xref: /freebsd/sys/dev/cxgbe/cxgbei/icl_cxgbei.c (revision 87322a907545fa76fbaf7949f80e85b1377a53ad)
1 /*-
2  * Copyright (c) 2012 The FreeBSD Foundation
3  * Copyright (c) 2015 Chelsio Communications, Inc.
4  * All rights reserved.
5  *
6  * This software was developed by Edward Tomasz Napierala under sponsorship
7  * from the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  */
31 
32 /*
33  * cxgbei implementation of iSCSI Common Layer kobj(9) interface.
34  */
35 
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 #include "opt_inet.h"
40 #include "opt_inet6.h"
41 
42 #ifdef TCP_OFFLOAD
43 #include <sys/param.h>
44 #include <sys/capsicum.h>
45 #include <sys/condvar.h>
46 #include <sys/conf.h>
47 #include <sys/file.h>
48 #include <sys/kernel.h>
49 #include <sys/kthread.h>
50 #include <sys/ktr.h>
51 #include <sys/lock.h>
52 #include <sys/mbuf.h>
53 #include <sys/mutex.h>
54 #include <sys/module.h>
55 #include <sys/protosw.h>
56 #include <sys/socket.h>
57 #include <sys/socketvar.h>
58 #include <sys/sysctl.h>
59 #include <sys/systm.h>
60 #include <sys/sx.h>
61 #include <sys/uio.h>
62 #include <machine/bus.h>
63 #include <vm/vm.h>
64 #include <vm/pmap.h>
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp.h>
68 #include <netinet/tcp_var.h>
69 #include <netinet/toecore.h>
70 
71 #include <dev/iscsi/icl.h>
72 #include <dev/iscsi/iscsi_proto.h>
73 #include <icl_conn_if.h>
74 
75 #include <cam/scsi/scsi_all.h>
76 #include <cam/scsi/scsi_da.h>
77 #include <cam/ctl/ctl_io.h>
78 #include <cam/ctl/ctl.h>
79 #include <cam/ctl/ctl_backend.h>
80 #include <cam/ctl/ctl_error.h>
81 #include <cam/ctl/ctl_frontend.h>
82 #include <cam/ctl/ctl_debug.h>
83 #include <cam/ctl/ctl_ha.h>
84 #include <cam/ctl/ctl_ioctl.h>
85 
86 #include <cam/cam.h>
87 #include <cam/cam_ccb.h>
88 #include <cam/cam_xpt.h>
89 #include <cam/cam_debug.h>
90 #include <cam/cam_sim.h>
91 #include <cam/cam_xpt_sim.h>
92 #include <cam/cam_xpt_periph.h>
93 #include <cam/cam_periph.h>
94 #include <cam/cam_compat.h>
95 #include <cam/scsi/scsi_message.h>
96 
97 #include "common/common.h"
98 #include "common/t4_tcb.h"
99 #include "tom/t4_tom.h"
100 #include "cxgbei.h"
101 
102 /*
103  * Use the page pod tag for the TT hash.
104  */
105 #define	TT_HASH(icc, tt)	(G_PPOD_TAG(tt) & (icc)->cmp_hash_mask)
106 
107 struct cxgbei_ddp_state {
108 	struct ppod_reservation prsv;
109 	struct cxgbei_cmp cmp;
110 };
111 
112 static MALLOC_DEFINE(M_CXGBEI, "cxgbei", "cxgbei(4)");
113 
114 SYSCTL_NODE(_kern_icl, OID_AUTO, cxgbei, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
115     "Chelsio iSCSI offload");
116 static int first_burst_length = 8192;
117 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, first_burst_length, CTLFLAG_RWTUN,
118     &first_burst_length, 0, "First burst length");
119 static int max_burst_length = 2 * 1024 * 1024;
120 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, max_burst_length, CTLFLAG_RWTUN,
121     &max_burst_length, 0, "Maximum burst length");
122 static int sendspace = 1048576;
123 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, sendspace, CTLFLAG_RWTUN,
124     &sendspace, 0, "Default send socket buffer size");
125 static int recvspace = 1048576;
126 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, recvspace, CTLFLAG_RWTUN,
127     &recvspace, 0, "Default receive socket buffer size");
128 
129 static volatile u_int icl_cxgbei_ncons;
130 
131 #define ICL_CONN_LOCK(X)		mtx_lock(X->ic_lock)
132 #define ICL_CONN_UNLOCK(X)		mtx_unlock(X->ic_lock)
133 #define ICL_CONN_LOCK_ASSERT(X)		mtx_assert(X->ic_lock, MA_OWNED)
134 #define ICL_CONN_LOCK_ASSERT_NOT(X)	mtx_assert(X->ic_lock, MA_NOTOWNED)
135 
136 static icl_conn_new_pdu_t	icl_cxgbei_conn_new_pdu;
137 static icl_conn_pdu_data_segment_length_t
138 				    icl_cxgbei_conn_pdu_data_segment_length;
139 static icl_conn_pdu_append_data_t	icl_cxgbei_conn_pdu_append_data;
140 static icl_conn_pdu_get_data_t	icl_cxgbei_conn_pdu_get_data;
141 static icl_conn_pdu_queue_t	icl_cxgbei_conn_pdu_queue;
142 static icl_conn_pdu_queue_cb_t	icl_cxgbei_conn_pdu_queue_cb;
143 static icl_conn_handoff_t	icl_cxgbei_conn_handoff;
144 static icl_conn_free_t		icl_cxgbei_conn_free;
145 static icl_conn_close_t		icl_cxgbei_conn_close;
146 static icl_conn_task_setup_t	icl_cxgbei_conn_task_setup;
147 static icl_conn_task_done_t	icl_cxgbei_conn_task_done;
148 static icl_conn_transfer_setup_t	icl_cxgbei_conn_transfer_setup;
149 static icl_conn_transfer_done_t	icl_cxgbei_conn_transfer_done;
150 
151 static kobj_method_t icl_cxgbei_methods[] = {
152 	KOBJMETHOD(icl_conn_new_pdu, icl_cxgbei_conn_new_pdu),
153 	KOBJMETHOD(icl_conn_pdu_free, icl_cxgbei_conn_pdu_free),
154 	KOBJMETHOD(icl_conn_pdu_data_segment_length,
155 	    icl_cxgbei_conn_pdu_data_segment_length),
156 	KOBJMETHOD(icl_conn_pdu_append_data, icl_cxgbei_conn_pdu_append_data),
157 	KOBJMETHOD(icl_conn_pdu_get_data, icl_cxgbei_conn_pdu_get_data),
158 	KOBJMETHOD(icl_conn_pdu_queue, icl_cxgbei_conn_pdu_queue),
