
888 	ICL_CONN_LOCK_ASSERT_NOT(ic);
889 
890 	so = ic->ic_socket;
891 
892 	SOCKBUF_LOCK(&so->so_snd);
893 	/*
894 	 * Check how much space do we have for transmit.  We can't just
895 	 * call sosend() and retry when we get EWOULDBLOCK or EMSGSIZE,
896 	 * as it always frees the mbuf chain passed to it, even in case
897 	 * of error.
898 	 */
899 	available = sbspace(&so->so_snd);
900 	isc->check_send_space = false;
901 
902 	/*
903 	 * Notify the socket upcall that we don't need wakeups
904 	 * for the time being.
905 	 */
906 	so->so_snd.sb_lowat = so->so_snd.sb_hiwat + 1;
907 	SOCKBUF_UNLOCK(&so->so_snd);
908 
909 	while (!STAILQ_EMPTY(queue)) {
910 		request = STAILQ_FIRST(queue);
911 		size = icl_pdu_size(request);
912 		if (available < size) {
913 			/*
914 			 * Set the low watermark, to be checked by
915 			 * sowriteable() in icl_soupcall_send()
916 			 * to avoid unnecessary wakeups until there
917 			 * is enough space for the PDU to fit.
918 			 */
919 			SOCKBUF_LOCK(&so->so_snd);
920 			available = sbspace(&so->so_snd);
921 			if (available < size) {
922 #if 1
923 				ICL_DEBUG("no space to send; "
924 				    "have %ld, need %ld",
925 				    available, size);
926 #endif
927 				so->so_snd.sb_lowat = max(size,
928 				    so->so_snd.sb_hiwat / 8);
929 				SOCKBUF_UNLOCK(&so->so_snd);
930 				return;
931 			}
932 			SOCKBUF_UNLOCK(&so->so_snd);
933 		}
934 		STAILQ_REMOVE_HEAD(queue, ip_next);
935 		error = icl_pdu_finalize(request);
936 		if (error != 0) {
937 			ICL_DEBUG("failed to finalize PDU; "
938 			    "dropping connection");
939 			icl_soft_pdu_done(request, EIO);
940 			icl_conn_fail(ic);
941 			return;
942 		}
943 		if (coalesce) {
944 			m = request->ip_bhs_mbuf;
945 			for (
946 #ifdef DEBUG_COALESCED
947 			    coalesced = 1
948 #endif
949 			    ; ;
950 #ifdef DEBUG_COALESCED
951 			    coalesced++
952 #endif
953 			    ) {
954 				request2 = STAILQ_FIRST(queue);
955 				if (request2 == NULL)
956 					break;
957 				size2 = icl_pdu_size(request2);
958 				if (available < size + size2)
959 					break;
960 				STAILQ_REMOVE_HEAD(queue, ip_next);
961 				error = icl_pdu_finalize(request2);
962 				if (error != 0) {
963 					ICL_DEBUG("failed to finalize PDU; "
964 					    "dropping connection");
965 					icl_soft_pdu_done(request, EIO);
966 					icl_soft_pdu_done(request2, EIO);
967 					icl_conn_fail(ic);
968 					return;
969 				}
970 				while (m->m_next)
971 					m = m->m_next;
972 				m_cat(m, request2->ip_bhs_mbuf);
973 				request2->ip_bhs_mbuf = NULL;
974 				request->ip_bhs_mbuf->m_pkthdr.len += size2;
975 				size += size2;
976 				icl_soft_pdu_done(request2, 0);
977 			}
978 #ifdef DEBUG_COALESCED
979 			if (coalesced > 1) {
980 				ICL_DEBUG("coalesced %d PDUs into %ld bytes",
981 				    coalesced, size);
982 			}
983 #endif
984 		}
985 		available -= size;
986 		error = sosend(so, NULL, NULL, request->ip_bhs_mbuf,
987 		    NULL, MSG_DONTWAIT, curthread);
988 		request->ip_bhs_mbuf = NULL; /* Sosend consumes the mbuf. */
989 		if (error != 0) {
990 			ICL_DEBUG("failed to send PDU, error %d; "
991 			    "dropping connection", error);
992 			icl_soft_pdu_done(request, error);
993 			icl_conn_fail(ic);
994 			return;
995 		}
996 		icl_soft_pdu_done(request, 0);
997 	}
998 }
999 
1000 static void
1001 icl_send_thread(void *arg)
1002 {
1003 	struct icl_soft_conn *isc;
1004 	struct icl_conn *ic;
1005 	struct icl_pdu_stailq queue;
1006 
1007 	isc = arg;
1008 	ic = &isc->ic;
1009 
1010 	STAILQ_INIT(&queue);
1011 
1012 	ICL_CONN_LOCK(ic);
1013 	for (;;) {
1014 		for (;;) {
1015 			/*
1016 			 * Populate the local queue from the main one.
