xref: /freebsd/sys/dev/cxgbe/cxgbei/icl_cxgbei.c (revision ac099daf6742ead81ea7ea86351a8ef4e783041b)
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 static icl_conn_new_pdu_t	icl_cxgbei_conn_new_pdu;
128 static icl_conn_pdu_data_segment_length_t
129 				    icl_cxgbei_conn_pdu_data_segment_length;
130 static icl_conn_pdu_append_data_t	icl_cxgbei_conn_pdu_append_data;
131 static icl_conn_pdu_get_data_t	icl_cxgbei_conn_pdu_get_data;
132 static icl_conn_pdu_queue_t	icl_cxgbei_conn_pdu_queue;
133 static icl_conn_handoff_t	icl_cxgbei_conn_handoff;
134 static icl_conn_free_t		icl_cxgbei_conn_free;
135 static icl_conn_close_t		icl_cxgbei_conn_close;
136 static icl_conn_task_setup_t	icl_cxgbei_conn_task_setup;
137 static icl_conn_task_done_t	icl_cxgbei_conn_task_done;
138 static icl_conn_transfer_setup_t	icl_cxgbei_conn_transfer_setup;
139 static icl_conn_transfer_done_t	icl_cxgbei_conn_transfer_done;
140 
141 static kobj_method_t icl_cxgbei_methods[] = {
142 	KOBJMETHOD(icl_conn_new_pdu, icl_cxgbei_conn_new_pdu),
143 	KOBJMETHOD(icl_conn_pdu_free, icl_cxgbei_conn_pdu_free),
144 	KOBJMETHOD(icl_conn_pdu_data_segment_length,
145 	    icl_cxgbei_conn_pdu_data_segment_length),
146 	KOBJMETHOD(icl_conn_pdu_append_data, icl_cxgbei_conn_pdu_append_data),
147 	KOBJMETHOD(icl_conn_pdu_get_data, icl_cxgbei_conn_pdu_get_data),
148 	KOBJMETHOD(icl_conn_pdu_queue, icl_cxgbei_conn_pdu_queue),
149 	KOBJMETHOD(icl_conn_handoff, icl_cxgbei_conn_handoff),
150 	KOBJMETHOD(icl_conn_free, icl_cxgbei_conn_free),
151 	KOBJMETHOD(icl_conn_close, icl_cxgbei_conn_close),
152 	KOBJMETHOD(icl_conn_task_setup, icl_cxgbei_conn_task_setup),
153 	KOBJMETHOD(icl_conn_task_done, icl_cxgbei_conn_task_done),
154 	KOBJMETHOD(icl_conn_transfer_setup, icl_cxgbei_conn_transfer_setup),
155 	KOBJMETHOD(icl_conn_transfer_done, icl_cxgbei_conn_transfer_done),
156 	{ 0, 0 }
157 };
158 
159 DEFINE_CLASS(icl_cxgbei, icl_cxgbei_methods, sizeof(struct icl_cxgbei_conn));
160 
161 void
162 icl_cxgbei_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip)
163 {
164 #ifdef INVARIANTS
165 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
166 #endif
167 
168 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
169 	MPASS(ic == ip->ip_conn);
170 	MPASS(ip->ip_bhs_mbuf != NULL);
171 
172 	m_freem(ip->ip_ahs_mbuf);
173 	m_freem(ip->ip_data_mbuf);
174 	m_freem(ip->ip_bhs_mbuf);	/* storage for icl_cxgbei_pdu itself */
175 
176 #ifdef DIAGNOSTIC
177 	if (__predict_true(ic != NULL))
178 		refcount_release(&ic->ic_outstanding_pdus);
179 #endif
180 }
181 
182 struct icl_pdu *
183 icl_cxgbei_new_pdu(int flags)
184 {
185 	struct icl_cxgbei_pdu *icp;
186 	struct icl_pdu *ip;
187 	struct mbuf *m;
188 	uintptr_t a;
189 
190 	m = m_gethdr(flags, MT_DATA);
191 	if (__predict_false(m == NULL))
192 		return (NULL);
193 
194 	a = roundup2(mtod(m, uintptr_t), _Alignof(struct icl_cxgbei_pdu));
195 	icp = (struct icl_cxgbei_pdu *)a;
196 	bzero(icp, sizeof(*icp));
197 
198 	icp->icp_signature = CXGBEI_PDU_SIGNATURE;
199 	ip = &icp->ip;
200 	ip->ip_bhs_mbuf = m;
201 
202 	a = roundup2((uintptr_t)(icp + 1), _Alignof(struct iscsi_bhs *));
203 	ip->ip_bhs = (struct iscsi_bhs *)a;
204 #ifdef INVARIANTS
205 	/* Everything must fit entirely in the mbuf. */
206 	a = (uintptr_t)(ip->ip_bhs + 1);
207 	MPASS(a <= (uintptr_t)m + MSIZE);
208 #endif
209 	bzero(ip->ip_bhs, sizeof(*ip->ip_bhs));
210 
211 	m->m_data = (void *)ip->ip_bhs;
212 	m->m_len = sizeof(struct iscsi_bhs);
213 	m->m_pkthdr.len = m->m_len;
214 
215 	return (ip);
216 }
217 
218 void
219 icl_cxgbei_new_pdu_set_conn(struct icl_pdu *ip, struct icl_conn *ic)
220 {
221 
222 	ip->ip_conn = ic;
223 #ifdef DIAGNOSTIC
224 	refcount_acquire(&ic->ic_outstanding_pdus);
225 #endif
226 }
227 
228 /*
229  * Allocate icl_pdu with empty BHS to fill up by the caller.