159 	KOBJMETHOD(icl_conn_pdu_queue_cb, icl_cxgbei_conn_pdu_queue_cb),
160 	KOBJMETHOD(icl_conn_handoff, icl_cxgbei_conn_handoff),
161 	KOBJMETHOD(icl_conn_free, icl_cxgbei_conn_free),
162 	KOBJMETHOD(icl_conn_close, icl_cxgbei_conn_close),
163 	KOBJMETHOD(icl_conn_task_setup, icl_cxgbei_conn_task_setup),
164 	KOBJMETHOD(icl_conn_task_done, icl_cxgbei_conn_task_done),
165 	KOBJMETHOD(icl_conn_transfer_setup, icl_cxgbei_conn_transfer_setup),
166 	KOBJMETHOD(icl_conn_transfer_done, icl_cxgbei_conn_transfer_done),
167 	{ 0, 0 }
168 };
169 
170 DEFINE_CLASS(icl_cxgbei, icl_cxgbei_methods, sizeof(struct icl_cxgbei_conn));
171 
172 void
173 icl_cxgbei_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip)
174 {
175 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
176 
177 	KASSERT(icp->ref_cnt != 0, ("freeing deleted PDU"));
178 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
179 	MPASS(ic == ip->ip_conn);
180 
181 	m_freem(ip->ip_ahs_mbuf);
182 	m_freem(ip->ip_data_mbuf);
183 	m_freem(ip->ip_bhs_mbuf);
184 
185 	KASSERT(ic != NULL || icp->ref_cnt == 1,
186 	    ("orphaned PDU has oustanding references"));
187 
188 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) != 1)
189 		return;
190 
191 	free(icp, M_CXGBEI);
192 #ifdef DIAGNOSTIC
193 	if (__predict_true(ic != NULL))
194 		refcount_release(&ic->ic_outstanding_pdus);
195 #endif
196 }
197 
198 static void
199 icl_cxgbei_pdu_call_cb(struct icl_pdu *ip)
200 {
201 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
202 
203 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
204 
205 	if (icp->cb != NULL)
206 		icp->cb(ip, icp->error);
207 #ifdef DIAGNOSTIC
208 	if (__predict_true(ip->ip_conn != NULL))
209 		refcount_release(&ip->ip_conn->ic_outstanding_pdus);
210 #endif
211 	free(icp, M_CXGBEI);
212 }
213 
214 static void
215 icl_cxgbei_pdu_done(struct icl_pdu *ip, int error)
216 {
217 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
218 
219 	if (error != 0)
220 		icp->error = error;
221 
222 	m_freem(ip->ip_ahs_mbuf);
223 	ip->ip_ahs_mbuf = NULL;
224 	m_freem(ip->ip_data_mbuf);
225 	ip->ip_data_mbuf = NULL;
226 	m_freem(ip->ip_bhs_mbuf);
227 	ip->ip_bhs_mbuf = NULL;
228 
229 	/*
230 	 * All other references to this PDU should have been dropped
231 	 * by the m_freem() of ip_data_mbuf.
232 	 */
233 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1)
234 		icl_cxgbei_pdu_call_cb(ip);
235 	else
236 		__assert_unreachable();
237 }
238 
239 static void
240 icl_cxgbei_mbuf_done(struct mbuf *mb)
241 {
242 
243 	struct icl_cxgbei_pdu *icp = (struct icl_cxgbei_pdu *)mb->m_ext.ext_arg1;
244 
245 	/*
246 	 * NB: mb_free_mext() might leave ref_cnt as 1 without
247 	 * decrementing it if it hits the fast path in the ref_cnt
248 	 * check.
249 	 */
250 	icl_cxgbei_pdu_call_cb(&icp->ip);
251 }
252 
253 struct icl_pdu *
254 icl_cxgbei_new_pdu(int flags)
255 {
256 	struct icl_cxgbei_pdu *icp;
257 	struct icl_pdu *ip;
258 	struct mbuf *m;
259 
260 	icp = malloc(sizeof(*icp), M_CXGBEI, flags | M_ZERO);
261 	if (__predict_false(icp == NULL))
262 		return (NULL);
263 
264 	icp->icp_signature = CXGBEI_PDU_SIGNATURE;
265 	icp->ref_cnt = 1;
266 	ip = &icp->ip;
267 
268 	m = m_gethdr(flags, MT_DATA);
269 	if (__predict_false(m == NULL)) {
270 		free(icp, M_CXGBEI);
271 		return (NULL);
272 	}
273 
274 	ip->ip_bhs_mbuf = m;
275 	ip->ip_bhs = mtod(m, struct iscsi_bhs *);
276 	memset(ip->ip_bhs, 0, sizeof(*ip->ip_bhs));
277 	m->m_len = sizeof(struct iscsi_bhs);
278 	m->m_pkthdr.len = m->m_len;
279 
280 	return (ip);
281 }
282 
283 void
284 icl_cxgbei_new_pdu_set_conn(struct icl_pdu *ip, struct icl_conn *ic)
285 {
286 
287 	ip->ip_conn = ic;
288 #ifdef DIAGNOSTIC
289 	refcount_acquire(&ic->ic_outstanding_pdus);
290 #endif
291 }
292 
293 /*
294  * Allocate icl_pdu with empty BHS to fill up by the caller.
295  */
296 static struct icl_pdu *
297 icl_cxgbei_conn_new_pdu(struct icl_conn *ic, int flags)
298 {
299 	struct icl_pdu *ip;
300 
301 	ip = icl_cxgbei_new_pdu(flags);
302 	if (__predict_false(ip == NULL))
303 		return (NULL);
304 	icl_cxgbei_new_pdu_set_conn(ip, ic);
305 
306 	return (ip);
307 }
308 
309 static size_t
310 icl_pdu_data_segment_length(const struct icl_pdu *request)
311 {
312 	uint32_t len = 0;
313 
314 	len += request->ip_bhs->bhs_data_segment_len[0];
315 	len <<= 8;
316 	len += request->ip_bhs->bhs_data_segment_len[1];
317 	len <<= 8;
318 	len += request->ip_bhs->bhs_data_segment_len[2];
319 
320 	return (len);
321 }
322 
323 size_t
324 icl_cxgbei_conn_pdu_data_segment_length(struct icl_conn *ic,
325     const struct icl_pdu *request)
326 {
327 
328 	return (icl_pdu_data_segment_length(request));
329 }
330 
331 static struct mbuf *
332 finalize_pdu(struct icl_cxgbei_conn *icc, struct icl_cxgbei_pdu *icp)
333 {
334 	struct icl_pdu *ip = &icp->ip;
335 	uint8_t ulp_submode, padding;
336 	struct mbuf *m, *last;
337 	struct iscsi_bhs *bhs;
338 
339 	/*
340 	 * Fix up the data segment mbuf first.
