xref: /freebsd/sys/dev/cxgbe/cxgbei/icl_cxgbei.c (revision 9b1bb0aee697352b39b3efa1843f581ca29068ba)
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_regs.h"
99 #include "common/t4_tcb.h"
100 #include "tom/t4_tom.h"
101 #include "cxgbei.h"
102 
103 /*
104  * Use the page pod tag for the TT hash.
105  */
106 #define	TT_HASH(icc, tt)	(G_PPOD_TAG(tt) & (icc)->cmp_hash_mask)
107 
108 struct cxgbei_ddp_state {
109 	struct ppod_reservation prsv;
110 	struct cxgbei_cmp cmp;
111 };
112 
113 static MALLOC_DEFINE(M_CXGBEI, "cxgbei", "cxgbei(4)");
114 
115 SYSCTL_NODE(_kern_icl, OID_AUTO, cxgbei, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
116     "Chelsio iSCSI offload");
117 static int first_burst_length = 8192;
118 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, first_burst_length, CTLFLAG_RWTUN,
119     &first_burst_length, 0, "First burst length");
120 static int max_burst_length = 2 * 1024 * 1024;
121 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, max_burst_length, CTLFLAG_RWTUN,
122     &max_burst_length, 0, "Maximum burst length");
123 static int sendspace = 1048576;
124 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, sendspace, CTLFLAG_RWTUN,
125     &sendspace, 0, "Default send socket buffer size");
126 static int recvspace = 1048576;
127 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, recvspace, CTLFLAG_RWTUN,
128     &recvspace, 0, "Default receive socket buffer size");
129 
130 static volatile u_int icl_cxgbei_ncons;
131 
132 #define ICL_CONN_LOCK(X)		mtx_lock(X->ic_lock)
133 #define ICL_CONN_UNLOCK(X)		mtx_unlock(X->ic_lock)
134 #define ICL_CONN_LOCK_ASSERT(X)		mtx_assert(X->ic_lock, MA_OWNED)
135 #define ICL_CONN_LOCK_ASSERT_NOT(X)	mtx_assert(X->ic_lock, MA_NOTOWNED)
136 
137 static icl_conn_new_pdu_t	icl_cxgbei_conn_new_pdu;
138 static icl_conn_pdu_data_segment_length_t
139 				    icl_cxgbei_conn_pdu_data_segment_length;
140 static icl_conn_pdu_append_data_t	icl_cxgbei_conn_pdu_append_data;
141 static icl_conn_pdu_get_data_t	icl_cxgbei_conn_pdu_get_data;
142 static icl_conn_pdu_queue_t	icl_cxgbei_conn_pdu_queue;
143 static icl_conn_pdu_queue_cb_t	icl_cxgbei_conn_pdu_queue_cb;
144 static icl_conn_handoff_t	icl_cxgbei_conn_handoff;
145 static icl_conn_free_t		icl_cxgbei_conn_free;
146 static icl_conn_close_t		icl_cxgbei_conn_close;
147 static icl_conn_task_setup_t	icl_cxgbei_conn_task_setup;
148 static icl_conn_task_done_t	icl_cxgbei_conn_task_done;
149 static icl_conn_transfer_setup_t	icl_cxgbei_conn_transfer_setup;
150 static icl_conn_transfer_done_t	icl_cxgbei_conn_transfer_done;
151 
152 static kobj_method_t icl_cxgbei_methods[] = {
153 	KOBJMETHOD(icl_conn_new_pdu, icl_cxgbei_conn_new_pdu),
154 	KOBJMETHOD(icl_conn_pdu_free, icl_cxgbei_conn_pdu_free),
155 	KOBJMETHOD(icl_conn_pdu_data_segment_length,
156 	    icl_cxgbei_conn_pdu_data_segment_length),
157 	KOBJMETHOD(icl_conn_pdu_append_data, icl_cxgbei_conn_pdu_append_data),
158 	KOBJMETHOD(icl_conn_pdu_get_data, icl_cxgbei_conn_pdu_get_data),
159 	KOBJMETHOD(icl_conn_pdu_queue, icl_cxgbei_conn_pdu_queue),
160 	KOBJMETHOD(icl_conn_pdu_queue_cb, icl_cxgbei_conn_pdu_queue_cb),
161 	KOBJMETHOD(icl_conn_handoff, icl_cxgbei_conn_handoff),
162 	KOBJMETHOD(icl_conn_free, icl_cxgbei_conn_free),
163 	KOBJMETHOD(icl_conn_close, icl_cxgbei_conn_close),
164 	KOBJMETHOD(icl_conn_task_setup, icl_cxgbei_conn_task_setup),
165 	KOBJMETHOD(icl_conn_task_done, icl_cxgbei_conn_task_done),
166 	KOBJMETHOD(icl_conn_transfer_setup, icl_cxgbei_conn_transfer_setup),
167 	KOBJMETHOD(icl_conn_transfer_done, icl_cxgbei_conn_transfer_done),
168 	{ 0, 0 }
169 };
170 
171 DEFINE_CLASS(icl_cxgbei, icl_cxgbei_methods, sizeof(struct icl_cxgbei_conn));
172 
173 void
174 icl_cxgbei_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip)
175 {
176 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
177 
178 	KASSERT(icp->ref_cnt != 0, ("freeing deleted PDU"));
179 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
180 	MPASS(ic == ip->ip_conn);
181 
182 	m_freem(ip->ip_ahs_mbuf);
183 	m_freem(ip->ip_data_mbuf);
184 	m_freem(ip->ip_bhs_mbuf);
185 
186 	KASSERT(ic != NULL || icp->ref_cnt == 1,
187 	    ("orphaned PDU has oustanding references"));
188 
189 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) != 1)
190 		return;
191 
192 	free(icp, M_CXGBEI);
193 #ifdef DIAGNOSTIC
194 	if (__predict_true(ic != NULL))
195 		refcount_release(&ic->ic_outstanding_pdus);
196 #endif
197 }
198 
199 static void
200 icl_cxgbei_pdu_call_cb(struct icl_pdu *ip)
201 {
202 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
203 
204 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
205 
206 	if (icp->cb != NULL)
207 		icp->cb(ip, icp->error);
208 #ifdef DIAGNOSTIC
209 	if (__predict_true(ip->ip_conn != NULL))
210 		refcount_release(&ip->ip_conn->ic_outstanding_pdus);
211 #endif
212 	free(icp, M_CXGBEI);
213 }
214 
215 static void
216 icl_cxgbei_pdu_done(struct icl_pdu *ip, int error)
217 {
218 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
219 
220 	if (error != 0)
221 		icp->error = error;
222 
223 	m_freem(ip->ip_ahs_mbuf);
224 	ip->ip_ahs_mbuf = NULL;
225 	m_freem(ip->ip_data_mbuf);
226 	ip->ip_data_mbuf = NULL;
227 	m_freem(ip->ip_bhs_mbuf);
228 	ip->ip_bhs_mbuf = NULL;
229 
230 	/*
231 	 * All other references to this PDU should have been dropped
232 	 * by the m_freem() of ip_data_mbuf.