230  */
231 static struct icl_pdu *
232 icl_cxgbei_conn_new_pdu(struct icl_conn *ic, int flags)
233 {
234 	struct icl_pdu *ip;
235 
236 	ip = icl_cxgbei_new_pdu(flags);
237 	if (__predict_false(ip == NULL))
238 		return (NULL);
239 	icl_cxgbei_new_pdu_set_conn(ip, ic);
240 
241 	return (ip);
242 }
243 
244 static size_t
245 icl_pdu_data_segment_length(const struct icl_pdu *request)
246 {
247 	uint32_t len = 0;
248 
249 	len += request->ip_bhs->bhs_data_segment_len[0];
250 	len <<= 8;
251 	len += request->ip_bhs->bhs_data_segment_len[1];
252 	len <<= 8;
253 	len += request->ip_bhs->bhs_data_segment_len[2];
254 
255 	return (len);
256 }
257 
258 size_t
259 icl_cxgbei_conn_pdu_data_segment_length(struct icl_conn *ic,
260     const struct icl_pdu *request)
261 {
262 
263 	return (icl_pdu_data_segment_length(request));
264 }
265 
266 static struct mbuf *
267 finalize_pdu(struct icl_cxgbei_conn *icc, struct icl_cxgbei_pdu *icp)
268 {
269 	struct icl_pdu *ip = &icp->ip;
270 	uint8_t ulp_submode, padding;
271 	struct mbuf *m, *last;
272 	struct iscsi_bhs *bhs;
273 
274 	/*
275 	 * Fix up the data segment mbuf first.
276 	 */
277 	m = ip->ip_data_mbuf;
278 	ulp_submode = icc->ulp_submode;
279 	if (m) {
280 		last = m_last(m);
281 
282 		/*
283 		 * Round up the data segment to a 4B boundary.  Pad with 0 if
284 		 * necessary.  There will definitely be room in the mbuf.
285 		 */
286 		padding = roundup2(ip->ip_data_len, 4) - ip->ip_data_len;
287 		if (padding) {
288 			bzero(mtod(last, uint8_t *) + last->m_len, padding);
289 			last->m_len += padding;
290 		}
291 	} else {
292 		MPASS(ip->ip_data_len == 0);
293 		ulp_submode &= ~ULP_CRC_DATA;
294 		padding = 0;
295 	}
296 
297 	/*
298 	 * Now the header mbuf that has the BHS.