341 	 */
342 	m = ip->ip_data_mbuf;
343 	ulp_submode = icc->ulp_submode;
344 	if (m) {
345 		last = m_last(m);
346 
347 		/*
348 		 * Round up the data segment to a 4B boundary.	Pad with 0 if
349 		 * necessary.  There will definitely be room in the mbuf.
350 		 */
351 		padding = roundup2(ip->ip_data_len, 4) - ip->ip_data_len;
352 		if (padding) {
353 			bzero(mtod(last, uint8_t *) + last->m_len, padding);
354 			last->m_len += padding;
355 		}
356 	} else {
357 		MPASS(ip->ip_data_len == 0);
358 		ulp_submode &= ~ULP_CRC_DATA;
359 		padding = 0;
360 	}
361 
362 	/*
363 	 * Now the header mbuf that has the BHS.
364 	 */
365 	m = ip->ip_bhs_mbuf;
366 	MPASS(m->m_pkthdr.len == sizeof(struct iscsi_bhs));
367 	MPASS(m->m_len == sizeof(struct iscsi_bhs));
368 
369 	bhs = ip->ip_bhs;
370 	bhs->bhs_data_segment_len[2] = ip->ip_data_len;
371 	bhs->bhs_data_segment_len[1] = ip->ip_data_len >> 8;
372 	bhs->bhs_data_segment_len[0] = ip->ip_data_len >> 16;
373 
374 	/*
375 	 * Extract mbuf chain from PDU.
376 	 */
377 	m->m_pkthdr.len += ip->ip_data_len + padding;
378 	m->m_next = ip->ip_data_mbuf;
379 	set_mbuf_ulp_submode(m, ulp_submode);
380 	ip->ip_bhs_mbuf = NULL;
381 	ip->ip_data_mbuf = NULL;
382 	ip->ip_bhs = NULL;
383 
384 	/*
385 	 * Drop PDU reference on icp.  Additional references might
386 	 * still be held by zero-copy PDU buffers (ICL_NOCOPY).
387 	 */
388 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1)
389 		icl_cxgbei_pdu_call_cb(ip);
390 
391 	return (m);
392 }
393 
394 int
395 icl_cxgbei_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *ip,
396     const void *addr, size_t len, int flags)
397 {
398 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
399 	struct mbuf *m, *m_tail;
400 	const char *src;
401 
402 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
403 	MPASS(ic == ip->ip_conn);
404 	KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len));
405 
406 	m_tail = ip->ip_data_mbuf;
407 	if (m_tail != NULL)
408 		for (; m_tail->m_next != NULL; m_tail = m_tail->m_next)
409 			;
410 
411 	if (flags & ICL_NOCOPY) {
412 		m = m_get(flags & ~ICL_NOCOPY, MT_DATA);
413 		if (m == NULL) {
414 			ICL_WARN("failed to allocate mbuf");
415 			return (ENOMEM);
416 		}
417 
418 		m->m_flags |= M_RDONLY;
419 		m_extaddref(m, __DECONST(char *, addr), len, &icp->ref_cnt,
420 		    icl_cxgbei_mbuf_done, icp, NULL);
421 		m->m_len = len;
422 		if (ip->ip_data_mbuf == NULL) {
423 			ip->ip_data_mbuf = m;
424 			ip->ip_data_len = len;
425 		} else {
426 			m_tail->m_next = m;
427 			m_tail = m_tail->m_next;
428 			ip->ip_data_len += len;
429 		}
430 
431 		return (0);
432 	}
433 
434 	src = (const char *)addr;
435 
436 	/* Allocate as jumbo mbufs of size MJUM16BYTES. */
437 	while (len >= MJUM16BYTES) {
438 		m = m_getjcl(M_NOWAIT, MT_DATA, 0, MJUM16BYTES);
439 		if (__predict_false(m == NULL)) {
440 			if ((flags & M_WAITOK) != 0) {
441 				/* Fall back to non-jumbo mbufs. */
442 				break;
443 			}
444 			return (ENOMEM);
445 		}
446 		memcpy(mtod(m, void *), src, MJUM16BYTES);
447 		m->m_len = MJUM16BYTES;
448 		if (ip->ip_data_mbuf == NULL) {
449 			ip->ip_data_mbuf = m_tail = m;
450 			ip->ip_data_len = MJUM16BYTES;
451 		} else {
452 			m_tail->m_next = m;
453 			m_tail = m_tail->m_next;
454 			ip->ip_data_len += MJUM16BYTES;
455 		}
456 		src += MJUM16BYTES;
457 		len -= MJUM16BYTES;
458 	}
459 
460 	/* Allocate mbuf chain for the remaining data. */
461 	if (len != 0) {
462 		m = m_getm2(NULL, len, flags, MT_DATA, 0);
463 		if (__predict_false(m == NULL))
464 			return (ENOMEM);
465 		if (ip->ip_data_mbuf == NULL) {
466 			ip->ip_data_mbuf = m;
467 			ip->ip_data_len = len;
468 		} else {
469 			m_tail->m_next = m;
470 			ip->ip_data_len += len;
471 		}
472 		for (; m != NULL; m = m->m_next) {
473 			m->m_len = min(len, M_SIZE(m));
474 			memcpy(mtod(m, void *), src, m->m_len);
475 			src += m->m_len;
476 			len -= m->m_len;
477 		}
478 		MPASS(len == 0);
479 	}
480 	MPASS(ip->ip_data_len <= ic->ic_max_send_data_segment_length);
481 
482 	return (0);
483 }
484 
485 void
486 icl_cxgbei_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip,
487     size_t off, void *addr, size_t len)
488 {
489 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
490 
491 	if (icp->icp_flags & ICPF_RX_DDP)
492 		return; /* data is DDP'ed, no need to copy */
493 	m_copydata(ip->ip_data_mbuf, off, len, addr);
494 }
495 
496 void
497 icl_cxgbei_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip)
498 {
499 	icl_cxgbei_conn_pdu_queue_cb(ic, ip, NULL);
500 }
501 
502 void
503 icl_cxgbei_conn_pdu_queue_cb(struct icl_conn *ic, struct icl_pdu *ip,
504 			     icl_pdu_cb cb)
505 {
506 	struct epoch_tracker et;
507 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
508 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
509 	struct socket *so = ic->ic_socket;
510 	struct toepcb *toep = icc->toep;
511 	struct inpcb *inp;
512 	struct mbuf *m;
513 
514 	MPASS(ic == ip->ip_conn);
515 	MPASS(ip->ip_bhs_mbuf != NULL);
516 	/* The kernel doesn't generate PDUs with AHS. */
517 	MPASS(ip->ip_ahs_mbuf == NULL && ip->ip_ahs_len == 0);
518 
519 	ICL_CONN_LOCK_ASSERT(ic);
520 
521 	icp->cb = cb;
522 
523 	/* NOTE: sowriteable without so_snd lock is a mostly harmless race. */
524 	if (ic->ic_disconnecting || so == NULL || !sowriteable(so)) {
525 		icl_cxgbei_pdu_done(ip, ENOTCONN);
526 		return;
527 	}
528 
529 	m = finalize_pdu(icc, icp);
530 	M_ASSERTPKTHDR(m);
531 	MPASS((m->m_pkthdr.len & 3) == 0);
532 
533 	/*
534 	 * Do not get inp from toep->inp as the toepcb might have detached
535 	 * already.