233 	 */
234 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1)
235 		icl_cxgbei_pdu_call_cb(ip);
236 	else
237 		__assert_unreachable();
238 }
239 
240 static void
241 icl_cxgbei_mbuf_done(struct mbuf *mb)
242 {
243 
244 	struct icl_cxgbei_pdu *icp = (struct icl_cxgbei_pdu *)mb->m_ext.ext_arg1;
245 
246 	/*
247 	 * NB: mb_free_mext() might leave ref_cnt as 1 without
248 	 * decrementing it if it hits the fast path in the ref_cnt
249 	 * check.
250 	 */
251 	icl_cxgbei_pdu_call_cb(&icp->ip);
252 }
253 
254 struct icl_pdu *
255 icl_cxgbei_new_pdu(int flags)
256 {
257 	struct icl_cxgbei_pdu *icp;
258 	struct icl_pdu *ip;
259 	struct mbuf *m;
260 
261 	icp = malloc(sizeof(*icp), M_CXGBEI, flags | M_ZERO);
262 	if (__predict_false(icp == NULL))
263 		return (NULL);
264 
265 	icp->icp_signature = CXGBEI_PDU_SIGNATURE;
266 	icp->ref_cnt = 1;
267 	ip = &icp->ip;
268 
269 	m = m_gethdr(flags, MT_DATA);
270 	if (__predict_false(m == NULL)) {
271 		free(icp, M_CXGBEI);
272 		return (NULL);
273 	}
274 
275 	ip->ip_bhs_mbuf = m;
276 	ip->ip_bhs = mtod(m, struct iscsi_bhs *);
277 	memset(ip->ip_bhs, 0, sizeof(*ip->ip_bhs));
278 	m->m_len = sizeof(struct iscsi_bhs);
279 	m->m_pkthdr.len = m->m_len;
280 
281 	return (ip);
282 }
283 
284 void
285 icl_cxgbei_new_pdu_set_conn(struct icl_pdu *ip, struct icl_conn *ic)
286 {
287 
288 	ip->ip_conn = ic;
289 #ifdef DIAGNOSTIC
290 	refcount_acquire(&ic->ic_outstanding_pdus);
291 #endif
292 }
293 
294 /*
295  * Allocate icl_pdu with empty BHS to fill up by the caller.
296  */
297 static struct icl_pdu *
298 icl_cxgbei_conn_new_pdu(struct icl_conn *ic, int flags)
299 {
300 	struct icl_pdu *ip;
301 
302 	ip = icl_cxgbei_new_pdu(flags);
303 	if (__predict_false(ip == NULL))
304 		return (NULL);
305 	icl_cxgbei_new_pdu_set_conn(ip, ic);
306 
307 	return (ip);
308 }
309 
310 static size_t
311 icl_pdu_data_segment_length(const struct icl_pdu *request)
312 {
313 	uint32_t len = 0;
314 
315 	len += request->ip_bhs->bhs_data_segment_len[0];
316 	len <<= 8;
317 	len += request->ip_bhs->bhs_data_segment_len[1];
318 	len <<= 8;
319 	len += request->ip_bhs->bhs_data_segment_len[2];
320 
321 	return (len);
322 }
323 
324 size_t
325 icl_cxgbei_conn_pdu_data_segment_length(struct icl_conn *ic,
326     const struct icl_pdu *request)
327 {
328 
329 	return (icl_pdu_data_segment_length(request));
330 }
331 
332 static struct mbuf *
333 finalize_pdu(struct icl_cxgbei_conn *icc, struct icl_cxgbei_pdu *icp)
334 {
335 	struct icl_pdu *ip = &icp->ip;
336 	uint8_t ulp_submode, padding;
337 	struct mbuf *m, *last;
338 	struct iscsi_bhs *bhs;
339 	int data_len;
340 
341 	/*
342 	 * Fix up the data segment mbuf first.
343 	 */
344 	m = ip->ip_data_mbuf;
345 	ulp_submode = icc->ulp_submode;
346 	if (m != NULL) {
347 		last = m_last(m);
348 
349 		/*
350 		 * Round up the data segment to a 4B boundary.	Pad with 0 if
351 		 * necessary.  There will definitely be room in the mbuf.
352 		 */
353 		padding = roundup2(ip->ip_data_len, 4) - ip->ip_data_len;
354 		if (padding != 0) {
355 			MPASS(padding <= M_TRAILINGSPACE(last));
356 			bzero(mtod(last, uint8_t *) + last->m_len, padding);
357 			last->m_len += padding;
358 		}
359 	} else {
360 		MPASS(ip->ip_data_len == 0);
361 		ulp_submode &= ~ULP_CRC_DATA;
362 		padding = 0;
363 	}
364 
365 	/*
366 	 * Now the header mbuf that has the BHS.
367 	 */
368 	m = ip->ip_bhs_mbuf;
369 	MPASS(m->m_pkthdr.len == sizeof(struct iscsi_bhs));
370 	MPASS(m->m_len == sizeof(struct iscsi_bhs));
371 
372 	bhs = ip->ip_bhs;
373 	data_len = ip->ip_data_len;
374 	if (data_len > icc->ic.ic_max_send_data_segment_length) {
375 		struct iscsi_bhs_data_in *bhsdi;
376 		int flags;
377 
378 		KASSERT(padding == 0, ("%s: ISO with padding %d for icp %p",
379 		    __func__, padding, icp));
380 		switch (bhs->bhs_opcode) {
381 		case ISCSI_BHS_OPCODE_SCSI_DATA_OUT:
382 			flags = 1;
383 			break;
384 		case ISCSI_BHS_OPCODE_SCSI_DATA_IN:
385 			flags = 2;
386 			break;
387 		default:
388 			panic("invalid opcode %#x for ISO", bhs->bhs_opcode);
389 		}
390 		data_len = icc->ic.ic_max_send_data_segment_length;
391 		bhsdi = (struct iscsi_bhs_data_in *)bhs;
392 		if (bhsdi->bhsdi_flags & BHSDI_FLAGS_F) {
393 			/*
394 			 * Firmware will set F on the final PDU in the
395 			 * burst.
396 			 */
397 			flags |= CXGBE_ISO_F;
398 			bhsdi->bhsdi_flags &= ~BHSDI_FLAGS_F;
399 		}
400 		set_mbuf_iscsi_iso(m, true);
401 		set_mbuf_iscsi_iso_flags(m, flags);
402 		set_mbuf_iscsi_iso_mss(m, data_len);
403 	}
404 
405 	bhs->bhs_data_segment_len[2] = data_len;
406 	bhs->bhs_data_segment_len[1] = data_len >> 8;
407 	bhs->bhs_data_segment_len[0] = data_len >> 16;
408 
409 	/*
410 	 * Extract mbuf chain from PDU.
