xref: /freebsd/sys/dev/cxgbe/cxgbei/icl_cxgbei.c (revision 4e99f45480598189d49d45a825533a6c9e12f02c)
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/uma.h>
64 #include <vm/vm.h>
65 #include <vm/pmap.h>
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp.h>
69 #include <netinet/tcp_var.h>
70 #include <netinet/toecore.h>
71 
72 #include <dev/iscsi/icl.h>
73 #include <dev/iscsi/iscsi_proto.h>
74 #include <icl_conn_if.h>
75 
76 #include <cam/scsi/scsi_all.h>
77 #include <cam/scsi/scsi_da.h>
78 #include <cam/ctl/ctl_io.h>
79 #include <cam/ctl/ctl.h>
80 #include <cam/ctl/ctl_backend.h>
81 #include <cam/ctl/ctl_error.h>
82 #include <cam/ctl/ctl_frontend.h>
83 #include <cam/ctl/ctl_debug.h>
84 #include <cam/ctl/ctl_ha.h>
85 #include <cam/ctl/ctl_ioctl.h>
86 
87 #include <cam/cam.h>
88 #include <cam/cam_ccb.h>
89 #include <cam/cam_xpt.h>
90 #include <cam/cam_debug.h>
91 #include <cam/cam_sim.h>
92 #include <cam/cam_xpt_sim.h>
93 #include <cam/cam_xpt_periph.h>
94 #include <cam/cam_periph.h>
95 #include <cam/cam_compat.h>
96 #include <cam/scsi/scsi_message.h>
97 
98 #include "common/common.h"
99 #include "common/t4_tcb.h"
100 #include "tom/t4_tom.h"
101 #include "cxgbei.h"
102 
103 SYSCTL_NODE(_kern_icl, OID_AUTO, cxgbei, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
104     "Chelsio iSCSI offload");
105 static int coalesce = 1;
106 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, coalesce, CTLFLAG_RWTUN,
107 	&coalesce, 0, "Try to coalesce PDUs before sending");
108 static int partial_receive_len = 128 * 1024;
109 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, partial_receive_len, CTLFLAG_RWTUN,
110     &partial_receive_len, 0, "Minimum read size for partially received "
111     "data segment");
112 static int sendspace = 1048576;
113 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, sendspace, CTLFLAG_RWTUN,
114     &sendspace, 0, "Default send socket buffer size");
115 static int recvspace = 1048576;
116 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, recvspace, CTLFLAG_RWTUN,
117     &recvspace, 0, "Default receive socket buffer size");
118 
119 static uma_zone_t prsv_zone;
120 static volatile u_int icl_cxgbei_ncons;
121 
122 #define ICL_CONN_LOCK(X)		mtx_lock(X->ic_lock)
123 #define ICL_CONN_UNLOCK(X)		mtx_unlock(X->ic_lock)
124 #define ICL_CONN_LOCK_ASSERT(X)		mtx_assert(X->ic_lock, MA_OWNED)
125 #define ICL_CONN_LOCK_ASSERT_NOT(X)	mtx_assert(X->ic_lock, MA_NOTOWNED)
126 
127 struct icl_pdu *icl_cxgbei_new_pdu(int);
128 void icl_cxgbei_new_pdu_set_conn(struct icl_pdu *, struct icl_conn *);
129 
130 static icl_conn_new_pdu_t	icl_cxgbei_conn_new_pdu;
131 icl_conn_pdu_free_t	icl_cxgbei_conn_pdu_free;
132 static icl_conn_pdu_data_segment_length_t
133 				    icl_cxgbei_conn_pdu_data_segment_length;
134 static icl_conn_pdu_append_data_t	icl_cxgbei_conn_pdu_append_data;
135 static icl_conn_pdu_get_data_t	icl_cxgbei_conn_pdu_get_data;
136 static icl_conn_pdu_queue_t	icl_cxgbei_conn_pdu_queue;
137 static icl_conn_handoff_t	icl_cxgbei_conn_handoff;
138 static icl_conn_free_t		icl_cxgbei_conn_free;
139 static icl_conn_close_t		icl_cxgbei_conn_close;
140 static icl_conn_task_setup_t	icl_cxgbei_conn_task_setup;
141 static icl_conn_task_done_t	icl_cxgbei_conn_task_done;
142 static icl_conn_transfer_setup_t	icl_cxgbei_conn_transfer_setup;
143 static icl_conn_transfer_done_t	icl_cxgbei_conn_transfer_done;
144 
145 static kobj_method_t icl_cxgbei_methods[] = {
146 	KOBJMETHOD(icl_conn_new_pdu, icl_cxgbei_conn_new_pdu),
147 	KOBJMETHOD(icl_conn_pdu_free, icl_cxgbei_conn_pdu_free),
148 	KOBJMETHOD(icl_conn_pdu_data_segment_length,
149 	    icl_cxgbei_conn_pdu_data_segment_length),