1017 			 * This way the icl_conn_send_pdus() can go through
1018 			 * all the queued PDUs without holding any locks.
1019 			 */
1020 			if (STAILQ_EMPTY(&queue) || isc->check_send_space)
1021 				STAILQ_CONCAT(&queue, &isc->to_send);
1022 
1023 			ICL_CONN_UNLOCK(ic);
1024 			icl_conn_send_pdus(isc, &queue);
1025 			ICL_CONN_LOCK(ic);
1026 
1027 			/*
1028 			 * The icl_soupcall_send() was called since the last
1029 			 * call to sbspace(); go around;
1030 			 */
1031 			if (isc->check_send_space)
1032 				continue;
1033 
1034 			/*
1035 			 * Local queue is empty, but we still have PDUs
1036 			 * in the main one; go around.
1037 			 */
1038 			if (STAILQ_EMPTY(&queue) &&
1039 			    !STAILQ_EMPTY(&isc->to_send))
1040 				continue;
1041 
1042 			/*
1043 			 * There might be some stuff in the local queue,
1044 			 * which didn't get sent due to not having enough send
1045 			 * space.  Wait for socket upcall.
1046 			 */
1047 			break;
1048 		}
1049 
1050 		if (ic->ic_disconnecting) {
1051 			//ICL_DEBUG("terminating");
1052 			break;
1053 		}
1054 
1055 		cv_wait(&isc->send_cv, ic->ic_lock);
1056 	}
1057 
1058 	/*
1059 	 * We're exiting; move PDUs back to the main queue, so they can
1060 	 * get freed properly.  At this point ordering doesn't matter.
1061 	 */
1062 	STAILQ_CONCAT(&isc->to_send, &queue);
1063 
1064 	isc->send_running = false;
1065 	cv_signal(&isc->send_cv);
1066 	ICL_CONN_UNLOCK(ic);
1067 	kthread_exit();
1068 }
1069 
1070 static int
1071 icl_soupcall_send(struct socket *so, void *arg, int waitflag)
1072 {
1073 	struct icl_soft_conn *isc;
1074 	struct icl_conn *ic;
1075 
1076 	if (!sowriteable(so))
1077 		return (SU_OK);
1078 
1079 	isc = arg;
1080 	ic = &isc->ic;
1081 
1082 	ICL_CONN_LOCK(ic);
1083 	isc->check_send_space = true;
1084 	ICL_CONN_UNLOCK(ic);
1085 
1086 	cv_signal(&isc->send_cv);
1087 
1088 	return (SU_OK);
1089 }
1090 
1091 static void
1092 icl_soft_free_mext_pg(struct mbuf *m)
1093 {
1094 	struct icl_soft_pdu *isp;
1095 
1096 	M_ASSERTEXTPG(m);
1097 
1098 	/*
1099 	 * Nothing to do for the pages; they are owned by the PDU /
1100 	 * I/O request.