411 	 */
412 	m->m_pkthdr.len += ip->ip_data_len + padding;
413 	m->m_next = ip->ip_data_mbuf;
414 	set_mbuf_ulp_submode(m, ulp_submode);
415 	ip->ip_bhs_mbuf = NULL;
416 	ip->ip_data_mbuf = NULL;
417 	ip->ip_bhs = NULL;
418 
419 	/*
420 	 * Drop PDU reference on icp.  Additional references might
421 	 * still be held by zero-copy PDU buffers (ICL_NOCOPY).
422 	 */
423 	if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1)
424 		icl_cxgbei_pdu_call_cb(ip);
425 
426 	return (m);
427 }
428 
429 int
430 icl_cxgbei_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *ip,
431     const void *addr, size_t len, int flags)
432 {
433 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
434 	struct mbuf *m, *m_tail;
435 	const char *src;
436 
437 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
438 	MPASS(ic == ip->ip_conn);
439 	KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len));
440 
441 	m_tail = ip->ip_data_mbuf;
442 	if (m_tail != NULL)
443 		for (; m_tail->m_next != NULL; m_tail = m_tail->m_next)
444 			;
445 
446 	if (flags & ICL_NOCOPY) {
447 		m = m_get(flags & ~ICL_NOCOPY, MT_DATA);
448 		if (m == NULL) {
449 			ICL_WARN("failed to allocate mbuf");
450 			return (ENOMEM);
451 		}
452 
453 		m->m_flags |= M_RDONLY;
454 		m_extaddref(m, __DECONST(char *, addr), len, &icp->ref_cnt,
455 		    icl_cxgbei_mbuf_done, icp, NULL);
456 		m->m_len = len;
457 		if (ip->ip_data_mbuf == NULL) {
458 			ip->ip_data_mbuf = m;
459 			ip->ip_data_len = len;
460 		} else {
461 			m_tail->m_next = m;
462 			m_tail = m_tail->m_next;
463 			ip->ip_data_len += len;
464 		}
465 
466 		return (0);
467 	}
468 
469 	src = (const char *)addr;
470 
471 	/* Allocate as jumbo mbufs of size MJUM16BYTES. */
472 	while (len >= MJUM16BYTES) {
473 		m = m_getjcl(M_NOWAIT, MT_DATA, 0, MJUM16BYTES);
474 		if (__predict_false(m == NULL)) {
475 			if ((flags & M_WAITOK) != 0) {
476 				/* Fall back to non-jumbo mbufs. */
477 				break;
478 			}
479 			return (ENOMEM);
480 		}
481 		memcpy(mtod(m, void *), src, MJUM16BYTES);
482 		m->m_len = MJUM16BYTES;
483 		if (ip->ip_data_mbuf == NULL) {
484 			ip->ip_data_mbuf = m_tail = m;
485 			ip->ip_data_len = MJUM16BYTES;
486 		} else {
487 			m_tail->m_next = m;
488 			m_tail = m_tail->m_next;
489 			ip->ip_data_len += MJUM16BYTES;
490 		}
491 		src += MJUM16BYTES;
492 		len -= MJUM16BYTES;
493 	}
494 
495 	/* Allocate mbuf chain for the remaining data. */
496 	if (len != 0) {
497 		m = m_getm2(NULL, len, flags, MT_DATA, 0);
498 		if (__predict_false(m == NULL))
499 			return (ENOMEM);
500 		if (ip->ip_data_mbuf == NULL) {
501 			ip->ip_data_mbuf = m;
502 			ip->ip_data_len = len;
503 		} else {
504 			m_tail->m_next = m;
505 			ip->ip_data_len += len;
506 		}
507 		for (; m != NULL; m = m->m_next) {
508 			m->m_len = min(len, M_SIZE(m));
509 			memcpy(mtod(m, void *), src, m->m_len);
510 			src += m->m_len;
511 			len -= m->m_len;
512 		}
513 		MPASS(len == 0);
514 	}
515 	MPASS(ip->ip_data_len <= max(ic->ic_max_send_data_segment_length,
516 	    ic->ic_hw_isomax));
517 
518 	return (0);
519 }
520 
521 void
522 icl_cxgbei_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip,
523     size_t off, void *addr, size_t len)
524 {
525 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
526 
527 	if (icp->icp_flags & ICPF_RX_DDP)
528 		return; /* data is DDP'ed, no need to copy */
529 	m_copydata(ip->ip_data_mbuf, off, len, addr);
530 }
531 
532 void
533 icl_cxgbei_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip)
534 {
535 	icl_cxgbei_conn_pdu_queue_cb(ic, ip, NULL);
536 }
537 
538 void
539 icl_cxgbei_conn_pdu_queue_cb(struct icl_conn *ic, struct icl_pdu *ip,
540 			     icl_pdu_cb cb)
541 {
542 	struct epoch_tracker et;
543 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
544 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
545 	struct socket *so = ic->ic_socket;
546 	struct toepcb *toep = icc->toep;
547 	struct inpcb *inp;
548 	struct mbuf *m;
549 
550 	MPASS(ic == ip->ip_conn);
551 	MPASS(ip->ip_bhs_mbuf != NULL);
552 	/* The kernel doesn't generate PDUs with AHS. */
553 	MPASS(ip->ip_ahs_mbuf == NULL && ip->ip_ahs_len == 0);
554 
555 	ICL_CONN_LOCK_ASSERT(ic);
556 
557 	icp->cb = cb;
558 
559 	/* NOTE: sowriteable without so_snd lock is a mostly harmless race. */
560 	if (ic->ic_disconnecting || so == NULL || !sowriteable(so)) {
561 		icl_cxgbei_pdu_done(ip, ENOTCONN);
562 		return;
563 	}
564 
565 	m = finalize_pdu(icc, icp);
566 	M_ASSERTPKTHDR(m);
567 	MPASS((m->m_pkthdr.len & 3) == 0);
568 
569 	/*
570 	 * Do not get inp from toep->inp as the toepcb might have detached
571 	 * already.