299 	 */
300 	m = ip->ip_bhs_mbuf;
301 	MPASS(m->m_pkthdr.len == sizeof(struct iscsi_bhs));
302 	MPASS(m->m_len == sizeof(struct iscsi_bhs));
303 
304 	bhs = ip->ip_bhs;
305 	bhs->bhs_data_segment_len[2] = ip->ip_data_len;
306 	bhs->bhs_data_segment_len[1] = ip->ip_data_len >> 8;
307 	bhs->bhs_data_segment_len[0] = ip->ip_data_len >> 16;
308 
309 	/* "Convert" PDU to mbuf chain.  Do not use icp/ip after this. */
310 	m->m_pkthdr.len = sizeof(struct iscsi_bhs) + ip->ip_data_len + padding;
311 	m->m_next = ip->ip_data_mbuf;
312 	set_mbuf_ulp_submode(m, ulp_submode);
313 #ifdef INVARIANTS
314 	bzero(icp, sizeof(*icp));
315 #endif
316 #ifdef DIAGNOSTIC
317 	refcount_release(&icc->ic.ic_outstanding_pdus);
318 #endif
319 
320 	return (m);
321 }
322 
323 int
324 icl_cxgbei_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *ip,
325     const void *addr, size_t len, int flags)
326 {
327 	struct mbuf *m;
328 #ifdef INVARIANTS
329 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
330 #endif
331 
332 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
333 	MPASS(ic == ip->ip_conn);
334 	KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len));
335 
336 	m = ip->ip_data_mbuf;
337 	if (m == NULL) {
338 		m = m_getjcl(M_NOWAIT, MT_DATA, 0, MJUM16BYTES);
339 		if (__predict_false(m == NULL))
340 			return (ENOMEM);
341 
342 		ip->ip_data_mbuf = m;
343 	}
344 
345 	if (__predict_true(m_append(m, len, addr) != 0)) {
346 		ip->ip_data_len += len;
347 		MPASS(ip->ip_data_len <= ic->ic_max_data_segment_length);
348 		return (0);
349 	} else {
350 	    	if (flags & M_WAITOK) {
351 			CXGBE_UNIMPLEMENTED("fail safe append");
352 		}
353 		ip->ip_data_len = m_length(m, NULL);
354 		return (1);
355 	}
356 }
357 
358 void
359 icl_cxgbei_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip,
360     size_t off, void *addr, size_t len)
361 {
362 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
363 
364 	if (icp->icp_flags & ICPF_RX_DDP)
365 		return; /* data is DDP'ed, no need to copy */
366 	m_copydata(ip->ip_data_mbuf, off, len, addr);
367 }
368 
369 void
370 icl_cxgbei_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip)
371 {
372 	struct epoch_tracker et;
373 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
374 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
375 	struct socket *so = ic->ic_socket;
376 	struct toepcb *toep = icc->toep;
377 	struct inpcb *inp;
378 	struct mbuf *m;
379 
380 	MPASS(ic == ip->ip_conn);
381 	MPASS(ip->ip_bhs_mbuf != NULL);
382 	/* The kernel doesn't generate PDUs with AHS. */
383 	MPASS(ip->ip_ahs_mbuf == NULL && ip->ip_ahs_len == 0);
384 
385 	ICL_CONN_LOCK_ASSERT(ic);
386 	/* NOTE: sowriteable without so_snd lock is a mostly harmless race. */
387 	if (ic->ic_disconnecting || so == NULL || !sowriteable(so)) {
388 		icl_cxgbei_conn_pdu_free(ic, ip);
389 		return;
390 	}
391 
392 	m = finalize_pdu(icc, icp);
393 	M_ASSERTPKTHDR(m);
394 	MPASS((m->m_pkthdr.len & 3) == 0);
395 
396 	/*
397 	 * Do not get inp from toep->inp as the toepcb might have detached
398 	 * already.
399 	 */
400 	inp = sotoinpcb(so);
401 	CURVNET_SET(toep->vnet);
402 	NET_EPOCH_ENTER(et);
403 	INP_WLOCK(inp);
404 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) ||
405 	    __predict_false((toep->flags & TPF_ATTACHED) == 0))
406 		m_freem(m);
407 	else {
408 		mbufq_enqueue(&toep->ulp_pduq, m);
409 		t4_push_pdus(icc->sc, toep, 0);
410 	}
411 	INP_WUNLOCK(inp);
412 	NET_EPOCH_EXIT(et);
413 	CURVNET_RESTORE();
414 }
415 
416 static struct icl_conn *
417 icl_cxgbei_new_conn(const char *name, struct mtx *lock)
418 {
419 	struct icl_cxgbei_conn *icc;
420 	struct icl_conn *ic;
421 
422 	refcount_acquire(&icl_cxgbei_ncons);
423 