150 	KOBJMETHOD(icl_conn_pdu_append_data, icl_cxgbei_conn_pdu_append_data),
151 	KOBJMETHOD(icl_conn_pdu_get_data, icl_cxgbei_conn_pdu_get_data),
152 	KOBJMETHOD(icl_conn_pdu_queue, icl_cxgbei_conn_pdu_queue),
153 	KOBJMETHOD(icl_conn_handoff, icl_cxgbei_conn_handoff),
154 	KOBJMETHOD(icl_conn_free, icl_cxgbei_conn_free),
155 	KOBJMETHOD(icl_conn_close, icl_cxgbei_conn_close),
156 	KOBJMETHOD(icl_conn_task_setup, icl_cxgbei_conn_task_setup),
157 	KOBJMETHOD(icl_conn_task_done, icl_cxgbei_conn_task_done),
158 	KOBJMETHOD(icl_conn_transfer_setup, icl_cxgbei_conn_transfer_setup),
159 	KOBJMETHOD(icl_conn_transfer_done, icl_cxgbei_conn_transfer_done),
160 	{ 0, 0 }
161 };
162 
163 DEFINE_CLASS(icl_cxgbei, icl_cxgbei_methods, sizeof(struct icl_cxgbei_conn));
164 
165 void
166 icl_cxgbei_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip)
167 {
168 #ifdef INVARIANTS
169 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
170 #endif
171 
172 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
173 	MPASS(ic == ip->ip_conn);
174 	MPASS(ip->ip_bhs_mbuf != NULL);
175 
176 	m_freem(ip->ip_ahs_mbuf);
177 	m_freem(ip->ip_data_mbuf);
178 	m_freem(ip->ip_bhs_mbuf);	/* storage for icl_cxgbei_pdu itself */
179 
180 #ifdef DIAGNOSTIC
181 	if (__predict_true(ic != NULL))
182 		refcount_release(&ic->ic_outstanding_pdus);
183 #endif
184 }
185 
186 struct icl_pdu *
187 icl_cxgbei_new_pdu(int flags)
188 {
189 	struct icl_cxgbei_pdu *icp;
190 	struct icl_pdu *ip;
191 	struct mbuf *m;
192 	uintptr_t a;
193 
194 	m = m_gethdr(flags, MT_DATA);
195 	if (__predict_false(m == NULL))
196 		return (NULL);
197 
198 	a = roundup2(mtod(m, uintptr_t), _Alignof(struct icl_cxgbei_pdu));
199 	icp = (struct icl_cxgbei_pdu *)a;
200 	bzero(icp, sizeof(*icp));
201 
202 	icp->icp_signature = CXGBEI_PDU_SIGNATURE;
203 	ip = &icp->ip;
204 	ip->ip_bhs_mbuf = m;
205 
206 	a = roundup2((uintptr_t)(icp + 1), _Alignof(struct iscsi_bhs *));
207 	ip->ip_bhs = (struct iscsi_bhs *)a;
208 #ifdef INVARIANTS
209 	/* Everything must fit entirely in the mbuf. */
210 	a = (uintptr_t)(ip->ip_bhs + 1);
211 	MPASS(a <= (uintptr_t)m + MSIZE);
212 #endif
213 	bzero(ip->ip_bhs, sizeof(*ip->ip_bhs));
214 
215 	m->m_data = (void *)ip->ip_bhs;
216 	m->m_len = sizeof(struct iscsi_bhs);
217 	m->m_pkthdr.len = m->m_len;
218 
219 	return (ip);
220 }
221 
222 void
223 icl_cxgbei_new_pdu_set_conn(struct icl_pdu *ip, struct icl_conn *ic)
224 {
225 
226 	ip->ip_conn = ic;
227 #ifdef DIAGNOSTIC
228 	refcount_acquire(&ic->ic_outstanding_pdus);
229 #endif
230 }
231 
232 /*
233  * Allocate icl_pdu with empty BHS to fill up by the caller.
234  */
235 static struct icl_pdu *
236 icl_cxgbei_conn_new_pdu(struct icl_conn *ic, int flags)
237 {
238 	struct icl_pdu *ip;
239 
240 	ip = icl_cxgbei_new_pdu(flags);
241 	if (__predict_false(ip == NULL))
242 		return (NULL);
243 	icl_cxgbei_new_pdu_set_conn(ip, ic);
244 
245 	return (ip);
246 }
247 
248 static size_t
249 icl_pdu_data_segment_length(const struct icl_pdu *request)
250 {
251 	uint32_t len = 0;
252 
253 	len += request->ip_bhs->bhs_data_segment_len[0];
254 	len <<= 8;
255 	len += request->ip_bhs->bhs_data_segment_len[1];
256 	len <<= 8;
257 	len += request->ip_bhs->bhs_data_segment_len[2];
258 
259 	return (len);
260 }
261 
262 size_t
263 icl_cxgbei_conn_pdu_data_segment_length(struct icl_conn *ic,
264     const struct icl_pdu *request)
265 {
266 
267 	return (icl_pdu_data_segment_length(request));
268 }
269 
270 static struct mbuf *
271 finalize_pdu(struct icl_cxgbei_conn *icc, struct icl_cxgbei_pdu *icp)
272 {
273 	struct icl_pdu *ip = &icp->ip;
274 	uint8_t ulp_submode, padding;
275 	struct mbuf *m, *last;
276 	struct iscsi_bhs *bhs;
277 
278 	/*
279 	 * Fix up the data segment mbuf first.