572 	 */
573 	inp = sotoinpcb(so);
574 	CURVNET_SET(toep->vnet);
575 	NET_EPOCH_ENTER(et);
576 	INP_WLOCK(inp);
577 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) ||
578 	    __predict_false((toep->flags & TPF_ATTACHED) == 0))
579 		m_freem(m);
580 	else {
581 		mbufq_enqueue(&toep->ulp_pduq, m);
582 		t4_push_pdus(icc->sc, toep, 0);
583 	}
584 	INP_WUNLOCK(inp);
585 	NET_EPOCH_EXIT(et);
586 	CURVNET_RESTORE();
587 }
588 
589 static struct icl_conn *
590 icl_cxgbei_new_conn(const char *name, struct mtx *lock)
591 {
592 	struct icl_cxgbei_conn *icc;
593 	struct icl_conn *ic;
594 
595 	refcount_acquire(&icl_cxgbei_ncons);
596 
597 	icc = (struct icl_cxgbei_conn *)kobj_create(&icl_cxgbei_class, M_CXGBE,
598 	    M_WAITOK | M_ZERO);
599 	icc->icc_signature = CXGBEI_CONN_SIGNATURE;
600 	STAILQ_INIT(&icc->rcvd_pdus);
601 
602 	icc->cmp_table = hashinit(64, M_CXGBEI, &icc->cmp_hash_mask);
603 	mtx_init(&icc->cmp_lock, "cxgbei_cmp", NULL, MTX_DEF);
604 
605 	ic = &icc->ic;
606 	ic->ic_lock = lock;
607 
608 #ifdef DIAGNOSTIC
609 	refcount_init(&ic->ic_outstanding_pdus, 0);
610 #endif
611 	ic->ic_name = name;
612 	ic->ic_offload = "cxgbei";
613 	ic->ic_unmapped = false;
614 
615 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
616 
617 	return (ic);
618 }
619 
620 void
621 icl_cxgbei_conn_free(struct icl_conn *ic)
622 {
623 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
624 
625 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
626 
627 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
628 
629 	mtx_destroy(&icc->cmp_lock);
630 	hashdestroy(icc->cmp_table, M_CXGBEI, icc->cmp_hash_mask);
631 	kobj_delete((struct kobj *)icc, M_CXGBE);
632 	refcount_release(&icl_cxgbei_ncons);
633 }
634 
635 static int
636 icl_cxgbei_setsockopt(struct icl_conn *ic, struct socket *so, int sspace,
637     int rspace)
638 {
639 	struct sockopt opt;
640 	int error, one = 1, ss, rs;
641 
642 	ss = max(sendspace, sspace);
643 	rs = max(recvspace, rspace);
644 
645 	error = soreserve(so, ss, rs);
646 	if (error != 0) {
647 		icl_cxgbei_conn_close(ic);
648 		return (error);
649 	}
650 	SOCKBUF_LOCK(&so->so_snd);
651 	so->so_snd.sb_flags |= SB_AUTOSIZE;
652 	SOCKBUF_UNLOCK(&so->so_snd);
653 	SOCKBUF_LOCK(&so->so_rcv);
654 	so->so_rcv.sb_flags |= SB_AUTOSIZE;
655 	SOCKBUF_UNLOCK(&so->so_rcv);
656 
657 	/*
658 	 * Disable Nagle.
659 	 */
660 	bzero(&opt, sizeof(opt));
661 	opt.sopt_dir = SOPT_SET;
662 	opt.sopt_level = IPPROTO_TCP;
663 	opt.sopt_name = TCP_NODELAY;
664 	opt.sopt_val = &one;
665 	opt.sopt_valsize = sizeof(one);
666 	error = sosetopt(so, &opt);
667 	if (error != 0) {
668 		icl_cxgbei_conn_close(ic);
669 		return (error);
670 	}
671 
672 	return (0);
673 }
674 
675 /*
676  * Request/response structure used to find out the adapter offloading a socket.
677  */
678 struct find_ofld_adapter_rr {
679 	struct socket *so;
680 	struct adapter *sc;	/* result */
681 };
682 
683 static void
684 find_offload_adapter(struct adapter *sc, void *arg)
685 {
686 	struct find_ofld_adapter_rr *fa = arg;
687 	struct socket *so = fa->so;
688 	struct tom_data *td = sc->tom_softc;
689 	struct tcpcb *tp;
690 	struct inpcb *inp;
691 
692 	/* Non-TCP were filtered out earlier. */
693 	MPASS(so->so_proto->pr_protocol == IPPROTO_TCP);
694 
695 	if (fa->sc != NULL)
696 		return;	/* Found already. */
697 
698 	if (td == NULL)
699 		return;	/* TOE not enabled on this adapter. */
700 
701 	inp = sotoinpcb(so);
702 	INP_WLOCK(inp);
703 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
704 		tp = intotcpcb(inp);
705 		if (tp->t_flags & TF_TOE && tp->tod == &td->tod)
706 			fa->sc = sc;	/* Found. */
707 	}
708 	INP_WUNLOCK(inp);
709 }
710 
711 static bool
712 is_memfree(struct adapter *sc)
713 {
714 	uint32_t em;
715 
716 	em = t4_read_reg(sc, A_MA_TARGET_MEM_ENABLE);
717 	if ((em & F_EXT_MEM_ENABLE) != 0)
718 		return (false);
719 	if (is_t5(sc) && (em & F_EXT_MEM1_ENABLE) != 0)
720 		return (false);
721 	return (true);
722 }
723 
724 /* XXXNP: move this to t4_tom. */
725 static void
726 send_iscsi_flowc_wr(struct adapter *sc, struct toepcb *toep, int maxlen)
727 {
728 	struct wrqe *wr;
729 	struct fw_flowc_wr *flowc;
730 	const u_int nparams = 1;
731 	u_int flowclen;
732 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
733 
734 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
735 
736 	wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq);
737 	if (wr == NULL) {
738 		/* XXX */
739 		panic("%s: allocation failure.", __func__);
740 	}
741 	flowc = wrtod(wr);
742 	memset(flowc, 0, wr->wr_len);
743 
744 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
745 	    V_FW_FLOWC_WR_NPARAMS(nparams));
746 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
747 	    V_FW_WR_FLOWID(toep->tid));
748 
749 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_TXDATAPLEN_MAX;
750 	flowc->mnemval[0].val = htobe32(maxlen);
751 
752 	txsd->tx_credits = howmany(flowclen, 16);
753 	txsd->plen = 0;
754 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
755 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
756 	toep->tx_credits -= txsd->tx_credits;
757 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
758 		toep->txsd_pidx = 0;
759 	toep->txsd_avail--;
760 
761 	t4_wrq_tx(sc, wr);
762 }
763 
764 static void
765 set_ulp_mode_iscsi(struct adapter *sc, struct toepcb *toep, u_int ulp_submode)
766 {
767 	uint64_t val;
768 
769 	CTR3(KTR_CXGBE, "%s: tid %u, ULP_MODE_ISCSI, submode=%#x",
770 	    __func__, toep->tid, ulp_submode);
771 
772 	val = V_TCB_ULP_TYPE(ULP_MODE_ISCSI) | V_TCB_ULP_RAW(ulp_submode);
773 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_ULP_TYPE,
774 	    V_TCB_ULP_TYPE(M_TCB_ULP_TYPE) | V_TCB_ULP_RAW(M_TCB_ULP_RAW), val,
775 	    0, 0);
776 
777 	val = V_TF_RX_FLOW_CONTROL_DISABLE(1ULL);
778 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, val, val, 0, 0);
779 }
780 
781 /*
782  * XXXNP: Who is responsible for cleaning up the socket if this returns with an
783  * error?  Review all error paths.
784  *
785  * XXXNP: What happens to the socket's fd reference if the operation is
786  * successful, and how does that affect the socket's life cycle?