424 	icc = (struct icl_cxgbei_conn *)kobj_create(&icl_cxgbei_class, M_CXGBE,
425 	    M_WAITOK | M_ZERO);
426 	icc->icc_signature = CXGBEI_CONN_SIGNATURE;
427 	STAILQ_INIT(&icc->rcvd_pdus);
428 
429 	ic = &icc->ic;
430 	ic->ic_lock = lock;
431 
432 	/* XXXNP: review.  Most of these icl_conn fields aren't really used */
433 	STAILQ_INIT(&ic->ic_to_send);
434 	cv_init(&ic->ic_send_cv, "icl_cxgbei_tx");
435 	cv_init(&ic->ic_receive_cv, "icl_cxgbei_rx");
436 #ifdef DIAGNOSTIC
437 	refcount_init(&ic->ic_outstanding_pdus, 0);
438 #endif
439 	/* This is a stop-gap value that will be corrected during handoff. */
440 	ic->ic_max_data_segment_length = 16384;
441 	ic->ic_name = name;
442 	ic->ic_offload = "cxgbei";
443 	ic->ic_unmapped = false;
444 
445 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
446 
447 	return (ic);
448 }
449 
450 void
451 icl_cxgbei_conn_free(struct icl_conn *ic)
452 {
453 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
454 
455 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
456 
457 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
458 
459 	cv_destroy(&ic->ic_send_cv);
460 	cv_destroy(&ic->ic_receive_cv);
461 
462 	kobj_delete((struct kobj *)icc, M_CXGBE);
463 	refcount_release(&icl_cxgbei_ncons);
464 }
465 
466 static int
467 icl_cxgbei_setsockopt(struct icl_conn *ic, struct socket *so, int sspace,
468     int rspace)
469 {
470 	struct sockopt opt;
471 	int error, one = 1, ss, rs;
472 
473 	ss = max(sendspace, sspace);
474 	rs = max(recvspace, rspace);
475 
476 	error = soreserve(so, ss, rs);
477 	if (error != 0) {
478 		icl_cxgbei_conn_close(ic);
479 		return (error);
480 	}
481 	SOCKBUF_LOCK(&so->so_snd);
482 	so->so_snd.sb_flags |= SB_AUTOSIZE;
483 	SOCKBUF_UNLOCK(&so->so_snd);
484 	SOCKBUF_LOCK(&so->so_rcv);
485 	so->so_rcv.sb_flags |= SB_AUTOSIZE;
486 	SOCKBUF_UNLOCK(&so->so_rcv);
487 
488 	/*
489 	 * Disable Nagle.
490 	 */
491 	bzero(&opt, sizeof(opt));
492 	opt.sopt_dir = SOPT_SET;
493 	opt.sopt_level = IPPROTO_TCP;
494 	opt.sopt_name = TCP_NODELAY;
495 	opt.sopt_val = &one;
496 	opt.sopt_valsize = sizeof(one);
497 	error = sosetopt(so, &opt);
498 	if (error != 0) {
499 		icl_cxgbei_conn_close(ic);
500 		return (error);
501 	}
502 
503 	return (0);
504 }
505 
506 /*
507  * Request/response structure used to find out the adapter offloading a socket.
508  */
509 struct find_ofld_adapter_rr {
510 	struct socket *so;
511 	struct adapter *sc;	/* result */
512 };
513 
514 static void
515 find_offload_adapter(struct adapter *sc, void *arg)
516 {
517 	struct find_ofld_adapter_rr *fa = arg;
518 	struct socket *so = fa->so;
519 	struct tom_data *td = sc->tom_softc;
520 	struct tcpcb *tp;
521 	struct inpcb *inp;
522 
523 	/* Non-TCP were filtered out earlier. */
524 	MPASS(so->so_proto->pr_protocol == IPPROTO_TCP);
525 
526 	if (fa->sc != NULL)
527 		return;	/* Found already. */
528 
529 	if (td == NULL)
530 		return;	/* TOE not enabled on this adapter. */
531 
532 	inp = sotoinpcb(so);
533 	INP_WLOCK(inp);
534 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
535 		tp = intotcpcb(inp);
536 		if (tp->t_flags & TF_TOE && tp->tod == &td->tod)
537 			fa->sc = sc;	/* Found. */
538 	}
539 	INP_WUNLOCK(inp);
540 }
541 
542 /* XXXNP: move this to t4_tom. */
543 static void
544 send_iscsi_flowc_wr(struct adapter *sc, struct toepcb *toep, int maxlen)
545 {
546 	struct wrqe *wr;
547 	struct fw_flowc_wr *flowc;
548 	const u_int nparams = 1;
549 	u_int flowclen;
550 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
551 
552 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
553 
554 	wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq);
555 	if (wr == NULL) {
556 		/* XXX */
557 		panic("%s: allocation failure.", __func__);
558 	}