1101 	 */
1102 
1103 	/* Drop reference on the PDU. */
1104 	isp = m->m_ext.ext_arg1;
1105 	if (atomic_fetchadd_int(&isp->ref_cnt, -1) == 1)
1106 		icl_soft_pdu_call_cb(&isp->ip);
1107 }
1108 
1109 static int
1110 icl_soft_conn_pdu_append_bio(struct icl_conn *ic, struct icl_pdu *request,
1111     struct bio *bp, size_t offset, size_t len, int flags)
1112 {
1113 	struct icl_soft_pdu *isp = (struct icl_soft_pdu *)request;
1114 	struct mbuf *m, *m_tail;
1115 	vm_offset_t vaddr;
1116 	size_t mtodo, page_offset, todo;
1117 	int i;
1118 
1119 	KASSERT(len > 0, ("len == 0"));
1120 
1121 	m_tail = request->ip_data_mbuf;
1122 	if (m_tail != NULL)
1123 		for (; m_tail->m_next != NULL; m_tail = m_tail->m_next)
1124 			;
1125 
1126 	MPASS(bp->bio_flags & BIO_UNMAPPED);
1127 	if (offset < PAGE_SIZE - bp->bio_ma_offset) {
1128 		page_offset = bp->bio_ma_offset + offset;
1129 		i = 0;
1130 	} else {
1131 		offset -= PAGE_SIZE - bp->bio_ma_offset;
1132 		for (i = 1; offset >= PAGE_SIZE; i++)
1133 			offset -= PAGE_SIZE;
1134 		page_offset = offset;
1135 	}
1136 
1137 	if (flags & ICL_NOCOPY) {
1138 		m = NULL;
1139 		while (len > 0) {
1140 			if (m == NULL) {
1141 				m = mb_alloc_ext_pgs(flags & ~ICL_NOCOPY,
1142 				    icl_soft_free_mext_pg);
1143 				if (__predict_false(m == NULL))
1144 					return (ENOMEM);
1145 				atomic_add_int(&isp->ref_cnt, 1);
1146 				m->m_ext.ext_arg1 = isp;
1147 				m->m_epg_1st_off = page_offset;
1148 			}
1149 
1150 			todo = MIN(len, PAGE_SIZE - page_offset);
1151 
1152 			m->m_epg_pa[m->m_epg_npgs] =
1153 			    VM_PAGE_TO_PHYS(bp->bio_ma[i]);
1154 			m->m_epg_npgs++;
1155 			m->m_epg_last_len = todo;
1156 			m->m_len += todo;
1157 			m->m_ext.ext_size += PAGE_SIZE;
1158 			MBUF_EXT_PGS_ASSERT_SANITY(m);
1159 
1160 			if (m->m_epg_npgs == MBUF_PEXT_MAX_PGS) {
1161 				if (m_tail != NULL)
1162 					m_tail->m_next = m;
1163 				else
1164 					request->ip_data_mbuf = m;
1165 				m_tail = m;
1166 				request->ip_data_len += m->m_len;
1167 				m = NULL;
1168 			}
1169 
1170 			page_offset = 0;
1171 			len -= todo;
1172 			i++;
1173 		}
1174 
1175 		if (m != NULL) {
1176 			if (m_tail != NULL)
1177 				m_tail->m_next = m;
1178 			else
1179 				request->ip_data_mbuf = m;
1180 			request->ip_data_len += m->m_len;
1181 		}
1182 		return (0);
1183 	}
1184 
1185 	m = m_getm2(NULL, len, flags, MT_DATA, 0);
1186 	if (__predict_false(m == NULL))
1187 		return (ENOMEM);
1188 
1189 	if (request->ip_data_mbuf == NULL) {
1190 		request->ip_data_mbuf = m;
1191 		request->ip_data_len = len;
1192 	} else {
1193 		m_tail->m_next = m;
1194 		request->ip_data_len += len;
1195 	}
1196 
1197 	while (len > 0) {
1198 		todo = MIN(len, PAGE_SIZE - page_offset);
1199 		vaddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(bp->bio_ma[i]));
1200 
1201 		do {
1202 			mtodo = min(todo, M_SIZE(m) - m->m_len);
1203 			memcpy(mtod(m, char *) + m->m_len, (char *)vaddr +
1204 			    page_offset, mtodo);
1205 			m->m_len += mtodo;
1206 			if (m->m_len == M_SIZE(m))
1207 				m = m->m_next;
1208 			page_offset += mtodo;
1209 			todo -= mtodo;
1210 		} while (todo > 0);
1211 
1212 		page_offset = 0;
1213 		len -= todo;
1214 		i++;
1215 	}
1216 
1217 	return (0);
1218 }
1219 
1220 static int
1221 icl_soft_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *request,
1222     const void *addr, size_t len, int flags)
1223 {