787  */
788 int
789 icl_cxgbei_conn_handoff(struct icl_conn *ic, int fd)
790 {
791 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
792 	struct cxgbei_data *ci;
793 	struct find_ofld_adapter_rr fa;
794 	struct file *fp;
795 	struct socket *so;
796 	struct inpcb *inp;
797 	struct tcpcb *tp;
798 	struct toepcb *toep;
799 	cap_rights_t rights;
800 	u_int max_rx_pdu_len, max_tx_pdu_len;
801 	int error, max_iso_pdus;
802 
803 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
804 	ICL_CONN_LOCK_ASSERT_NOT(ic);
805 
806 	/*
807 	 * Steal the socket from userland.
808 	 */
809 	error = fget(curthread, fd,
810 	    cap_rights_init_one(&rights, CAP_SOCK_CLIENT), &fp);
811 	if (error != 0)
812 		return (error);
813 	if (fp->f_type != DTYPE_SOCKET) {
814 		fdrop(fp, curthread);
815 		return (EINVAL);
816 	}
817 	so = fp->f_data;
818 	if (so->so_type != SOCK_STREAM ||
819 	    so->so_proto->pr_protocol != IPPROTO_TCP) {
820 		fdrop(fp, curthread);
821 		return (EINVAL);
822 	}
823 
824 	ICL_CONN_LOCK(ic);
825 	if (ic->ic_socket != NULL) {
826 		ICL_CONN_UNLOCK(ic);
827 		fdrop(fp, curthread);
828 		return (EBUSY);
829 	}
830 	ic->ic_disconnecting = false;
831 	ic->ic_socket = so;
832 	fp->f_ops = &badfileops;
833 	fp->f_data = NULL;
834 	fdrop(fp, curthread);
835 	ICL_CONN_UNLOCK(ic);
836 
837 	/* Find the adapter offloading this socket. */
838 	fa.sc = NULL;
839 	fa.so = so;
840 	t4_iterate(find_offload_adapter, &fa);
841 	if (fa.sc == NULL)
842 		return (EINVAL);
843 	icc->sc = fa.sc;
844 	ci = icc->sc->iscsi_ulp_softc;
845 
846 	max_rx_pdu_len = ISCSI_BHS_SIZE + ic->ic_max_recv_data_segment_length;
847 	max_tx_pdu_len = ISCSI_BHS_SIZE + ic->ic_max_send_data_segment_length;
848 	if (ic->ic_header_crc32c) {
849 		max_rx_pdu_len += ISCSI_HEADER_DIGEST_SIZE;
850 		max_tx_pdu_len += ISCSI_HEADER_DIGEST_SIZE;
851 	}
852 	if (ic->ic_data_crc32c) {
853 		max_rx_pdu_len += ISCSI_DATA_DIGEST_SIZE;
854 		max_tx_pdu_len += ISCSI_DATA_DIGEST_SIZE;
855 	}
856 
857 	inp = sotoinpcb(so);
858 	INP_WLOCK(inp);
859 	tp = intotcpcb(inp);
860 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))
861 		error = EBUSY;
862 	else {
863 		/*
864 		 * socket could not have been "unoffloaded" if here.
865 		 */
866 		MPASS(tp->t_flags & TF_TOE);
867 		MPASS(tp->tod != NULL);
868 		MPASS(tp->t_toe != NULL);
869 		toep = tp->t_toe;
870 		MPASS(toep->vi->adapter == icc->sc);
871 		icc->toep = toep;
872 		icc->cwt = cxgbei_select_worker_thread(icc);
873 
874 		icc->ulp_submode = 0;
875 		if (ic->ic_header_crc32c)
876 			icc->ulp_submode |= ULP_CRC_HEADER;
877 		if (ic->ic_data_crc32c)
878 			icc->ulp_submode |= ULP_CRC_DATA;
879 
880 		if (icc->sc->tt.iso && chip_id(icc->sc) >= CHELSIO_T5 &&
881 		    !is_memfree(icc->sc)) {
882 			max_iso_pdus = CXGBEI_MAX_ISO_PAYLOAD /
883 			    max_tx_pdu_len;
884 			ic->ic_hw_isomax = max_iso_pdus *
885 			    ic->ic_max_send_data_segment_length;
886 		} else
887 			max_iso_pdus = 1;
888 
889 		so->so_options |= SO_NO_DDP;
890 		toep->params.ulp_mode = ULP_MODE_ISCSI;
891 		toep->ulpcb = icc;
892 
893 		send_iscsi_flowc_wr(icc->sc, toep,
894 		    roundup(max_iso_pdus * max_tx_pdu_len, tp->t_maxseg));
895 		set_ulp_mode_iscsi(icc->sc, toep, icc->ulp_submode);
896 		error = 0;
897 	}
898 	INP_WUNLOCK(inp);
899 
900 	if (error == 0) {
901 		error = icl_cxgbei_setsockopt(ic, so, max_tx_pdu_len,
902 		    max_rx_pdu_len);
903 	}
904 
905 	return (error);
906 }
907 
908 void
909 icl_cxgbei_conn_close(struct icl_conn *ic)
910 {
911 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
912 	struct icl_pdu *ip;
913 	struct socket *so;
914 	struct sockbuf *sb;
915 	struct inpcb *inp;
916 	struct toepcb *toep = icc->toep;
917 
918 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
919 	ICL_CONN_LOCK_ASSERT_NOT(ic);
920 
921 	ICL_CONN_LOCK(ic);
922 	so = ic->ic_socket;
923 	if (ic->ic_disconnecting || so == NULL) {
924 		CTR4(KTR_CXGBE, "%s: icc %p (disconnecting = %d), so %p",
925 		    __func__, icc, ic->ic_disconnecting, so);
926 		ICL_CONN_UNLOCK(ic);
927 		return;
928 	}
929 	ic->ic_disconnecting = true;
930 
931 #ifdef DIAGNOSTIC
932 	KASSERT(ic->ic_outstanding_pdus == 0,
933 	    ("destroying session with %d outstanding PDUs",
934 	     ic->ic_outstanding_pdus));
935 #endif
936 	ICL_CONN_UNLOCK(ic);
937 
938 	CTR3(KTR_CXGBE, "%s: tid %d, icc %p", __func__, toep ? toep->tid : -1,
939 	    icc);
940 	inp = sotoinpcb(so);
941 	sb = &so->so_rcv;
942 	INP_WLOCK(inp);
943 	if (toep != NULL) {	/* NULL if connection was never offloaded. */
944 		toep->ulpcb = NULL;
945 
946 		/* Discard PDUs queued for TX. */
947 		mbufq_drain(&toep->ulp_pduq);
948 
949 		/*
950 		 * Wait for the cwt threads to stop processing this
951 		 * connection.