280 	 */
281 	m = ip->ip_data_mbuf;
282 	ulp_submode = icc->ulp_submode;
283 	if (m) {
284 		last = m_last(m);
285 
286 		/*
287 		 * Round up the data segment to a 4B boundary.  Pad with 0 if
288 		 * necessary.  There will definitely be room in the mbuf.
289 		 */
290 		padding = roundup2(ip->ip_data_len, 4) - ip->ip_data_len;
291 		if (padding) {
292 			bzero(mtod(last, uint8_t *) + last->m_len, padding);
293 			last->m_len += padding;
294 		}
295 	} else {
296 		MPASS(ip->ip_data_len == 0);
297 		ulp_submode &= ~ULP_CRC_DATA;
298 		padding = 0;
299 	}
300 
301 	/*
302 	 * Now the header mbuf that has the BHS.
303 	 */
304 	m = ip->ip_bhs_mbuf;
305 	MPASS(m->m_pkthdr.len == sizeof(struct iscsi_bhs));
306 	MPASS(m->m_len == sizeof(struct iscsi_bhs));
307 
308 	bhs = ip->ip_bhs;
309 	bhs->bhs_data_segment_len[2] = ip->ip_data_len;
310 	bhs->bhs_data_segment_len[1] = ip->ip_data_len >> 8;
311 	bhs->bhs_data_segment_len[0] = ip->ip_data_len >> 16;
312 
313 	/* "Convert" PDU to mbuf chain.  Do not use icp/ip after this. */
314 	m->m_pkthdr.len = sizeof(struct iscsi_bhs) + ip->ip_data_len + padding;
315 	m->m_next = ip->ip_data_mbuf;
316 	set_mbuf_ulp_submode(m, ulp_submode);
317 #ifdef INVARIANTS
318 	bzero(icp, sizeof(*icp));
319 #endif
320 #ifdef DIAGNOSTIC
321 	refcount_release(&icc->ic.ic_outstanding_pdus);
322 #endif
323 
324 	return (m);
325 }
326 
327 int
328 icl_cxgbei_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *ip,
329     const void *addr, size_t len, int flags)
330 {
331 	struct mbuf *m;
332 #ifdef INVARIANTS
333 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
334 #endif
335 
336 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
337 	MPASS(ic == ip->ip_conn);
338 	KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len));
339 
340 	m = ip->ip_data_mbuf;
341 	if (m == NULL) {
342 		m = m_getjcl(M_NOWAIT, MT_DATA, 0, MJUM16BYTES);
343 		if (__predict_false(m == NULL))
344 			return (ENOMEM);
345 
346 		ip->ip_data_mbuf = m;
347 	}
348 
349 	if (__predict_true(m_append(m, len, addr) != 0)) {
350 		ip->ip_data_len += len;
351 		MPASS(ip->ip_data_len <= ic->ic_max_data_segment_length);
352 		return (0);
353 	} else {
354 	    	if (flags & M_WAITOK) {
355 			CXGBE_UNIMPLEMENTED("fail safe append");
356 		}
357 		ip->ip_data_len = m_length(m, NULL);
358 		return (1);
359 	}
360 }
361 
362 void
363 icl_cxgbei_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip,
364     size_t off, void *addr, size_t len)
365 {
366 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
367 
368 	if (icp->icp_flags & ICPF_RX_DDP)
369 		return; /* data is DDP'ed, no need to copy */
370 	m_copydata(ip->ip_data_mbuf, off, len, addr);
371 }
372 
373 void
374 icl_cxgbei_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip)
375 {
376 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
377 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
378 	struct socket *so = ic->ic_socket;
379 	struct toepcb *toep = icc->toep;
380 	struct inpcb *inp;
381 	struct mbuf *m;
382 
383 	MPASS(ic == ip->ip_conn);
384 	MPASS(ip->ip_bhs_mbuf != NULL);
385 	/* The kernel doesn't generate PDUs with AHS. */
386 	MPASS(ip->ip_ahs_mbuf == NULL && ip->ip_ahs_len == 0);
387 
388 	ICL_CONN_LOCK_ASSERT(ic);
389 	/* NOTE: sowriteable without so_snd lock is a mostly harmless race. */
390 	if (ic->ic_disconnecting || so == NULL || !sowriteable(so)) {
391 		icl_cxgbei_conn_pdu_free(ic, ip);
392 		return;
393 	}
394 
395 	m = finalize_pdu(icc, icp);
396 	M_ASSERTPKTHDR(m);
397 	MPASS((m->m_pkthdr.len & 3) == 0);
398 
399 	/*
400 	 * Do not get inp from toep->inp as the toepcb might have detached
401 	 * already.