1224 	struct icl_soft_pdu *isp = (struct icl_soft_pdu *)request;
1225 	struct mbuf *mb, *newmb;
1226 	size_t copylen, off = 0;
1227 
1228 	KASSERT(len > 0, ("len == 0"));
1229 
1230 	if (flags & ICL_NOCOPY) {
1231 		newmb = m_get(flags & ~ICL_NOCOPY, MT_DATA);
1232 		if (newmb == NULL) {
1233 			ICL_WARN("failed to allocate mbuf");
1234 			return (ENOMEM);
1235 		}
1236 
1237 		newmb->m_flags |= M_RDONLY;
1238 		m_extaddref(newmb, __DECONST(char *, addr), len, &isp->ref_cnt,
1239 		    icl_soft_mbuf_done, isp, NULL);
1240 		newmb->m_len = len;
1241 	} else {
1242 		newmb = m_getm2(NULL, len, flags, MT_DATA, 0);
1243 		if (newmb == NULL) {
1244 			ICL_WARN("failed to allocate mbuf for %zd bytes", len);
1245 			return (ENOMEM);
1246 		}
1247 
1248 		for (mb = newmb; mb != NULL; mb = mb->m_next) {
1249 			copylen = min(M_TRAILINGSPACE(mb), len - off);
1250 			memcpy(mtod(mb, char *), (const char *)addr + off, copylen);
1251 			mb->m_len = copylen;
1252 			off += copylen;
1253 		}
1254 		KASSERT(off == len, ("%s: off != len", __func__));
1255 	}
1256 
1257 	if (request->ip_data_mbuf == NULL) {
1258 		request->ip_data_mbuf = newmb;
1259 		request->ip_data_len = len;
1260 	} else {
1261 		m_cat(request->ip_data_mbuf, newmb);
1262 		request->ip_data_len += len;
1263 	}
1264 
1265 	return (0);
1266 }
1267 
1268 void
1269 icl_soft_conn_pdu_get_bio(struct icl_conn *ic, struct icl_pdu *ip,
1270     size_t pdu_off, struct bio *bp, size_t bio_off, size_t len)
1271 {
1272 	vm_offset_t vaddr;
1273 	size_t page_offset, todo;
1274 	int i __unused;
1275 
1276 	MPASS(bp->bio_flags & BIO_UNMAPPED);
1277 	if (bio_off < PAGE_SIZE - bp->bio_ma_offset) {
1278 		page_offset = bp->bio_ma_offset + bio_off;
1279 		i = 0;
1280 	} else {
1281 		bio_off -= PAGE_SIZE - bp->bio_ma_offset;
1282 		for (i = 1; bio_off >= PAGE_SIZE; i++)
1283 			bio_off -= PAGE_SIZE;
1284 		page_offset = bio_off;
1285 	}
1286 
1287 	while (len > 0) {
1288 		todo = MIN(len, PAGE_SIZE - page_offset);
1289 
1290 		vaddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(bp->bio_ma[i]));
1291 		m_copydata(ip->ip_data_mbuf, pdu_off, todo, (char *)vaddr +
1292 		    page_offset);
1293 
1294 		page_offset = 0;
1295 		pdu_off += todo;
1296 		len -= todo;
1297 		i++;
1298 	}
1299 }
1300 
1301 void
1302 icl_soft_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip,
1303     size_t off, void *addr, size_t len)
1304 {
1305 
1306 	m_copydata(ip->ip_data_mbuf, off, len, addr);
1307 }
1308 
1309 static void
1310 icl_soft_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip)
1311 {
1312 
1313 	icl_soft_conn_pdu_queue_cb(ic, ip, NULL);
1314 }
1315 
1316 static void
1317 icl_soft_conn_pdu_queue_cb(struct icl_conn *ic, struct icl_pdu *ip,
1318     icl_pdu_cb cb)
1319 {
1320 	struct icl_soft_conn *isc = (struct icl_soft_conn *)ic;
1321 	struct icl_soft_pdu *isp = (struct icl_soft_pdu *)ip;
1322 
1323 	ICL_CONN_LOCK_ASSERT(ic);
1324 	isp->ref_cnt++;
1325 	isp->cb = cb;
1326 
1327 	if (ic->ic_disconnecting || ic->ic_socket == NULL) {
1328 		ICL_DEBUG("icl_pdu_queue on closed connection");
1329 		icl_soft_pdu_done(ip, ENOTCONN);
1330 		return;
1331 	}
1332 
1333 	if (!STAILQ_EMPTY(&isc->to_send)) {
1334 		STAILQ_INSERT_TAIL(&isc->to_send, ip, ip_next);
1335 		/*
1336 		 * If the queue is not empty, someone else had already
1337 		 * signaled the send thread; no need to do that again,
1338 		 * just return.