952 		 */
953 		SOCKBUF_LOCK(sb);
954 		if (icc->rx_flags & RXF_ACTIVE) {
955 			volatile u_int *p = &icc->rx_flags;
956 
957 			SOCKBUF_UNLOCK(sb);
958 			INP_WUNLOCK(inp);
959 
960 			while (*p & RXF_ACTIVE)
961 				pause("conclo", 1);
962 
963 			INP_WLOCK(inp);
964 			SOCKBUF_LOCK(sb);
965 		}
966 
967 		/*
968 		 * Discard received PDUs not passed to the iSCSI
969 		 * layer.
970 		 */
971 		while (!STAILQ_EMPTY(&icc->rcvd_pdus)) {
972 			ip = STAILQ_FIRST(&icc->rcvd_pdus);
973 			STAILQ_REMOVE_HEAD(&icc->rcvd_pdus, ip_next);
974 			icl_cxgbei_pdu_done(ip, ENOTCONN);
975 		}
976 		SOCKBUF_UNLOCK(sb);
977 
978 		/*
979 		 * Grab a reference to use when waiting for the final
980 		 * CPL to be received.  If toep->inp is NULL, then
981 		 * final_cpl_received() has already been called (e.g.
982 		 * due to the peer sending a RST).
983 		 */
984 		if (toep->inp != NULL) {
985 			toep = hold_toepcb(toep);
986 			toep->flags |= TPF_WAITING_FOR_FINAL;
987 		} else
988 			toep = NULL;
989 	}
990 	INP_WUNLOCK(inp);
991 
992 	ICL_CONN_LOCK(ic);
993 	ic->ic_socket = NULL;
994 	ICL_CONN_UNLOCK(ic);
995 
996 	/*
997 	 * XXXNP: we should send RST instead of FIN when PDUs held in various
998 	 * queues were purged instead of delivered reliably but soabort isn't
999 	 * really general purpose and wouldn't do the right thing here.
1000 	 */
1001 	soclose(so);
1002 
1003 	/*
1004 	 * Wait for the socket to fully close.  This ensures any
1005 	 * pending received data has been received (and in particular,
1006 	 * any data that would be received by DDP has been handled).
1007 	 * Callers assume that it is safe to free buffers for tasks
1008 	 * and transfers after this function returns.
1009 	 */
1010 	if (toep != NULL) {
1011 		struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep);
1012 
1013 		mtx_lock(lock);
1014 		while ((toep->flags & TPF_WAITING_FOR_FINAL) != 0)
1015 			mtx_sleep(toep, lock, PSOCK, "conclo2", 0);
1016 		mtx_unlock(lock);
1017 		free_toepcb(toep);
1018 	}
1019 }
1020 
1021 static void
1022 cxgbei_insert_cmp(struct icl_cxgbei_conn *icc, struct cxgbei_cmp *cmp,
1023     uint32_t tt)
1024 {
1025 #ifdef INVARIANTS
1026 	struct cxgbei_cmp *cmp2;
1027 #endif
1028 
1029 	cmp->tt = tt;
1030 
1031 	mtx_lock(&icc->cmp_lock);
1032 #ifdef INVARIANTS
1033 	LIST_FOREACH(cmp2, &icc->cmp_table[TT_HASH(icc, tt)], link) {
1034 		KASSERT(cmp2->tt != tt, ("%s: duplicate cmp", __func__));
1035 	}
1036 #endif
1037 	LIST_INSERT_HEAD(&icc->cmp_table[TT_HASH(icc, tt)], cmp, link);
1038 	mtx_unlock(&icc->cmp_lock);
1039 }
1040 
1041 struct cxgbei_cmp *
1042 cxgbei_find_cmp(struct icl_cxgbei_conn *icc, uint32_t tt)
1043 {
1044 	struct cxgbei_cmp *cmp;
1045 
1046 	mtx_lock(&icc->cmp_lock);
1047 	LIST_FOREACH(cmp, &icc->cmp_table[TT_HASH(icc, tt)], link) {
1048 		if (cmp->tt == tt)
1049 			break;
1050 	}
1051 	mtx_unlock(&icc->cmp_lock);
1052 	return (cmp);
1053 }
1054 
1055 static void
1056 cxgbei_rm_cmp(struct icl_cxgbei_conn *icc, struct cxgbei_cmp *cmp)
1057 {
1058 #ifdef INVARIANTS
1059 	struct cxgbei_cmp *cmp2;
1060 #endif
1061 
1062 	mtx_lock(&icc->cmp_lock);
1063 
1064 #ifdef INVARIANTS
1065 	LIST_FOREACH(cmp2, &icc->cmp_table[TT_HASH(icc, cmp->tt)], link) {
1066 		if (cmp2 == cmp)
1067 			goto found;
1068 	}
1069 	panic("%s: could not find cmp", __func__);
1070 found:
1071 #endif
1072 	LIST_REMOVE(cmp, link);
1073 	mtx_unlock(&icc->cmp_lock);
1074 }
1075 
1076 int
1077 icl_cxgbei_conn_task_setup(struct icl_conn *ic, struct icl_pdu *ip,
1078     struct ccb_scsiio *csio, uint32_t *ittp, void **arg)
1079 {
1080 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
1081 	struct toepcb *toep = icc->toep;
1082 	struct adapter *sc = icc->sc;
1083 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
1084 	struct ppod_region *pr = &ci->pr;
1085 	struct cxgbei_ddp_state *ddp;
1086 	struct ppod_reservation *prsv;
1087 	struct inpcb *inp;
1088 	struct mbufq mq;
1089 	uint32_t itt;
1090 	int rc = 0;
1091 
1092 	ICL_CONN_LOCK_ASSERT(ic);
1093 
1094 	/* This is for the offload driver's state.  Must not be set already. */
1095 	MPASS(arg != NULL);
1096 	MPASS(*arg == NULL);
1097 
1098 	if (ic->ic_disconnecting || ic->ic_socket == NULL)
1099 		return (ECONNRESET);
1100 
1101 	if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_IN ||
1102 	    csio->dxfer_len < ci->ddp_threshold) {
1103 no_ddp:
1104 		/*
1105 		 * No DDP for this I/O.	 Allocate an ITT (based on the one
1106 		 * passed in) that cannot be a valid hardware DDP tag in the
1107 		 * iSCSI region.
1108 		 */
1109 		itt = *ittp & M_PPOD_TAG;
1110 		itt = V_PPOD_TAG(itt) | pr->pr_invalid_bit;
1111 		*ittp = htobe32(itt);
1112 		MPASS(*arg == NULL);	/* State is maintained for DDP only. */
1113 		if (rc != 0)
1114 			counter_u64_add(
1115 			    toep->ofld_rxq->rx_iscsi_ddp_setup_error, 1);
1116 		return (0);
1117 	}
1118 
1119 	/*
1120 	 * Reserve resources for DDP, update the itt that should be used in the
1121 	 * PDU, and save DDP specific state for this I/O in *arg.