536 	 */
537 	inp = sotoinpcb(so);
538 	CURVNET_SET(toep->vnet);
539 	NET_EPOCH_ENTER(et);
540 	INP_WLOCK(inp);
541 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) ||
542 	    __predict_false((toep->flags & TPF_ATTACHED) == 0))
543 		m_freem(m);
544 	else {
545 		mbufq_enqueue(&toep->ulp_pduq, m);
546 		t4_push_pdus(icc->sc, toep, 0);
547 	}
548 	INP_WUNLOCK(inp);
549 	NET_EPOCH_EXIT(et);
550 	CURVNET_RESTORE();
551 }
552 
553 static struct icl_conn *
554 icl_cxgbei_new_conn(const char *name, struct mtx *lock)
555 {
556 	struct icl_cxgbei_conn *icc;
557 	struct icl_conn *ic;
558 
559 	refcount_acquire(&icl_cxgbei_ncons);
560 
561 	icc = (struct icl_cxgbei_conn *)kobj_create(&icl_cxgbei_class, M_CXGBE,
562 	    M_WAITOK | M_ZERO);
563 	icc->icc_signature = CXGBEI_CONN_SIGNATURE;
564 	STAILQ_INIT(&icc->rcvd_pdus);
565 
566 	icc->cmp_table = hashinit(64, M_CXGBEI, &icc->cmp_hash_mask);
567 	mtx_init(&icc->cmp_lock, "cxgbei_cmp", NULL, MTX_DEF);
568 
569 	ic = &icc->ic;
570 	ic->ic_lock = lock;
571 
572 #ifdef DIAGNOSTIC
573 	refcount_init(&ic->ic_outstanding_pdus, 0);
574 #endif
575 	ic->ic_name = name;
576 	ic->ic_offload = "cxgbei";
577 	ic->ic_unmapped = false;
578 
579 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
580 
581 	return (ic);
582 }
583 
584 void
585 icl_cxgbei_conn_free(struct icl_conn *ic)
586 {
587 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
588 
589 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
590 
591 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
592 
593 	mtx_destroy(&icc->cmp_lock);
594 	hashdestroy(icc->cmp_table, M_CXGBEI, icc->cmp_hash_mask);
595 	kobj_delete((struct kobj *)icc, M_CXGBE);
596 	refcount_release(&icl_cxgbei_ncons);
597 }
598 
599 static int
600 icl_cxgbei_setsockopt(struct icl_conn *ic, struct socket *so, int sspace,
601     int rspace)
602 {
603 	struct sockopt opt;
604 	int error, one = 1, ss, rs;
605 
606 	ss = max(sendspace, sspace);
607 	rs = max(recvspace, rspace);
608 
609 	error = soreserve(so, ss, rs);
610 	if (error != 0) {
611 		icl_cxgbei_conn_close(ic);
612 		return (error);
613 	}
614 	SOCKBUF_LOCK(&so->so_snd);
615 	so->so_snd.sb_flags |= SB_AUTOSIZE;
616 	SOCKBUF_UNLOCK(&so->so_snd);
617 	SOCKBUF_LOCK(&so->so_rcv);
618 	so->so_rcv.sb_flags |= SB_AUTOSIZE;
619 	SOCKBUF_UNLOCK(&so->so_rcv);
620 
621 	/*
622 	 * Disable Nagle.
623 	 */
624 	bzero(&opt, sizeof(opt));
625 	opt.sopt_dir = SOPT_SET;
626 	opt.sopt_level = IPPROTO_TCP;
627 	opt.sopt_name = TCP_NODELAY;
628 	opt.sopt_val = &one;
629 	opt.sopt_valsize = sizeof(one);
630 	error = sosetopt(so, &opt);
631 	if (error != 0) {
632 		icl_cxgbei_conn_close(ic);
633 		return (error);
634 	}
635 
636 	return (0);
637 }
638 
639 /*
640  * Request/response structure used to find out the adapter offloading a socket.
641  */
642 struct find_ofld_adapter_rr {
643 	struct socket *so;
644 	struct adapter *sc;	/* result */
645 };
646 
647 static void
648 find_offload_adapter(struct adapter *sc, void *arg)
649 {
650 	struct find_ofld_adapter_rr *fa = arg;
651 	struct socket *so = fa->so;
652 	struct tom_data *td = sc->tom_softc;
653 	struct tcpcb *tp;
654 	struct inpcb *inp;
655 
656 	/* Non-TCP were filtered out earlier. */
657 	MPASS(so->so_proto->pr_protocol == IPPROTO_TCP);
658 
659 	if (fa->sc != NULL)
660 		return;	/* Found already. */
661 
662 	if (td == NULL)
663 		return;	/* TOE not enabled on this adapter. */
664 
665 	inp = sotoinpcb(so);
666 	INP_WLOCK(inp);
667 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
668 		tp = intotcpcb(inp);
669 		if (tp->t_flags & TF_TOE && tp->tod == &td->tod)
670 			fa->sc = sc;	/* Found. */
671 	}
672 	INP_WUNLOCK(inp);
673 }
674 
675 /* XXXNP: move this to t4_tom. */
676 static void
677 send_iscsi_flowc_wr(struct adapter *sc, struct toepcb *toep, int maxlen)
678 {
679 	struct wrqe *wr;
680 	struct fw_flowc_wr *flowc;
681 	const u_int nparams = 1;
682 	u_int flowclen;
683 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
684 
685 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
686 
687 	wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq);
688 	if (wr == NULL) {
689 		/* XXX */
690 		panic("%s: allocation failure.", __func__);
691 	}
692 	flowc = wrtod(wr);
693 	memset(flowc, 0, wr->wr_len);
694 
695 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
696 	    V_FW_FLOWC_WR_NPARAMS(nparams));
697 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
698 	    V_FW_WR_FLOWID(toep->tid));
699 
700 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_TXDATAPLEN_MAX;
701 	flowc->mnemval[0].val = htobe32(maxlen);
702 
703 	txsd->tx_credits = howmany(flowclen, 16);
704 	txsd->plen = 0;
705 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
706 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
707 	toep->tx_credits -= txsd->tx_credits;
708 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
709 		toep->txsd_pidx = 0;
710 	toep->txsd_avail--;
711 
712 	t4_wrq_tx(sc, wr);
713 }
714 
715 static void
716 set_ulp_mode_iscsi(struct adapter *sc, struct toepcb *toep, u_int ulp_submode)
717 {
718 	uint64_t val;
719 
720 	CTR3(KTR_CXGBE, "%s: tid %u, ULP_MODE_ISCSI, submode=%#x",
721 	    __func__, toep->tid, ulp_submode);
722 
723 	val = V_TCB_ULP_TYPE(ULP_MODE_ISCSI) | V_TCB_ULP_RAW(ulp_submode);
724 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_ULP_TYPE,
725 	    V_TCB_ULP_TYPE(M_TCB_ULP_TYPE) | V_TCB_ULP_RAW(M_TCB_ULP_RAW), val,
726 	    0, 0);
727 
728 	val = V_TF_RX_FLOW_CONTROL_DISABLE(1ULL);
729 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, val, val, 0, 0);
730 }
731 
732 /*
733  * XXXNP: Who is responsible for cleaning up the socket if this returns with an
734  * error?  Review all error paths.