402 	 */
403 	inp = sotoinpcb(so);
404 	INP_WLOCK(inp);
405 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) ||
406 	    __predict_false((toep->flags & TPF_ATTACHED) == 0))
407 		m_freem(m);
408 	else {
409 		mbufq_enqueue(&toep->ulp_pduq, m);
410 		t4_push_pdus(icc->sc, toep, 0);
411 	}
412 	INP_WUNLOCK(inp);
413 }
414 
415 static struct icl_conn *
416 icl_cxgbei_new_conn(const char *name, struct mtx *lock)
417 {
418 	struct icl_cxgbei_conn *icc;
419 	struct icl_conn *ic;
420 
421 	refcount_acquire(&icl_cxgbei_ncons);
422 
423 	icc = (struct icl_cxgbei_conn *)kobj_create(&icl_cxgbei_class, M_CXGBE,
424 	    M_WAITOK | M_ZERO);
425 	icc->icc_signature = CXGBEI_CONN_SIGNATURE;
426 	STAILQ_INIT(&icc->rcvd_pdus);
427 
428 	ic = &icc->ic;
429 	ic->ic_lock = lock;
430 
431 	/* XXXNP: review.  Most of these icl_conn fields aren't really used */
432 	STAILQ_INIT(&ic->ic_to_send);
433 	cv_init(&ic->ic_send_cv, "icl_cxgbei_tx");
434 	cv_init(&ic->ic_receive_cv, "icl_cxgbei_rx");
435 #ifdef DIAGNOSTIC
436 	refcount_init(&ic->ic_outstanding_pdus, 0);
437 #endif
438 	/* This is a stop-gap value that will be corrected during handoff. */
439 	ic->ic_max_data_segment_length = 16384;
440 	ic->ic_name = name;
441 	ic->ic_offload = "cxgbei";
442 	ic->ic_unmapped = false;
443 
444 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
445 
446 	return (ic);
447 }
448 
449 void
450 icl_cxgbei_conn_free(struct icl_conn *ic)
451 {
452 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
453 
454 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
455 
456 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
457 
458 	cv_destroy(&ic->ic_send_cv);
459 	cv_destroy(&ic->ic_receive_cv);
460 
461 	kobj_delete((struct kobj *)icc, M_CXGBE);
462 	refcount_release(&icl_cxgbei_ncons);
463 }
464 
465 static int
466 icl_cxgbei_setsockopt(struct icl_conn *ic, struct socket *so, int sspace,
467     int rspace)
468 {
469 	struct sockopt opt;
470 	int error, one = 1, ss, rs;
471 
472 	ss = max(sendspace, sspace);
473 	rs = max(recvspace, rspace);
474 
475 	error = soreserve(so, ss, rs);
476 	if (error != 0) {
477 		icl_cxgbei_conn_close(ic);
478 		return (error);
479 	}
480 	SOCKBUF_LOCK(&so->so_snd);
481 	so->so_snd.sb_flags |= SB_AUTOSIZE;
482 	SOCKBUF_UNLOCK(&so->so_snd);
483 	SOCKBUF_LOCK(&so->so_rcv);
484 	so->so_rcv.sb_flags |= SB_AUTOSIZE;
485 	SOCKBUF_UNLOCK(&so->so_rcv);
486 
487 	/*
488 	 * Disable Nagle.
489 	 */
490 	bzero(&opt, sizeof(opt));
491 	opt.sopt_dir = SOPT_SET;
492 	opt.sopt_level = IPPROTO_TCP;
493 	opt.sopt_name = TCP_NODELAY;
494 	opt.sopt_val = &one;
495 	opt.sopt_valsize = sizeof(one);
496 	error = sosetopt(so, &opt);
497 	if (error != 0) {
498 		icl_cxgbei_conn_close(ic);
499 		return (error);
500 	}
501 
502 	return (0);
503 }
504 
505 /*
506  * Request/response structure used to find out the adapter offloading a socket.
507  */
508 struct find_ofld_adapter_rr {
509 	struct socket *so;
510 	struct adapter *sc;	/* result */
511 };
512 
513 static void
514 find_offload_adapter(struct adapter *sc, void *arg)
515 {
516 	struct find_ofld_adapter_rr *fa = arg;
517 	struct socket *so = fa->so;
518 	struct tom_data *td = sc->tom_softc;
519 	struct tcpcb *tp;
520 	struct inpcb *inp;
521 
522 	/* Non-TCP were filtered out earlier. */
523 	MPASS(so->so_proto->pr_protocol == IPPROTO_TCP);
524 
525 	if (fa->sc != NULL)
526 		return;	/* Found already. */
527 
528 	if (td == NULL)
529 		return;	/* TOE not enabled on this adapter. */
530 
531 	inp = sotoinpcb(so);
532 	INP_WLOCK(inp);
533 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
534 		tp = intotcpcb(inp);
535 		if (tp->t_flags & TF_TOE && tp->tod == &td->tod)
536 			fa->sc = sc;	/* Found. */
537 	}
538 	INP_WUNLOCK(inp);
539 }
540 
541 /* XXXNP: move this to t4_tom. */
542 static void
543 send_iscsi_flowc_wr(struct adapter *sc, struct toepcb *toep, int maxlen)
544 {
545 	struct wrqe *wr;
546 	struct fw_flowc_wr *flowc;
547 	const u_int nparams = 1;
548 	u_int flowclen;
549 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
550 
551 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
552 
553 	wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
554 	if (wr == NULL) {
555 		/* XXX */
556 		panic("%s: allocation failure.", __func__);
557 	}
558 	flowc = wrtod(wr);
559 	memset(flowc, 0, wr->wr_len);
560 
561 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
562 	    V_FW_FLOWC_WR_NPARAMS(nparams));
563 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
564 	    V_FW_WR_FLOWID(toep->tid));
565 
566 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_TXDATAPLEN_MAX;
567 	flowc->mnemval[0].val = htobe32(maxlen);
568 
569 	txsd->tx_credits = howmany(flowclen, 16);
570 	txsd->plen = 0;
571 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
572 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
573 	toep->tx_credits -= txsd->tx_credits;
574 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
575 		toep->txsd_pidx = 0;
576 	toep->txsd_avail--;
577 
578         t4_wrq_tx(sc, wr);
579 }
580 
581 static void
582 set_ulp_mode_iscsi(struct adapter *sc, struct toepcb *toep, int hcrc, int dcrc)
583 {
584 	uint64_t val = ULP_MODE_ISCSI;
585 
586 	if (hcrc)
587 		val |= ULP_CRC_HEADER << 4;
588 	if (dcrc)
589 		val |= ULP_CRC_DATA << 4;
590 
591 	CTR4(KTR_CXGBE, "%s: tid %u, ULP_MODE_ISCSI, CRC hdr=%d data=%d",
592 	    __func__, toep->tid, hcrc, dcrc);
593 
594 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_ULP_TYPE,
595 	    V_TCB_ULP_TYPE(M_TCB_ULP_TYPE) | V_TCB_ULP_RAW(M_TCB_ULP_RAW), val,
596 	    0, 0);
597 }
598 
599 /*
600  * XXXNP: Who is responsible for cleaning up the socket if this returns with an
601  * error?  Review all error paths.