1122 	 */
1123 	ddp = malloc(sizeof(*ddp), M_CXGBEI, M_NOWAIT | M_ZERO);
1124 	if (ddp == NULL) {
1125 		rc = ENOMEM;
1126 		goto no_ddp;
1127 	}
1128 	prsv = &ddp->prsv;
1129 
1130 	/* XXX add support for all CAM_DATA_ types */
1131 	MPASS((csio->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_VADDR);
1132 	rc = t4_alloc_page_pods_for_buf(pr, (vm_offset_t)csio->data_ptr,
1133 	    csio->dxfer_len, prsv);
1134 	if (rc != 0) {
1135 		free(ddp, M_CXGBEI);
1136 		goto no_ddp;
1137 	}
1138 
1139 	mbufq_init(&mq, INT_MAX);
1140 	rc = t4_write_page_pods_for_buf(sc, toep, prsv,
1141 	    (vm_offset_t)csio->data_ptr, csio->dxfer_len, &mq);
1142 	if (__predict_false(rc != 0)) {
1143 		mbufq_drain(&mq);
1144 		t4_free_page_pods(prsv);
1145 		free(ddp, M_CXGBEI);
1146 		goto no_ddp;
1147 	}
1148 
1149 	/*
1150 	 * Do not get inp from toep->inp as the toepcb might have
1151 	 * detached already.
1152 	 */
1153 	inp = sotoinpcb(ic->ic_socket);
1154 	INP_WLOCK(inp);
1155 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) != 0) {
1156 		INP_WUNLOCK(inp);
1157 		mbufq_drain(&mq);
1158 		t4_free_page_pods(prsv);
1159 		free(ddp, M_CXGBEI);
1160 		return (ECONNRESET);
1161 	}
1162 	mbufq_concat(&toep->ulp_pduq, &mq);
1163 	INP_WUNLOCK(inp);
1164 
1165 	ddp->cmp.last_datasn = -1;
1166 	cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag);
1167 	*ittp = htobe32(prsv->prsv_tag);
1168 	*arg = prsv;
1169 	counter_u64_add(toep->ofld_rxq->rx_iscsi_ddp_setup_ok, 1);
1170 	return (0);
1171 }
1172 
1173 void
1174 icl_cxgbei_conn_task_done(struct icl_conn *ic, void *arg)
1175 {
1176 
1177 	if (arg != NULL) {
1178 		struct cxgbei_ddp_state *ddp = arg;
1179 
1180 		cxgbei_rm_cmp(ic_to_icc(ic), &ddp->cmp);
1181 		t4_free_page_pods(&ddp->prsv);
1182 		free(ddp, M_CXGBEI);
1183 	}
1184 }
1185 
1186 static inline bool
1187 ddp_sgl_check(struct ctl_sg_entry *sg, int entries, int xferlen)
1188 {
1189 	int total_len = 0;
1190 
1191 	MPASS(entries > 0);
1192 	if (((vm_offset_t)sg[--entries].addr & 3U) != 0)
1193 		return (false);
1194 
1195 	total_len += sg[entries].len;
1196 
1197 	while (--entries >= 0) {
1198 		if (((vm_offset_t)sg[entries].addr & PAGE_MASK) != 0 ||
1199 		    (sg[entries].len % PAGE_SIZE) != 0)
1200 			return (false);
1201 		total_len += sg[entries].len;
1202 	}
1203 
1204 	MPASS(total_len == xferlen);
1205 	return (true);
1206 }
1207 
1208 /* XXXNP: PDU should be passed in as parameter, like on the initiator. */
1209 #define io_to_request_pdu(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND].ptr)
1210 #define io_to_ddp_state(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND2].ptr)
1211 
1212 int
1213 icl_cxgbei_conn_transfer_setup(struct icl_conn *ic, union ctl_io *io,
1214     uint32_t *tttp, void **arg)
1215 {
1216 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
1217 	struct toepcb *toep = icc->toep;
1218 	struct ctl_scsiio *ctsio = &io->scsiio;
1219 	struct adapter *sc = icc->sc;
1220 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
1221 	struct ppod_region *pr = &ci->pr;
1222 	struct cxgbei_ddp_state *ddp;
1223 	struct ppod_reservation *prsv;
1224 	struct ctl_sg_entry *sgl, sg_entry;
1225 	struct inpcb *inp;
1226 	struct mbufq mq;
1227 	int sg_entries = ctsio->kern_sg_entries;
1228 	uint32_t ttt;
1229 	int xferlen, rc = 0, alias;
1230 
1231 	/* This is for the offload driver's state.  Must not be set already. */
1232 	MPASS(arg != NULL);
1233 	MPASS(*arg == NULL);
1234 
1235 	if (ctsio->ext_data_filled == 0) {
1236 		int first_burst;
1237 		struct icl_pdu *ip = io_to_request_pdu(io);
1238 #ifdef INVARIANTS
1239 		struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
1240 
1241 		MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
1242 		MPASS(ic == ip->ip_conn);
1243 		MPASS(ip->ip_bhs_mbuf != NULL);
1244 #endif
1245 		first_burst = icl_pdu_data_segment_length(ip);
1246 
1247 		/*
1248 		 * Note that ICL calls conn_transfer_setup even if the first
1249 		 * burst had everything and there's nothing left to transfer.
1250 		 *
1251 		 * NB: The CTL frontend might have provided a buffer
1252 		 * whose length (kern_data_len) is smaller than the
1253 		 * FirstBurstLength of unsolicited data.  Treat those
1254 		 * as an empty transfer.
1255 		 */
1256 		xferlen = ctsio->kern_data_len;
1257 		if (xferlen < first_burst ||
1258 		    xferlen - first_burst < ci->ddp_threshold) {
1259 no_ddp:
1260 			/*
1261 			 * No DDP for this transfer.  Allocate a TTT (based on
1262 			 * the one passed in) that cannot be a valid hardware
1263 			 * DDP tag in the iSCSI region.
1264 			 */
1265 			ttt = *tttp & M_PPOD_TAG;
1266 			ttt = V_PPOD_TAG(ttt) | pr->pr_invalid_bit;
1267 			*tttp = htobe32(ttt);
1268 			MPASS(io_to_ddp_state(io) == NULL);
1269 			if (rc != 0)
1270 				counter_u64_add(
1271 				    toep->ofld_rxq->rx_iscsi_ddp_setup_error, 1);
1272 			return (0);
1273 		}
1274 
1275 		if (sg_entries == 0) {
1276 			sgl = &sg_entry;
1277 			sgl->len = xferlen;
1278 			sgl->addr = (void *)ctsio->kern_data_ptr;
1279 			sg_entries = 1;
1280 		} else
1281 			sgl = (void *)ctsio->kern_data_ptr;
1282 
1283 		if (!ddp_sgl_check(sgl, sg_entries, xferlen))
1284 			goto no_ddp;
1285 
1286 		/*
1287 		 * Reserve resources for DDP, update the ttt that should be used
1288 		 * in the PDU, and save DDP specific state for this I/O.