735  *
736  * XXXNP: What happens to the socket's fd reference if the operation is
737  * successful, and how does that affect the socket's life cycle?
738  */
739 int
740 icl_cxgbei_conn_handoff(struct icl_conn *ic, int fd)
741 {
742 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
743 	struct cxgbei_data *ci;
744 	struct find_ofld_adapter_rr fa;
745 	struct file *fp;
746 	struct socket *so;
747 	struct inpcb *inp;
748 	struct tcpcb *tp;
749 	struct toepcb *toep;
750 	cap_rights_t rights;
751 	int error;
752 
753 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
754 	ICL_CONN_LOCK_ASSERT_NOT(ic);
755 
756 	/*
757 	 * Steal the socket from userland.
758 	 */
759 	error = fget(curthread, fd,
760 	    cap_rights_init_one(&rights, CAP_SOCK_CLIENT), &fp);
761 	if (error != 0)
762 		return (error);
763 	if (fp->f_type != DTYPE_SOCKET) {
764 		fdrop(fp, curthread);
765 		return (EINVAL);
766 	}
767 	so = fp->f_data;
768 	if (so->so_type != SOCK_STREAM ||
769 	    so->so_proto->pr_protocol != IPPROTO_TCP) {
770 		fdrop(fp, curthread);
771 		return (EINVAL);
772 	}
773 
774 	ICL_CONN_LOCK(ic);
775 	if (ic->ic_socket != NULL) {
776 		ICL_CONN_UNLOCK(ic);
777 		fdrop(fp, curthread);
778 		return (EBUSY);
779 	}
780 	ic->ic_disconnecting = false;
781 	ic->ic_socket = so;
782 	fp->f_ops = &badfileops;
783 	fp->f_data = NULL;
784 	fdrop(fp, curthread);
785 	ICL_CONN_UNLOCK(ic);
786 
787 	/* Find the adapter offloading this socket. */
788 	fa.sc = NULL;
789 	fa.so = so;
790 	t4_iterate(find_offload_adapter, &fa);
791 	if (fa.sc == NULL)
792 		return (EINVAL);
793 	icc->sc = fa.sc;
794 	ci = icc->sc->iscsi_ulp_softc;
795 
796 	inp = sotoinpcb(so);
797 	INP_WLOCK(inp);
798 	tp = intotcpcb(inp);
799 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))
800 		error = EBUSY;
801 	else {
802 		/*
803 		 * socket could not have been "unoffloaded" if here.
804 		 */
805 		MPASS(tp->t_flags & TF_TOE);
806 		MPASS(tp->tod != NULL);
807 		MPASS(tp->t_toe != NULL);
808 		toep = tp->t_toe;
809 		MPASS(toep->vi->adapter == icc->sc);
810 		icc->toep = toep;
811 		icc->cwt = cxgbei_select_worker_thread(icc);
812 
813 		icc->ulp_submode = 0;
814 		if (ic->ic_header_crc32c)
815 			icc->ulp_submode |= ULP_CRC_HEADER;
816 		if (ic->ic_data_crc32c)
817 			icc->ulp_submode |= ULP_CRC_DATA;
818 		so->so_options |= SO_NO_DDP;
819 		toep->params.ulp_mode = ULP_MODE_ISCSI;
820 		toep->ulpcb = icc;
821 
822 		send_iscsi_flowc_wr(icc->sc, toep, roundup(ci->max_tx_pdu_len,
823 		    tp->t_maxseg));
824 		set_ulp_mode_iscsi(icc->sc, toep, icc->ulp_submode);
825 		error = 0;
826 	}
827 	INP_WUNLOCK(inp);
828 
829 	if (error == 0) {
830 		error = icl_cxgbei_setsockopt(ic, so, ci->max_tx_pdu_len,
831 		    ci->max_rx_pdu_len);
832 	}
833 
834 	return (error);
835 }
836 
837 void
838 icl_cxgbei_conn_close(struct icl_conn *ic)
839 {
840 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
841 	struct icl_pdu *ip;
842 	struct socket *so;
843 	struct sockbuf *sb;
844 	struct inpcb *inp;
845 	struct toepcb *toep = icc->toep;
846 
847 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
848 	ICL_CONN_LOCK_ASSERT_NOT(ic);
849 
850 	ICL_CONN_LOCK(ic);
851 	so = ic->ic_socket;
852 	if (ic->ic_disconnecting || so == NULL) {
853 		CTR4(KTR_CXGBE, "%s: icc %p (disconnecting = %d), so %p",
854 		    __func__, icc, ic->ic_disconnecting, so);
855 		ICL_CONN_UNLOCK(ic);
856 		return;
857 	}
858 	ic->ic_disconnecting = true;
859 
860 #ifdef DIAGNOSTIC
861 	KASSERT(ic->ic_outstanding_pdus == 0,
862 	    ("destroying session with %d outstanding PDUs",
863 	     ic->ic_outstanding_pdus));
864 #endif
865 	ICL_CONN_UNLOCK(ic);
866 
867 	CTR3(KTR_CXGBE, "%s: tid %d, icc %p", __func__, toep ? toep->tid : -1,
868 	    icc);
869 	inp = sotoinpcb(so);
870 	sb = &so->so_rcv;
871 	INP_WLOCK(inp);
872 	if (toep != NULL) {	/* NULL if connection was never offloaded. */
873 		toep->ulpcb = NULL;
874 
875 		/* Discard PDUs queued for TX. */
876 		mbufq_drain(&toep->ulp_pduq);
877 
878 		/*
879 		 * Wait for the cwt threads to stop processing this
880 		 * connection.
881 		 */
882 		SOCKBUF_LOCK(sb);
883 		if (icc->rx_flags & RXF_ACTIVE) {
884 			volatile u_int *p = &icc->rx_flags;
885 
886 			SOCKBUF_UNLOCK(sb);
887 			INP_WUNLOCK(inp);
888 
889 			while (*p & RXF_ACTIVE)
890 				pause("conclo", 1);
891 
892 			INP_WLOCK(inp);
893 			SOCKBUF_LOCK(sb);
894 		}
895 
896 		/*
897 		 * Discard received PDUs not passed to the iSCSI
898 		 * layer.