602  *
603  * XXXNP: What happens to the socket's fd reference if the operation is
604  * successful, and how does that affect the socket's life cycle?
605  */
606 int
607 icl_cxgbei_conn_handoff(struct icl_conn *ic, int fd)
608 {
609 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
610 	struct cxgbei_data *ci;
611 	struct find_ofld_adapter_rr fa;
612 	struct file *fp;
613 	struct socket *so;
614 	struct inpcb *inp;
615 	struct tcpcb *tp;
616 	struct toepcb *toep;
617 	cap_rights_t rights;
618 	int error;
619 
620 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
621 	ICL_CONN_LOCK_ASSERT_NOT(ic);
622 
623 	/*
624 	 * Steal the socket from userland.
625 	 */
626 	error = fget(curthread, fd,
627 	    cap_rights_init(&rights, CAP_SOCK_CLIENT), &fp);
628 	if (error != 0)
629 		return (error);
630 	if (fp->f_type != DTYPE_SOCKET) {
631 		fdrop(fp, curthread);
632 		return (EINVAL);
633 	}
634 	so = fp->f_data;
635 	if (so->so_type != SOCK_STREAM ||
636 	    so->so_proto->pr_protocol != IPPROTO_TCP) {
637 		fdrop(fp, curthread);
638 		return (EINVAL);
639 	}
640 
641 	ICL_CONN_LOCK(ic);
642 	if (ic->ic_socket != NULL) {
643 		ICL_CONN_UNLOCK(ic);
644 		fdrop(fp, curthread);
645 		return (EBUSY);
646 	}
647 	ic->ic_disconnecting = false;
648 	ic->ic_socket = so;
649 	fp->f_ops = &badfileops;
650 	fp->f_data = NULL;
651 	fdrop(fp, curthread);
652 	ICL_CONN_UNLOCK(ic);
653 
654 	/* Find the adapter offloading this socket. */
655 	fa.sc = NULL;
656 	fa.so = so;
657 	t4_iterate(find_offload_adapter, &fa);
658 	if (fa.sc == NULL)
659 		return (EINVAL);
660 	icc->sc = fa.sc;
661 	ci = icc->sc->iscsi_ulp_softc;
662 
663 	inp = sotoinpcb(so);
664 	INP_WLOCK(inp);
665 	tp = intotcpcb(inp);
666 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))
667 		error = EBUSY;
668 	else {
669 		/*
670 		 * socket could not have been "unoffloaded" if here.
671 		 */
672 		MPASS(tp->t_flags & TF_TOE);
673 		MPASS(tp->tod != NULL);
674 		MPASS(tp->t_toe != NULL);
675 		toep = tp->t_toe;
676 		MPASS(toep->vi->pi->adapter == icc->sc);
677 		icc->toep = toep;
678 		icc->cwt = cxgbei_select_worker_thread(icc);
679 
680 		/*
681 		 * We maintain the _send_ DSL in this field just to have a
682 		 * convenient way to assert that the kernel never sends
683 		 * oversized PDUs.  This field is otherwise unused in the driver
684 		 * or the kernel.