1339 		 */
1340 		return;
1341 	}
1342 
1343 	STAILQ_INSERT_TAIL(&isc->to_send, ip, ip_next);
1344 	cv_signal(&isc->send_cv);
1345 }
1346 
1347 static struct icl_conn *
1348 icl_soft_new_conn(const char *name, struct mtx *lock)
1349 {
1350 	struct icl_soft_conn *isc;
1351 	struct icl_conn *ic;
1352 
1353 	refcount_acquire(&icl_ncons);
1354 
1355 	isc = (struct icl_soft_conn *)kobj_create(&icl_soft_class, M_ICL_SOFT,
1356 	    M_WAITOK | M_ZERO);
1357 
1358 	STAILQ_INIT(&isc->to_send);
1359 	cv_init(&isc->send_cv, "icl_tx");
1360 	cv_init(&isc->receive_cv, "icl_rx");
1361 
1362 	ic = &isc->ic;
1363 	ic->ic_lock = lock;
1364 #ifdef DIAGNOSTIC
1365 	refcount_init(&ic->ic_outstanding_pdus, 0);
1366 #endif
1367 	ic->ic_name = name;
1368 	ic->ic_offload = "None";
1369 	ic->ic_unmapped = PMAP_HAS_DMAP;
1370 
1371 	return (ic);
1372 }
1373 
1374 void
1375 icl_soft_conn_free(struct icl_conn *ic)
1376 {
1377 	struct icl_soft_conn *isc = (struct icl_soft_conn *)ic;
1378 
1379 #ifdef DIAGNOSTIC
1380 	KASSERT(ic->ic_outstanding_pdus == 0,
1381 	    ("destroying session with %d outstanding PDUs",
1382 	     ic->ic_outstanding_pdus));
1383 #endif
1384 	cv_destroy(&isc->send_cv);
1385 	cv_destroy(&isc->receive_cv);
1386 	kobj_delete((struct kobj *)isc, M_ICL_SOFT);
1387 	refcount_release(&icl_ncons);
1388 }
1389 
1390 static int
1391 icl_conn_start(struct icl_conn *ic)
1392 {
1393 	struct icl_soft_conn *isc = (struct icl_soft_conn *)ic;
1394 	size_t minspace;
1395 	struct sockopt opt;
1396 	int error, one = 1;
1397 
1398 	ICL_CONN_LOCK(ic);
1399 
1400 	/*
1401 	 * XXX: Ugly hack.
1402 	 */
1403 	if (ic->ic_socket == NULL) {
1404 		ICL_CONN_UNLOCK(ic);
1405 		return (EINVAL);
1406 	}
1407 
1408 	isc->receive_state = ICL_CONN_STATE_BHS;
1409 	isc->receive_len = sizeof(struct iscsi_bhs);
1410 	ic->ic_disconnecting = false;
1411 
1412 	ICL_CONN_UNLOCK(ic);
1413 
1414 	/*
1415 	 * For sendspace, this is required because the current code cannot
1416 	 * send a PDU in pieces; thus, the minimum buffer size is equal
1417 	 * to the maximum PDU size.  "+4" is to account for possible padding.
1418 	 */
1419 	minspace = sizeof(struct iscsi_bhs) +
1420 	    ic->ic_max_send_data_segment_length +
1421 	    ISCSI_HEADER_DIGEST_SIZE + ISCSI_DATA_DIGEST_SIZE + 4;
1422 	if (sendspace < minspace) {
1423 		ICL_WARN("kern.icl.sendspace too low; must be at least %zd",
1424 		    minspace);
1425 		sendspace = minspace;
1426 	}
1427 	minspace = sizeof(struct iscsi_bhs) +
1428 	    ic->ic_max_recv_data_segment_length +
1429 	    ISCSI_HEADER_DIGEST_SIZE + ISCSI_DATA_DIGEST_SIZE + 4;
1430 	if (recvspace < minspace) {
1431 		ICL_WARN("kern.icl.recvspace too low; must be at least %zd",
1432 		    minspace);
1433 		recvspace = minspace;
1434 	}
1435 
1436 	error = soreserve(ic->ic_socket, sendspace, recvspace);
1437 	if (error != 0) {
1438 		ICL_WARN("soreserve failed with error %d", error);
1439 		icl_soft_conn_close(ic);
1440 		return (error);
1441 	}
1442 	ic->ic_socket->so_snd.sb_flags |= SB_AUTOSIZE;
1443 	ic->ic_socket->so_rcv.sb_flags |= SB_AUTOSIZE;
1444 
1445 	/*
1446 	 * Disable Nagle.