899 		 */
900 		while (!STAILQ_EMPTY(&icc->rcvd_pdus)) {
901 			ip = STAILQ_FIRST(&icc->rcvd_pdus);
902 			STAILQ_REMOVE_HEAD(&icc->rcvd_pdus, ip_next);
903 			icl_cxgbei_pdu_done(ip, ENOTCONN);
904 		}
905 		SOCKBUF_UNLOCK(sb);
906 	}
907 	INP_WUNLOCK(inp);
908 
909 	ICL_CONN_LOCK(ic);
910 	ic->ic_socket = NULL;
911 	ICL_CONN_UNLOCK(ic);
912 
913 	/*
914 	 * XXXNP: we should send RST instead of FIN when PDUs held in various
915 	 * queues were purged instead of delivered reliably but soabort isn't
916 	 * really general purpose and wouldn't do the right thing here.
917 	 */
918 	soref(so);
919 	soclose(so);
920 
921 	/*
922 	 * Wait for the socket to fully close.  This ensures any
923 	 * pending received data has been received (and in particular,
924 	 * any data that would be received by DDP has been handled).
925 	 * Callers assume that it is safe to free buffers for tasks
926 	 * and transfers after this function returns.
927 	 */
928 	SOCK_LOCK(so);
929 	while ((so->so_state & SS_ISDISCONNECTED) == 0)
930 		mtx_sleep(&so->so_timeo, SOCK_MTX(so), PSOCK, "conclo2", 0);
931 	CURVNET_SET(so->so_vnet);
932 	sorele(so);
933 	CURVNET_RESTORE();
934 }
935 
936 static void
937 cxgbei_insert_cmp(struct icl_cxgbei_conn *icc, struct cxgbei_cmp *cmp,
938     uint32_t tt)
939 {
940 #ifdef INVARIANTS
941 	struct cxgbei_cmp *cmp2;
942 #endif
943 
944 	cmp->tt = tt;
945 
946 	mtx_lock(&icc->cmp_lock);
947 #ifdef INVARIANTS
948 	LIST_FOREACH(cmp2, &icc->cmp_table[TT_HASH(icc, tt)], link) {
949 		KASSERT(cmp2->tt != tt, ("%s: duplicate cmp", __func__));
950 	}
951 #endif
952 	LIST_INSERT_HEAD(&icc->cmp_table[TT_HASH(icc, tt)], cmp, link);
953 	mtx_unlock(&icc->cmp_lock);
954 }
955 
956 struct cxgbei_cmp *
957 cxgbei_find_cmp(struct icl_cxgbei_conn *icc, uint32_t tt)
958 {
959 	struct cxgbei_cmp *cmp;
960 
961 	mtx_lock(&icc->cmp_lock);
962 	LIST_FOREACH(cmp, &icc->cmp_table[TT_HASH(icc, tt)], link) {
963 		if (cmp->tt == tt)
964 			break;
965 	}
966 	mtx_unlock(&icc->cmp_lock);
967 	return (cmp);
968 }
969 
970 static void
971 cxgbei_rm_cmp(struct icl_cxgbei_conn *icc, struct cxgbei_cmp *cmp)
972 {
973 #ifdef INVARIANTS
974 	struct cxgbei_cmp *cmp2;
975 #endif
976 
977 	mtx_lock(&icc->cmp_lock);
978 
979 #ifdef INVARIANTS
980 	LIST_FOREACH(cmp2, &icc->cmp_table[TT_HASH(icc, cmp->tt)], link) {
981 		if (cmp2 == cmp)
982 			goto found;
983 	}
984 	panic("%s: could not find cmp", __func__);
985 found:
986 #endif
987 	LIST_REMOVE(cmp, link);
988 	mtx_unlock(&icc->cmp_lock);
989 }
990 
991 int
992 icl_cxgbei_conn_task_setup(struct icl_conn *ic, struct icl_pdu *ip,
993     struct ccb_scsiio *csio, uint32_t *ittp, void **arg)
994 {
995 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
996 	struct toepcb *toep = icc->toep;
997 	struct adapter *sc = icc->sc;
998 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
999 	struct ppod_region *pr = &ci->pr;
1000 	struct cxgbei_ddp_state *ddp;
1001 	struct ppod_reservation *prsv;
1002 	struct inpcb *inp;
1003 	struct mbufq mq;
1004 	uint32_t itt;
1005 	int rc = 0;
1006 
1007 	ICL_CONN_LOCK_ASSERT(ic);
1008 
1009 	/* This is for the offload driver's state.  Must not be set already. */
1010 	MPASS(arg != NULL);
1011 	MPASS(*arg == NULL);
1012 
1013 	if (ic->ic_disconnecting || ic->ic_socket == NULL)
1014 		return (ECONNRESET);
1015 
1016 	if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_IN ||
1017 	    csio->dxfer_len < ci->ddp_threshold) {
1018 no_ddp:
1019 		/*
1020 		 * No DDP for this I/O.	 Allocate an ITT (based on the one
1021 		 * passed in) that cannot be a valid hardware DDP tag in the
1022 		 * iSCSI region.
1023 		 */
1024 		itt = *ittp & M_PPOD_TAG;
1025 		itt = V_PPOD_TAG(itt) | pr->pr_invalid_bit;
1026 		*ittp = htobe32(itt);
1027 		MPASS(*arg == NULL);	/* State is maintained for DDP only. */
1028 		if (rc != 0)
1029 			counter_u64_add(
1030 			    toep->ofld_rxq->rx_iscsi_ddp_setup_error, 1);
1031 		return (0);
1032 	}
1033 
1034 	/*
1035 	 * Reserve resources for DDP, update the itt that should be used in the
1036 	 * PDU, and save DDP specific state for this I/O in *arg.
1037 	 */
1038 	ddp = malloc(sizeof(*ddp), M_CXGBEI, M_NOWAIT | M_ZERO);
1039 	if (ddp == NULL) {
1040 		rc = ENOMEM;
1041 		goto no_ddp;
1042 	}
1043 	prsv = &ddp->prsv;
1044 
1045 	/* XXX add support for all CAM_DATA_ types */
1046 	MPASS((csio->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_VADDR);
1047 	rc = t4_alloc_page_pods_for_buf(pr, (vm_offset_t)csio->data_ptr,
1048 	    csio->dxfer_len, prsv);
1049 	if (rc != 0) {
1050 		free(ddp, M_CXGBEI);
1051 		goto no_ddp;
1052 	}
1053 
1054 	mbufq_init(&mq, INT_MAX);
1055 	rc = t4_write_page_pods_for_buf(sc, toep, prsv,
1056 	    (vm_offset_t)csio->data_ptr, csio->dxfer_len, &mq);
1057 	if (__predict_false(rc != 0)) {
1058 		mbufq_drain(&mq);
1059 		t4_free_page_pods(prsv);
1060 		free(ddp, M_CXGBEI);
1061 		goto no_ddp;
1062 	}
1063 
1064 	/*
1065 	 * Do not get inp from toep->inp as the toepcb might have
1066 	 * detached already.