685 		 */
686 		ic->ic_max_data_segment_length = ci->max_tx_pdu_len -
687 		    ISCSI_BHS_SIZE;
688 
689 		icc->ulp_submode = 0;
690 		if (ic->ic_header_crc32c) {
691 			icc->ulp_submode |= ULP_CRC_HEADER;
692 			ic->ic_max_data_segment_length -=
693 			    ISCSI_HEADER_DIGEST_SIZE;
694 		}
695 		if (ic->ic_data_crc32c) {
696 			icc->ulp_submode |= ULP_CRC_DATA;
697 			ic->ic_max_data_segment_length -=
698 			    ISCSI_DATA_DIGEST_SIZE;
699 		}
700 		so->so_options |= SO_NO_DDP;
701 		toep->params.ulp_mode = ULP_MODE_ISCSI;
702 		toep->ulpcb = icc;
703 
704 		send_iscsi_flowc_wr(icc->sc, toep, ci->max_tx_pdu_len);
705 		set_ulp_mode_iscsi(icc->sc, toep, ic->ic_header_crc32c,
706 		    ic->ic_data_crc32c);
707 		error = 0;
708 	}
709 	INP_WUNLOCK(inp);
710 
711 	if (error == 0) {
712 		error = icl_cxgbei_setsockopt(ic, so, ci->max_tx_pdu_len,
713 		    ci->max_rx_pdu_len);
714 	}
715 
716 	return (error);
717 }
718 
719 void
720 icl_cxgbei_conn_close(struct icl_conn *ic)
721 {
722 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
723 	struct icl_pdu *ip;
724 	struct socket *so;
725 	struct sockbuf *sb;
726 	struct inpcb *inp;
727 	struct toepcb *toep = icc->toep;
728 
729 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
730 	ICL_CONN_LOCK_ASSERT_NOT(ic);
731 
732 	ICL_CONN_LOCK(ic);
733 	so = ic->ic_socket;
734 	if (ic->ic_disconnecting || so == NULL) {
735 		CTR4(KTR_CXGBE, "%s: icc %p (disconnecting = %d), so %p",
736 		    __func__, icc, ic->ic_disconnecting, so);
737 		ICL_CONN_UNLOCK(ic);
738 		return;
739 	}
740 	ic->ic_disconnecting = true;
741 
742 	/* These are unused in this driver right now. */
743 	MPASS(STAILQ_EMPTY(&ic->ic_to_send));
744 	MPASS(ic->ic_receive_pdu == NULL);
745 
746 #ifdef DIAGNOSTIC
747 	KASSERT(ic->ic_outstanding_pdus == 0,
748 	    ("destroying session with %d outstanding PDUs",
749 	     ic->ic_outstanding_pdus));
750 #endif
751 	ICL_CONN_UNLOCK(ic);
752 
753 	CTR3(KTR_CXGBE, "%s: tid %d, icc %p", __func__, toep ? toep->tid : -1,
754 	    icc);
755 	inp = sotoinpcb(so);
756 	sb = &so->so_rcv;
757 	INP_WLOCK(inp);
758 	if (toep != NULL) {	/* NULL if connection was never offloaded. */
759 		toep->ulpcb = NULL;
760 		mbufq_drain(&toep->ulp_pduq);
761 		SOCKBUF_LOCK(sb);
762 		if (icc->rx_flags & RXF_ACTIVE) {
763 			volatile u_int *p = &icc->rx_flags;
764 
765 			SOCKBUF_UNLOCK(sb);
766 			INP_WUNLOCK(inp);
767 
768 			while (*p & RXF_ACTIVE)
769 				pause("conclo", 1);
770 
771 			INP_WLOCK(inp);
772 			SOCKBUF_LOCK(sb);
773 		}
774 
775 		while (!STAILQ_EMPTY(&icc->rcvd_pdus)) {
776 			ip = STAILQ_FIRST(&icc->rcvd_pdus);
777 			STAILQ_REMOVE_HEAD(&icc->rcvd_pdus, ip_next);
778 			icl_cxgbei_conn_pdu_free(ic, ip);
779 		}
780 		SOCKBUF_UNLOCK(sb);
781 	}
782 	INP_WUNLOCK(inp);
783 
784 	ICL_CONN_LOCK(ic);
785 	ic->ic_socket = NULL;
786 	ICL_CONN_UNLOCK(ic);
787 
788 	/*
789 	 * XXXNP: we should send RST instead of FIN when PDUs held in various
790 	 * queues were purged instead of delivered reliably but soabort isn't
791 	 * really general purpose and wouldn't do the right thing here.
792 	 */
793 	soclose(so);
794 }
795 
796 int
797 icl_cxgbei_conn_task_setup(struct icl_conn *ic, struct icl_pdu *ip,
798     struct ccb_scsiio *csio, uint32_t *ittp, void **arg)
799 {
800 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
801 	struct toepcb *toep = icc->toep;
802 	struct adapter *sc = icc->sc;
803 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
804 	struct ppod_region *pr = &ci->pr;
805 	struct ppod_reservation *prsv;
806 	uint32_t itt;
807 	int rc = 0;
808 
809 	/* This is for the offload driver's state.  Must not be set already. */
810 	MPASS(arg != NULL);
811 	MPASS(*arg == NULL);
812 
813 	if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_IN ||
814 	    csio->dxfer_len < ci->ddp_threshold) {
815 no_ddp:
816 		/*
817 		 * No DDP for this I/O.  Allocate an ITT (based on the one
818 		 * passed in) that cannot be a valid hardware DDP tag in the
819 		 * iSCSI region.
820 		 */
821 		itt = *ittp & M_PPOD_TAG;
822 		itt = V_PPOD_TAG(itt) | pr->pr_invalid_bit;
823 		*ittp = htobe32(itt);
824 		MPASS(*arg == NULL);	/* State is maintained for DDP only. */
825 		if (rc != 0)
826 			counter_u64_add(ci->ddp_setup_error, 1);
827 		return (0);
828 	}
829 
830 	/*
831 	 * Reserve resources for DDP, update the itt that should be used in the
832 	 * PDU, and save DDP specific state for this I/O in *arg.