1447 	 */
1448 	bzero(&opt, sizeof(opt));
1449 	opt.sopt_dir = SOPT_SET;
1450 	opt.sopt_level = IPPROTO_TCP;
1451 	opt.sopt_name = TCP_NODELAY;
1452 	opt.sopt_val = &one;
1453 	opt.sopt_valsize = sizeof(one);
1454 	error = sosetopt(ic->ic_socket, &opt);
1455 	if (error != 0) {
1456 		ICL_WARN("disabling TCP_NODELAY failed with error %d", error);
1457 		icl_soft_conn_close(ic);
1458 		return (error);
1459 	}
1460 
1461 	/*
1462 	 * Register socket upcall, to get notified about incoming PDUs
1463 	 * and free space to send outgoing ones.
1464 	 */
1465 	SOCKBUF_LOCK(&ic->ic_socket->so_snd);
1466 	soupcall_set(ic->ic_socket, SO_SND, icl_soupcall_send, isc);
1467 	SOCKBUF_UNLOCK(&ic->ic_socket->so_snd);
1468 	SOCKBUF_LOCK(&ic->ic_socket->so_rcv);
1469 	soupcall_set(ic->ic_socket, SO_RCV, icl_soupcall_receive, isc);
1470 	SOCKBUF_UNLOCK(&ic->ic_socket->so_rcv);
1471 
1472 	/*
1473 	 * Start threads.
1474 	 */
1475 	ICL_CONN_LOCK(ic);
1476 	isc->send_running = isc->receive_running = true;
1477 	ICL_CONN_UNLOCK(ic);
1478 	error = kthread_add(icl_send_thread, ic, NULL, NULL, 0, 0, "%stx",
1479 	    ic->ic_name);
1480 	if (error != 0) {
1481 		ICL_WARN("kthread_add(9) failed with error %d", error);
1482 		ICL_CONN_LOCK(ic);
1483 		isc->send_running = isc->receive_running = false;
1484 		cv_signal(&isc->send_cv);
1485 		ICL_CONN_UNLOCK(ic);
1486 		icl_soft_conn_close(ic);
1487 		return (error);
1488 	}
1489 	error = kthread_add(icl_receive_thread, ic, NULL, NULL, 0, 0, "%srx",
1490 	    ic->ic_name);
1491 	if (error != 0) {
1492 		ICL_WARN("kthread_add(9) failed with error %d", error);
1493 		ICL_CONN_LOCK(ic);
1494 		isc->receive_running = false;
1495 		cv_signal(&isc->send_cv);
1496 		ICL_CONN_UNLOCK(ic);
1497 		icl_soft_conn_close(ic);
1498 		return (error);
1499 	}
1500 
1501 	return (0);
1502 }
1503 
1504 int
1505 icl_soft_conn_handoff(struct icl_conn *ic, int fd)
1506 {
1507 	struct file *fp;
1508 	struct socket *so;
1509 	cap_rights_t rights;
1510 	int error;
1511 
1512 	ICL_CONN_LOCK_ASSERT_NOT(ic);
1513 
1514 #ifdef ICL_KERNEL_PROXY
1515 	/*
1516 	 * We're transitioning to Full Feature phase, and we don't
1517 	 * really care.
1518 	 */
1519 	if (fd == 0) {
1520 		ICL_CONN_LOCK(ic);
1521 		if (ic->ic_socket == NULL) {
1522 			ICL_CONN_UNLOCK(ic);
1523 			ICL_WARN("proxy handoff without connect");
1524 			return (EINVAL);
1525 		}
1526 		ICL_CONN_UNLOCK(ic);
1527 		return (0);
1528 	}
1529 #endif
1530 
1531 	/*
1532 	 * Steal the socket from userland.