1067 	 */
1068 	inp = sotoinpcb(ic->ic_socket);
1069 	INP_WLOCK(inp);
1070 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) != 0) {
1071 		INP_WUNLOCK(inp);
1072 		mbufq_drain(&mq);
1073 		t4_free_page_pods(prsv);
1074 		free(ddp, M_CXGBEI);
1075 		return (ECONNRESET);
1076 	}
1077 	mbufq_concat(&toep->ulp_pduq, &mq);
1078 	INP_WUNLOCK(inp);
1079 
1080 	ddp->cmp.last_datasn = -1;
1081 	cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag);
1082 	*ittp = htobe32(prsv->prsv_tag);
1083 	*arg = prsv;
1084 	counter_u64_add(toep->ofld_rxq->rx_iscsi_ddp_setup_ok, 1);
1085 	return (0);
1086 }
1087 
1088 void
1089 icl_cxgbei_conn_task_done(struct icl_conn *ic, void *arg)
1090 {
1091 
1092 	if (arg != NULL) {
1093 		struct cxgbei_ddp_state *ddp = arg;
1094 
1095 		cxgbei_rm_cmp(ic_to_icc(ic), &ddp->cmp);
1096 		t4_free_page_pods(&ddp->prsv);
1097 		free(ddp, M_CXGBEI);
1098 	}
1099 }
1100 
1101 static inline bool
1102 ddp_sgl_check(struct ctl_sg_entry *sg, int entries, int xferlen)
1103 {
1104 	int total_len = 0;
1105 
1106 	MPASS(entries > 0);
1107 	if (((vm_offset_t)sg[--entries].addr & 3U) != 0)
1108 		return (false);
1109 
1110 	total_len += sg[entries].len;
1111 
1112 	while (--entries >= 0) {
1113 		if (((vm_offset_t)sg[entries].addr & PAGE_MASK) != 0 ||
1114 		    (sg[entries].len % PAGE_SIZE) != 0)
1115 			return (false);
1116 		total_len += sg[entries].len;
1117 	}
1118 
1119 	MPASS(total_len == xferlen);
1120 	return (true);
1121 }
1122 
1123 /* XXXNP: PDU should be passed in as parameter, like on the initiator. */
1124 #define io_to_request_pdu(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND].ptr)
1125 #define io_to_ddp_state(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND2].ptr)
1126 
1127 int
1128 icl_cxgbei_conn_transfer_setup(struct icl_conn *ic, union ctl_io *io,
1129     uint32_t *tttp, void **arg)
1130 {
1131 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
1132 	struct toepcb *toep = icc->toep;
1133 	struct ctl_scsiio *ctsio = &io->scsiio;
1134 	struct adapter *sc = icc->sc;
1135 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
1136 	struct ppod_region *pr = &ci->pr;
1137 	struct cxgbei_ddp_state *ddp;
1138 	struct ppod_reservation *prsv;
1139 	struct ctl_sg_entry *sgl, sg_entry;
1140 	struct inpcb *inp;
1141 	struct mbufq mq;
1142 	int sg_entries = ctsio->kern_sg_entries;
1143 	uint32_t ttt;
1144 	int xferlen, rc = 0, alias;
1145 
1146 	/* This is for the offload driver's state.  Must not be set already. */
1147 	MPASS(arg != NULL);
1148 	MPASS(*arg == NULL);
1149 
1150 	if (ctsio->ext_data_filled == 0) {
1151 		int first_burst;
1152 		struct icl_pdu *ip = io_to_request_pdu(io);
1153 #ifdef INVARIANTS
1154 		struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
1155 
1156 		MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
1157 		MPASS(ic == ip->ip_conn);
1158 		MPASS(ip->ip_bhs_mbuf != NULL);
1159 #endif
1160 		first_burst = icl_pdu_data_segment_length(ip);
1161 
1162 		/*
1163 		 * Note that ICL calls conn_transfer_setup even if the first
1164 		 * burst had everything and there's nothing left to transfer.
1165 		 *
1166 		 * NB: The CTL frontend might have provided a buffer
1167 		 * whose length (kern_data_len) is smaller than the
1168 		 * FirstBurstLength of unsolicited data.  Treat those
1169 		 * as an empty transfer.
1170 		 */
1171 		xferlen = ctsio->kern_data_len;
1172 		if (xferlen < first_burst ||
1173 		    xferlen - first_burst < ci->ddp_threshold) {
1174 no_ddp:
1175 			/*
1176 			 * No DDP for this transfer.  Allocate a TTT (based on
1177 			 * the one passed in) that cannot be a valid hardware
1178 			 * DDP tag in the iSCSI region.
1179 			 */
1180 			ttt = *tttp & M_PPOD_TAG;
1181 			ttt = V_PPOD_TAG(ttt) | pr->pr_invalid_bit;
1182 			*tttp = htobe32(ttt);
1183 			MPASS(io_to_ddp_state(io) == NULL);
1184 			if (rc != 0)
1185 				counter_u64_add(
1186 				    toep->ofld_rxq->rx_iscsi_ddp_setup_error, 1);
1187 			return (0);
1188 		}
1189 
1190 		if (sg_entries == 0) {
1191 			sgl = &sg_entry;
1192 			sgl->len = xferlen;
1193 			sgl->addr = (void *)ctsio->kern_data_ptr;
1194 			sg_entries = 1;
1195 		} else
1196 			sgl = (void *)ctsio->kern_data_ptr;
1197 
1198 		if (!ddp_sgl_check(sgl, sg_entries, xferlen))
1199 			goto no_ddp;
1200 
1201 		/*
1202 		 * Reserve resources for DDP, update the ttt that should be used
1203 		 * in the PDU, and save DDP specific state for this I/O.