833 	 */
834 
835 	prsv = uma_zalloc(prsv_zone, M_NOWAIT);
836 	if (prsv == NULL) {
837 		rc = ENOMEM;
838 		goto no_ddp;
839 	}
840 
841 	/* XXX add support for all CAM_DATA_ types */
842 	MPASS((csio->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_VADDR);
843 	rc = t4_alloc_page_pods_for_buf(pr, (vm_offset_t)csio->data_ptr,
844 	    csio->dxfer_len, prsv);
845 	if (rc != 0) {
846 		uma_zfree(prsv_zone, prsv);
847 		goto no_ddp;
848 	}
849 
850 	rc = t4_write_page_pods_for_buf(sc, toep->ofld_txq, toep->tid, prsv,
851 	    (vm_offset_t)csio->data_ptr, csio->dxfer_len);
852 	if (rc != 0) {
853 		t4_free_page_pods(prsv);
854 		uma_zfree(prsv_zone, prsv);
855 		goto no_ddp;
856 	}
857 
858 	*ittp = htobe32(prsv->prsv_tag);
859 	*arg = prsv;
860 	counter_u64_add(ci->ddp_setup_ok, 1);
861 	return (0);
862 }
863 
864 void
865 icl_cxgbei_conn_task_done(struct icl_conn *ic, void *arg)
866 {
867 
868 	if (arg != NULL) {
869 		struct ppod_reservation *prsv = arg;
870 
871 		t4_free_page_pods(prsv);
872 		uma_zfree(prsv_zone, prsv);
873 	}
874 }
875 
876 /* XXXNP: PDU should be passed in as parameter, like on the initiator. */
877 #define io_to_request_pdu(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND].ptr)
878 #define io_to_ppod_reservation(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND2].ptr)
879 
880 int
881 icl_cxgbei_conn_transfer_setup(struct icl_conn *ic, union ctl_io *io,
882     uint32_t *tttp, void **arg)
883 {
884 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
885 	struct toepcb *toep = icc->toep;
886 	struct ctl_scsiio *ctsio = &io->scsiio;
887 	struct adapter *sc = icc->sc;
888 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
889 	struct ppod_region *pr = &ci->pr;
890 	struct ppod_reservation *prsv;
891 	uint32_t ttt;
892 	int xferlen, rc = 0, alias;
893 
894 	/* This is for the offload driver's state.  Must not be set already. */
895 	MPASS(arg != NULL);
896 	MPASS(*arg == NULL);
897 
898 	if (ctsio->ext_data_filled == 0) {
899 		int first_burst;
900 		struct icl_pdu *ip = io_to_request_pdu(io);
901 		vm_offset_t buf;
902 #ifdef INVARIANTS
903 		struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
904 
905 		MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
906 		MPASS(ic == ip->ip_conn);
907 		MPASS(ip->ip_bhs_mbuf != NULL);
908 #endif
909 		first_burst = icl_pdu_data_segment_length(ip);
910 
911 		/*
912 		 * Note that ICL calls conn_transfer_setup even if the first
913 		 * burst had everything and there's nothing left to transfer.
914 		 */
915 		MPASS(ctsio->kern_data_len >= first_burst);
916 		xferlen = ctsio->kern_data_len;
917 		if (xferlen - first_burst < ci->ddp_threshold) {
918 no_ddp:
919 			/*
920 			 * No DDP for this transfer.  Allocate a TTT (based on
921 			 * the one passed in) that cannot be a valid hardware
922 			 * DDP tag in the iSCSI region.
923 			 */
924 			ttt = *tttp & M_PPOD_TAG;
925 			ttt = V_PPOD_TAG(ttt) | pr->pr_invalid_bit;
926 			*tttp = htobe32(ttt);
927 			MPASS(io_to_ppod_reservation(io) == NULL);
928 			if (rc != 0)
929 				counter_u64_add(ci->ddp_setup_error, 1);
930 			return (0);
931 		}
932 
933 		if (ctsio->kern_sg_entries == 0)
934 			buf = (vm_offset_t)ctsio->kern_data_ptr;
935 		else if (ctsio->kern_sg_entries == 1) {
936 			struct ctl_sg_entry *sgl = (void *)ctsio->kern_data_ptr;
937 
938 			MPASS(sgl->len == xferlen);
939 			buf = (vm_offset_t)sgl->addr;
940 		} else {
941 			rc = EAGAIN;	/* XXX implement */
942 			goto no_ddp;
943 		}
944 
945 
946 		/*
947 		 * Reserve resources for DDP, update the ttt that should be used
948 		 * in the PDU, and save DDP specific state for this I/O.