1533 	 */
1534 	error = fget(curthread, fd,
1535 	    cap_rights_init_one(&rights, CAP_SOCK_CLIENT), &fp);
1536 	if (error != 0)
1537 		return (error);
1538 	if (fp->f_type != DTYPE_SOCKET) {
1539 		fdrop(fp, curthread);
1540 		return (EINVAL);
1541 	}
1542 	so = fp->f_data;
1543 	if (so->so_type != SOCK_STREAM) {
1544 		fdrop(fp, curthread);
1545 		return (EINVAL);
1546 	}
1547 
1548 	ICL_CONN_LOCK(ic);
1549 
1550 	if (ic->ic_socket != NULL) {
1551 		ICL_CONN_UNLOCK(ic);
1552 		fdrop(fp, curthread);
1553 		return (EBUSY);
1554 	}
1555 
1556 	ic->ic_socket = fp->f_data;
1557 	fp->f_ops = &badfileops;
1558 	fp->f_data = NULL;
1559 	fdrop(fp, curthread);
1560 	ICL_CONN_UNLOCK(ic);
1561 
1562 	error = icl_conn_start(ic);
1563 
1564 	return (error);
1565 }
1566 
1567 void
1568 icl_soft_conn_close(struct icl_conn *ic)
1569 {
1570 	struct icl_soft_conn *isc = (struct icl_soft_conn *)ic;
1571 	struct icl_pdu *pdu;
1572 	struct socket *so;
1573 
1574 	/*
1575 	 * Wake up the threads, so they can properly terminate.
1576 	 * Receive thread sleeps on so->so_rcv lock, send on ic->ic_lock.
1577 	 */
1578 	ICL_CONN_LOCK(ic);
1579 	if (!ic->ic_disconnecting) {
1580 		so = ic->ic_socket;
1581 		if (so)
1582 			SOCKBUF_LOCK(&so->so_rcv);
1583 		ic->ic_disconnecting = true;
1584 		if (so)
1585 			SOCKBUF_UNLOCK(&so->so_rcv);
1586 	}
1587 	while (isc->receive_running || isc->send_running) {
1588 		cv_signal(&isc->receive_cv);
1589 		cv_signal(&isc->send_cv);
1590 		cv_wait(&isc->send_cv, ic->ic_lock);
1591 	}
1592 
1593 	/* Some other thread could close the connection same time. */
1594 	so = ic->ic_socket;
1595 	if (so == NULL) {
1596 		ICL_CONN_UNLOCK(ic);
1597 		return;
1598 	}
1599 	ic->ic_socket = NULL;
1600 
1601 	/*
1602 	 * Deregister socket upcalls.
1603 	 */
1604 	ICL_CONN_UNLOCK(ic);
1605 	SOCKBUF_LOCK(&so->so_snd);
1606 	if (so->so_snd.sb_upcall != NULL)
1607 		soupcall_clear(so, SO_SND);
1608 	SOCKBUF_UNLOCK(&so->so_snd);
1609 	SOCKBUF_LOCK(&so->so_rcv);
1610 	if (so->so_rcv.sb_upcall != NULL)
1611 		soupcall_clear(so, SO_RCV);
1612 	SOCKBUF_UNLOCK(&so->so_rcv);
1613 	soclose(so);
1614 	ICL_CONN_LOCK(ic);
1615 
1616 	if (isc->receive_pdu != NULL) {
1617 		//ICL_DEBUG("freeing partially received PDU");
1618 		icl_soft_conn_pdu_free(ic, isc->receive_pdu);
1619 		isc->receive_pdu = NULL;
1620 	}
1621 
1622 	/*
1623 	 * Remove any outstanding PDUs from the send queue.