1289 		 */
1290 		MPASS(io_to_ddp_state(io) == NULL);
1291 		ddp = malloc(sizeof(*ddp), M_CXGBEI, M_NOWAIT | M_ZERO);
1292 		if (ddp == NULL) {
1293 			rc = ENOMEM;
1294 			goto no_ddp;
1295 		}
1296 		prsv = &ddp->prsv;
1297 
1298 		rc = t4_alloc_page_pods_for_sgl(pr, sgl, sg_entries, prsv);
1299 		if (rc != 0) {
1300 			free(ddp, M_CXGBEI);
1301 			goto no_ddp;
1302 		}
1303 
1304 		mbufq_init(&mq, INT_MAX);
1305 		rc = t4_write_page_pods_for_sgl(sc, toep, prsv, sgl, sg_entries,
1306 		    xferlen, &mq);
1307 		if (__predict_false(rc != 0)) {
1308 			mbufq_drain(&mq);
1309 			t4_free_page_pods(prsv);
1310 			free(ddp, M_CXGBEI);
1311 			goto no_ddp;
1312 		}
1313 
1314 		/*
1315 		 * Do not get inp from toep->inp as the toepcb might
1316 		 * have detached already.
1317 		 */
1318 		ICL_CONN_LOCK(ic);
1319 		if (ic->ic_disconnecting || ic->ic_socket == NULL) {
1320 			ICL_CONN_UNLOCK(ic);
1321 			mbufq_drain(&mq);
1322 			t4_free_page_pods(prsv);
1323 			free(ddp, M_CXGBEI);
1324 			return (ECONNRESET);
1325 		}
1326 		inp = sotoinpcb(ic->ic_socket);
1327 		INP_WLOCK(inp);
1328 		ICL_CONN_UNLOCK(ic);
1329 		if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) != 0) {
1330 			INP_WUNLOCK(inp);
1331 			mbufq_drain(&mq);
1332 			t4_free_page_pods(prsv);
1333 			free(ddp, M_CXGBEI);
1334 			return (ECONNRESET);
1335 		}
1336 		mbufq_concat(&toep->ulp_pduq, &mq);
1337 		INP_WUNLOCK(inp);
1338 
1339 		ddp->cmp.next_buffer_offset = ctsio->kern_rel_offset +
1340 		    first_burst;
1341 		ddp->cmp.last_datasn = -1;
1342 		cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag);
1343 		*tttp = htobe32(prsv->prsv_tag);
1344 		io_to_ddp_state(io) = ddp;
1345 		*arg = ctsio;
1346 		counter_u64_add(toep->ofld_rxq->rx_iscsi_ddp_setup_ok, 1);
1347 		return (0);
1348 	}
1349 
1350 	/*
1351 	 * In the middle of an I/O.  A non-NULL page pod reservation indicates
1352 	 * that a DDP buffer is being used for the I/O.
1353 	 */
1354 	ddp = io_to_ddp_state(ctsio);
1355 	if (ddp == NULL)
1356 		goto no_ddp;
1357 	prsv = &ddp->prsv;
1358 
1359 	alias = (prsv->prsv_tag & pr->pr_alias_mask) >> pr->pr_alias_shift;
1360 	alias++;
1361 	prsv->prsv_tag &= ~pr->pr_alias_mask;
1362 	prsv->prsv_tag |= alias << pr->pr_alias_shift & pr->pr_alias_mask;
1363 
1364 	ddp->cmp.last_datasn = -1;
1365 	cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag);
1366 	*tttp = htobe32(prsv->prsv_tag);
1367 	*arg = ctsio;
1368 
1369 	return (0);
1370 }
1371 
1372 void
1373 icl_cxgbei_conn_transfer_done(struct icl_conn *ic, void *arg)
1374 {
1375 	struct ctl_scsiio *ctsio = arg;
1376 
1377 	if (ctsio != NULL) {
1378 		struct cxgbei_ddp_state *ddp;
1379 
1380 		ddp = io_to_ddp_state(ctsio);
1381 		MPASS(ddp != NULL);
1382 
1383 		cxgbei_rm_cmp(ic_to_icc(ic), &ddp->cmp);
1384 		if (ctsio->kern_data_len == ctsio->ext_data_filled ||
1385 		    ic->ic_disconnecting) {
1386 			t4_free_page_pods(&ddp->prsv);
1387 			free(ddp, M_CXGBEI);
1388 			io_to_ddp_state(ctsio) = NULL;
1389 		}
1390 	}
1391 }
1392 
1393 static void
1394 cxgbei_limits(struct adapter *sc, void *arg)
1395 {
1396 	struct icl_drv_limits *idl = arg;
1397 	struct cxgbei_data *ci;
1398 	int max_dsl;
1399 
1400 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4lims") != 0)
1401 		return;
1402 
1403 	if (uld_active(sc, ULD_ISCSI)) {
1404 		ci = sc->iscsi_ulp_softc;
1405 		MPASS(ci != NULL);
1406 
1407 
1408 		max_dsl = ci->max_rx_data_len;
1409 		if (idl->idl_max_recv_data_segment_length > max_dsl)
1410 			idl->idl_max_recv_data_segment_length = max_dsl;
1411 
1412 		max_dsl = ci->max_tx_data_len;
1413 		if (idl->idl_max_send_data_segment_length > max_dsl)
1414 			idl->idl_max_send_data_segment_length = max_dsl;
1415 	}
1416 
1417 	end_synchronized_op(sc, LOCK_HELD);
1418 }
1419 
1420 static int
1421 icl_cxgbei_limits(struct icl_drv_limits *idl)
1422 {
1423 
1424 	/* Maximum allowed by the RFC.	cxgbei_limits will clip them. */
1425 	idl->idl_max_recv_data_segment_length = (1 << 24) - 1;
1426 	idl->idl_max_send_data_segment_length = (1 << 24) - 1;
1427 
1428 	/* These are somewhat arbitrary. */
1429 	idl->idl_max_burst_length = max_burst_length;
1430 	idl->idl_first_burst_length = first_burst_length;
1431 
1432 	t4_iterate(cxgbei_limits, idl);
1433 
1434 	return (0);
1435 }
1436 
1437 int
1438 icl_cxgbei_mod_load(void)
1439 {
1440 	int rc;
1441 
1442 	refcount_init(&icl_cxgbei_ncons, 0);
1443 
1444 	rc = icl_register("cxgbei", false, -100, icl_cxgbei_limits,
1445 	    icl_cxgbei_new_conn);
1446 
1447 	return (rc);
1448 }
1449 
1450 int
1451 icl_cxgbei_mod_unload(void)
1452 {
1453 
1454 	if (icl_cxgbei_ncons != 0)
1455 		return (EBUSY);
1456 
1457 	icl_unregister("cxgbei", false);
1458 
1459 	return (0);
1460 }
1461 #endif
1462