949 		 */
950 
951 		MPASS(io_to_ppod_reservation(io) == NULL);
952 		prsv = uma_zalloc(prsv_zone, M_NOWAIT);
953 		if (prsv == NULL) {
954 			rc = ENOMEM;
955 			goto no_ddp;
956 		}
957 
958 		rc = t4_alloc_page_pods_for_buf(pr, buf, xferlen, prsv);
959 		if (rc != 0) {
960 			uma_zfree(prsv_zone, prsv);
961 			goto no_ddp;
962 		}
963 
964 		rc = t4_write_page_pods_for_buf(sc, toep->ofld_txq, toep->tid,
965 		    prsv, buf, xferlen);
966 		if (rc != 0) {
967 			t4_free_page_pods(prsv);
968 			uma_zfree(prsv_zone, prsv);
969 			goto no_ddp;
970 		}
971 
972 		*tttp = htobe32(prsv->prsv_tag);
973 		io_to_ppod_reservation(io) = prsv;
974 		*arg = ctsio;
975 		counter_u64_add(ci->ddp_setup_ok, 1);
976 		return (0);
977 	}
978 
979 	/*
980 	 * In the middle of an I/O.  A non-NULL page pod reservation indicates
981 	 * that a DDP buffer is being used for the I/O.
982 	 */
983 
984 	prsv = io_to_ppod_reservation(ctsio);
985 	if (prsv == NULL)
986 		goto no_ddp;
987 
988 	alias = (prsv->prsv_tag & pr->pr_alias_mask) >> pr->pr_alias_shift;
989 	alias++;
990 	prsv->prsv_tag &= ~pr->pr_alias_mask;
991 	prsv->prsv_tag |= alias << pr->pr_alias_shift & pr->pr_alias_mask;
992 
993 	*tttp = htobe32(prsv->prsv_tag);
994 	*arg = ctsio;
995 
996 	return (0);
997 }
998 
999 void
1000 icl_cxgbei_conn_transfer_done(struct icl_conn *ic, void *arg)
1001 {
1002 	struct ctl_scsiio *ctsio = arg;
1003 
1004 	if (ctsio != NULL && ctsio->kern_data_len == ctsio->ext_data_filled) {
1005 		struct ppod_reservation *prsv;
1006 
1007 		prsv = io_to_ppod_reservation(ctsio);
1008 		MPASS(prsv != NULL);
1009 
1010 		t4_free_page_pods(prsv);
1011 		uma_zfree(prsv_zone, prsv);
1012 	}
1013 }
1014 
1015 static void
1016 cxgbei_limits(struct adapter *sc, void *arg)
1017 {
1018 	struct icl_drv_limits *idl = arg;
1019 	struct cxgbei_data *ci;
1020 	int max_dsl;
1021 
1022 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4lims") != 0)
1023 		return;
1024 
1025 	if (uld_active(sc, ULD_ISCSI)) {
1026 		ci = sc->iscsi_ulp_softc;
1027 		MPASS(ci != NULL);
1028 
1029 		/*
1030 		 * AHS is not supported by the kernel so we'll not account for
1031 		 * it either in our PDU len -> data segment len conversions.
1032 		 */
1033 
1034 		max_dsl = ci->max_rx_pdu_len - ISCSI_BHS_SIZE -
1035 		    ISCSI_HEADER_DIGEST_SIZE - ISCSI_DATA_DIGEST_SIZE;
1036 		if (idl->idl_max_recv_data_segment_length > max_dsl)
1037 			idl->idl_max_recv_data_segment_length = max_dsl;
1038 
1039 		max_dsl = ci->max_tx_pdu_len - ISCSI_BHS_SIZE -
1040 		    ISCSI_HEADER_DIGEST_SIZE - ISCSI_DATA_DIGEST_SIZE;
1041 		if (idl->idl_max_send_data_segment_length > max_dsl)
1042 			idl->idl_max_send_data_segment_length = max_dsl;
1043 	}
1044 
1045 	end_synchronized_op(sc, LOCK_HELD);
1046 }
1047 
1048 static int
1049 icl_cxgbei_limits(struct icl_drv_limits *idl)
1050 {
1051 
1052 	/* Maximum allowed by the RFC.  cxgbei_limits will clip them. */
1053 	idl->idl_max_recv_data_segment_length = (1 << 24) - 1;
1054 	idl->idl_max_send_data_segment_length = (1 << 24) - 1;
1055 
1056 	/* These are somewhat arbitrary. */
1057 	idl->idl_max_burst_length = 2 * 1024 * 1024;
1058 	idl->idl_first_burst_length = 8192;
1059 
1060 	t4_iterate(cxgbei_limits, idl);
1061 
1062 	return (0);
1063 }
1064 
1065 int
1066 icl_cxgbei_mod_load(void)
1067 {
1068 	int rc;
1069 
1070 	/*
1071 	 * Space to track pagepod reservations.
1072 	 */
1073 	prsv_zone = uma_zcreate("Pagepod reservations",
1074 	    sizeof(struct ppod_reservation), NULL, NULL, NULL, NULL,
1075 	    UMA_ALIGN_CACHE, 0);
1076 
1077 	refcount_init(&icl_cxgbei_ncons, 0);
1078 
1079 	rc = icl_register("cxgbei", false, -100, icl_cxgbei_limits,
1080 	    icl_cxgbei_new_conn);
1081 
1082 	return (rc);
1083 }
1084 
1085 int
1086 icl_cxgbei_mod_unload(void)
1087 {
1088 
1089 	if (icl_cxgbei_ncons != 0)
1090 		return (EBUSY);
1091 
1092 	icl_unregister("cxgbei", false);
1093 
1094 	uma_zdestroy(prsv_zone);
1095 
1096 	return (0);
1097 }
1098 #endif
1099