1204 		 */
1205 		MPASS(io_to_ddp_state(io) == NULL);
1206 		ddp = malloc(sizeof(*ddp), M_CXGBEI, M_NOWAIT | M_ZERO);
1207 		if (ddp == NULL) {
1208 			rc = ENOMEM;
1209 			goto no_ddp;
1210 		}
1211 		prsv = &ddp->prsv;
1212 
1213 		rc = t4_alloc_page_pods_for_sgl(pr, sgl, sg_entries, prsv);
1214 		if (rc != 0) {
1215 			free(ddp, M_CXGBEI);
1216 			goto no_ddp;
1217 		}
1218 
1219 		mbufq_init(&mq, INT_MAX);
1220 		rc = t4_write_page_pods_for_sgl(sc, toep, prsv, sgl, sg_entries,
1221 		    xferlen, &mq);
1222 		if (__predict_false(rc != 0)) {
1223 			mbufq_drain(&mq);
1224 			t4_free_page_pods(prsv);
1225 			free(ddp, M_CXGBEI);
1226 			goto no_ddp;
1227 		}
1228 
1229 		/*
1230 		 * Do not get inp from toep->inp as the toepcb might
1231 		 * have detached already.
1232 		 */
1233 		ICL_CONN_LOCK(ic);
1234 		if (ic->ic_disconnecting || ic->ic_socket == NULL) {
1235 			ICL_CONN_UNLOCK(ic);
1236 			mbufq_drain(&mq);
1237 			t4_free_page_pods(prsv);
1238 			free(ddp, M_CXGBEI);
1239 			return (ECONNRESET);
1240 		}
1241 		inp = sotoinpcb(ic->ic_socket);
1242 		INP_WLOCK(inp);
1243 		ICL_CONN_UNLOCK(ic);
1244 		if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) != 0) {
1245 			INP_WUNLOCK(inp);
1246 			mbufq_drain(&mq);
1247 			t4_free_page_pods(prsv);
1248 			free(ddp, M_CXGBEI);
1249 			return (ECONNRESET);
1250 		}
1251 		mbufq_concat(&toep->ulp_pduq, &mq);
1252 		INP_WUNLOCK(inp);
1253 
1254 		ddp->cmp.next_buffer_offset = ctsio->kern_rel_offset +
1255 		    first_burst;
1256 		ddp->cmp.last_datasn = -1;
1257 		cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag);
1258 		*tttp = htobe32(prsv->prsv_tag);
1259 		io_to_ddp_state(io) = ddp;
1260 		*arg = ctsio;
1261 		counter_u64_add(toep->ofld_rxq->rx_iscsi_ddp_setup_ok, 1);
1262 		return (0);
1263 	}
1264 
1265 	/*
1266 	 * In the middle of an I/O.  A non-NULL page pod reservation indicates
1267 	 * that a DDP buffer is being used for the I/O.
1268 	 */
1269 	ddp = io_to_ddp_state(ctsio);
1270 	if (ddp == NULL)
1271 		goto no_ddp;
1272 	prsv = &ddp->prsv;
1273 
1274 	alias = (prsv->prsv_tag & pr->pr_alias_mask) >> pr->pr_alias_shift;
1275 	alias++;
1276 	prsv->prsv_tag &= ~pr->pr_alias_mask;
1277 	prsv->prsv_tag |= alias << pr->pr_alias_shift & pr->pr_alias_mask;
1278 
1279 	ddp->cmp.next_datasn = 0;
1280 	ddp->cmp.last_datasn = -1;
1281 	cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag);
1282 	*tttp = htobe32(prsv->prsv_tag);
1283 	*arg = ctsio;
1284 
1285 	return (0);
1286 }
1287 
1288 void
1289 icl_cxgbei_conn_transfer_done(struct icl_conn *ic, void *arg)
1290 {
1291 	struct ctl_scsiio *ctsio = arg;
1292 
1293 	if (ctsio != NULL) {
1294 		struct cxgbei_ddp_state *ddp;
1295 
1296 		ddp = io_to_ddp_state(ctsio);
1297 		MPASS(ddp != NULL);
1298 
1299 		cxgbei_rm_cmp(ic_to_icc(ic), &ddp->cmp);
1300 		if (ctsio->kern_data_len == ctsio->ext_data_filled ||
1301 		    ic->ic_disconnecting) {
1302 			t4_free_page_pods(&ddp->prsv);
1303 			free(ddp, M_CXGBEI);
1304 			io_to_ddp_state(ctsio) = NULL;
1305 		}
1306 	}
1307 }
1308 
1309 static void
1310 cxgbei_limits(struct adapter *sc, void *arg)
1311 {
1312 	struct icl_drv_limits *idl = arg;
1313 	struct cxgbei_data *ci;
1314 	int max_dsl;
1315 
1316 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4lims") != 0)
1317 		return;
1318 
1319 	if (uld_active(sc, ULD_ISCSI)) {
1320 		ci = sc->iscsi_ulp_softc;
1321 		MPASS(ci != NULL);
1322 
1323 		/*
1324 		 * AHS is not supported by the kernel so we'll not account for
1325 		 * it either in our PDU len -> data segment len conversions.
1326 		 */
1327 
1328 		max_dsl = ci->max_rx_pdu_len - ISCSI_BHS_SIZE -
1329 		    ISCSI_HEADER_DIGEST_SIZE - ISCSI_DATA_DIGEST_SIZE;
1330 		if (idl->idl_max_recv_data_segment_length > max_dsl)
1331 			idl->idl_max_recv_data_segment_length = max_dsl;
1332 
1333 		max_dsl = ci->max_tx_pdu_len - ISCSI_BHS_SIZE -
1334 		    ISCSI_HEADER_DIGEST_SIZE - ISCSI_DATA_DIGEST_SIZE;
1335 		if (idl->idl_max_send_data_segment_length > max_dsl)
1336 			idl->idl_max_send_data_segment_length = max_dsl;
1337 	}
1338 
1339 	end_synchronized_op(sc, LOCK_HELD);
1340 }
1341 
1342 static int
1343 icl_cxgbei_limits(struct icl_drv_limits *idl)
1344 {
1345 
1346 	/* Maximum allowed by the RFC.	cxgbei_limits will clip them. */
1347 	idl->idl_max_recv_data_segment_length = (1 << 24) - 1;
1348 	idl->idl_max_send_data_segment_length = (1 << 24) - 1;
1349 
1350 	/* These are somewhat arbitrary. */
1351 	idl->idl_max_burst_length = max_burst_length;
1352 	idl->idl_first_burst_length = first_burst_length;
1353 
1354 	t4_iterate(cxgbei_limits, idl);
1355 
1356 	return (0);
1357 }
1358 
1359 int
1360 icl_cxgbei_mod_load(void)
1361 {
1362 	int rc;
1363 
1364 	refcount_init(&icl_cxgbei_ncons, 0);
1365 
1366 	rc = icl_register("cxgbei", false, -100, icl_cxgbei_limits,
1367 	    icl_cxgbei_new_conn);
1368 
1369 	return (rc);
1370 }
1371 
1372 int
1373 icl_cxgbei_mod_unload(void)
1374 {
1375 
1376 	if (icl_cxgbei_ncons != 0)
1377 		return (EBUSY);
1378 
1379 	icl_unregister("cxgbei", false);
1380 
1381 	return (0);
1382 }
1383 #endif
1384