1624 	 */
1625 	while (!STAILQ_EMPTY(&isc->to_send)) {
1626 		pdu = STAILQ_FIRST(&isc->to_send);
1627 		STAILQ_REMOVE_HEAD(&isc->to_send, ip_next);
1628 		icl_soft_pdu_done(pdu, ENOTCONN);
1629 	}
1630 
1631 	KASSERT(STAILQ_EMPTY(&isc->to_send),
1632 	    ("destroying session with non-empty send queue"));
1633 	ICL_CONN_UNLOCK(ic);
1634 }
1635 
1636 int
1637 icl_soft_conn_task_setup(struct icl_conn *ic, struct icl_pdu *ip,
1638     struct ccb_scsiio *csio, uint32_t *task_tagp, void **prvp)
1639 {
1640 
1641 	return (0);
1642 }
1643 
1644 void
1645 icl_soft_conn_task_done(struct icl_conn *ic, void *prv)
1646 {
1647 }
1648 
1649 int
1650 icl_soft_conn_transfer_setup(struct icl_conn *ic, struct icl_pdu *ip,
1651     union ctl_io *io, uint32_t *transfer_tag, void **prvp)
1652 {
1653 
1654 	return (0);
1655 }
1656 
1657 void
1658 icl_soft_conn_transfer_done(struct icl_conn *ic, void *prv)
1659 {
1660 }
1661 
1662 static int
1663 icl_soft_limits(struct icl_drv_limits *idl, int socket)
1664 {
1665 
1666 	idl->idl_max_recv_data_segment_length = max_data_segment_length;
1667 	idl->idl_max_send_data_segment_length = max_data_segment_length;
1668 	idl->idl_max_burst_length = max_burst_length;
1669 	idl->idl_first_burst_length = first_burst_length;
1670 
1671 	return (0);
1672 }
1673 
1674 #ifdef ICL_KERNEL_PROXY
1675 int
1676 icl_soft_conn_connect(struct icl_conn *ic, int domain, int socktype,
1677     int protocol, struct sockaddr *from_sa, struct sockaddr *to_sa)
1678 {
1679 
1680 	return (icl_soft_proxy_connect(ic, domain, socktype, protocol,
1681 	    from_sa, to_sa));
1682 }
1683 
1684 int
1685 icl_soft_handoff_sock(struct icl_conn *ic, struct socket *so)
1686 {
1687 	int error;
1688 
1689 	ICL_CONN_LOCK_ASSERT_NOT(ic);
1690 
1691 	if (so->so_type != SOCK_STREAM)
1692 		return (EINVAL);
1693 
1694 	ICL_CONN_LOCK(ic);
1695 	if (ic->ic_socket != NULL) {
1696 		ICL_CONN_UNLOCK(ic);
1697 		return (EBUSY);
1698 	}
1699 	ic->ic_socket = so;
1700 	ICL_CONN_UNLOCK(ic);
1701 
1702 	error = icl_conn_start(ic);
1703 
1704 	return (error);
1705 }
1706 #endif /* ICL_KERNEL_PROXY */
1707 
1708 static int
1709 icl_soft_load(void)
1710 {
1711 	int error;
1712 
1713 	icl_soft_pdu_zone = uma_zcreate("icl_soft_pdu",
1714 	    sizeof(struct icl_soft_pdu), NULL, NULL, NULL, NULL,
1715 	    UMA_ALIGN_PTR, 0);
1716 	refcount_init(&icl_ncons, 0);
1717 
1718 	/*
1719 	 * The reason we call this "none" is that to the user,
1720 	 * it's known as "offload driver"; "offload driver: soft"
1721 	 * doesn't make much sense.
1722 	 */
1723 	error = icl_register("none", false, 0,
1724 	    icl_soft_limits, icl_soft_new_conn);
1725 	KASSERT(error == 0, ("failed to register"));
1726 
1727 #if defined(ICL_KERNEL_PROXY) && 0
1728 	/*
1729 	 * Debugging aid for kernel proxy functionality.
1730 	 */
1731 	error = icl_register("proxytest", true, 0,
1732 	    icl_soft_limits, icl_soft_new_conn);
1733 	KASSERT(error == 0, ("failed to register"));
1734 #endif
1735 
1736 	return (error);
1737 }
1738 
1739 static int
1740 icl_soft_unload(void)
1741 {
1742 
1743 	if (icl_ncons != 0)
1744 		return (EBUSY);
1745 
1746 	icl_unregister("none", false);
1747 #if defined(ICL_KERNEL_PROXY) && 0
1748 	icl_unregister("proxytest", true);
1749 #endif
1750 
1751 	uma_zdestroy(icl_soft_pdu_zone);
1752 
1753 	return (0);
1754 }
1755 
1756 static int
1757 icl_soft_modevent(module_t mod, int what, void *arg)
1758 {
1759 
1760 	switch (what) {
1761 	case MOD_LOAD:
1762 		return (icl_soft_load());
1763 	case MOD_UNLOAD:
1764 		return (icl_soft_unload());
1765 	default:
1766 		return (EINVAL);
1767 	}
1768 }
1769 
1770 moduledata_t icl_soft_data = {
1771 	"icl_soft",
1772 	icl_soft_modevent,
1773 	0
1774 };
1775 
1776 DECLARE_MODULE(icl_soft, icl_soft_data, SI_SUB_DRIVERS, SI_ORDER_MIDDLE);
1777 MODULE_DEPEND(icl_soft, icl, 1, 1, 1);
1778 MODULE_VERSION(icl_soft, 1);
1779