559 	flowc = wrtod(wr);
560 	memset(flowc, 0, wr->wr_len);
561 
562 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
563 	    V_FW_FLOWC_WR_NPARAMS(nparams));
564 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
565 	    V_FW_WR_FLOWID(toep->tid));
566 
567 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_TXDATAPLEN_MAX;
568 	flowc->mnemval[0].val = htobe32(maxlen);
569 
570 	txsd->tx_credits = howmany(flowclen, 16);
571 	txsd->plen = 0;
572 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
573 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
574 	toep->tx_credits -= txsd->tx_credits;
575 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
576 		toep->txsd_pidx = 0;
577 	toep->txsd_avail--;
578 
579         t4_wrq_tx(sc, wr);
580 }
581 
582 static void
583 set_ulp_mode_iscsi(struct adapter *sc, struct toepcb *toep, u_int ulp_submode)
584 {
585 	uint64_t val;
586 
587 	CTR3(KTR_CXGBE, "%s: tid %u, ULP_MODE_ISCSI, submode=%#x",
588 	    __func__, toep->tid, ulp_submode);
589 
590 	val = V_TCB_ULP_TYPE(ULP_MODE_ISCSI) | V_TCB_ULP_RAW(ulp_submode);
591 	t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_ULP_TYPE,
592 	    V_TCB_ULP_TYPE(M_TCB_ULP_TYPE) | V_TCB_ULP_RAW(M_TCB_ULP_RAW), val,
593 	    0, 0);
594 }
595 
596 /*
597  * XXXNP: Who is responsible for cleaning up the socket if this returns with an
598  * error?  Review all error paths.
599  *
600  * XXXNP: What happens to the socket's fd reference if the operation is
601  * successful, and how does that affect the socket's life cycle?
602  */
603 int
604 icl_cxgbei_conn_handoff(struct icl_conn *ic, int fd)
605 {
606 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
607 	struct cxgbei_data *ci;
608 	struct find_ofld_adapter_rr fa;
609 	struct file *fp;
610 	struct socket *so;
611 	struct inpcb *inp;
612 	struct tcpcb *tp;
613 	struct toepcb *toep;
614 	cap_rights_t rights;
615 	int error;
616 
617 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
618 	ICL_CONN_LOCK_ASSERT_NOT(ic);
619 
620 	/*
621 	 * Steal the socket from userland.
622 	 */
623 	error = fget(curthread, fd,
624 	    cap_rights_init_one(&rights, CAP_SOCK_CLIENT), &fp);
625 	if (error != 0)
626 		return (error);
627 	if (fp->f_type != DTYPE_SOCKET) {
628 		fdrop(fp, curthread);
629 		return (EINVAL);
630 	}
631 	so = fp->f_data;
632 	if (so->so_type != SOCK_STREAM ||
633 	    so->so_proto->pr_protocol != IPPROTO_TCP) {
634 		fdrop(fp, curthread);
635 		return (EINVAL);
636 	}
637 
638 	ICL_CONN_LOCK(ic);
639 	if (ic->ic_socket != NULL) {
640 		ICL_CONN_UNLOCK(ic);
641 		fdrop(fp, curthread);
642 		return (EBUSY);
643 	}
644 	ic->ic_disconnecting = false;
645 	ic->ic_socket = so;
646 	fp->f_ops = &badfileops;
647 	fp->f_data = NULL;
648 	fdrop(fp, curthread);
649 	ICL_CONN_UNLOCK(ic);
650 
651 	/* Find the adapter offloading this socket. */
652 	fa.sc = NULL;
653 	fa.so = so;
654 	t4_iterate(find_offload_adapter, &fa);
655 	if (fa.sc == NULL)
656 		return (EINVAL);
657 	icc->sc = fa.sc;
658 	ci = icc->sc->iscsi_ulp_softc;
659 
660 	inp = sotoinpcb(so);
661 	INP_WLOCK(inp);
662 	tp = intotcpcb(inp);
663 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))
664 		error = EBUSY;
665 	else {
666 		/*
667 		 * socket could not have been "unoffloaded" if here.
668 		 */
669 		MPASS(tp->t_flags & TF_TOE);
670 		MPASS(tp->tod != NULL);
671 		MPASS(tp->t_toe != NULL);
672 		toep = tp->t_toe;
673 		MPASS(toep->vi->adapter == icc->sc);
674 		icc->toep = toep;
675 		icc->cwt = cxgbei_select_worker_thread(icc);
676 
677 		/*
678 		 * We maintain the _send_ DSL in this field just to have a
679 		 * convenient way to assert that the kernel never sends
680 		 * oversized PDUs.  This field is otherwise unused in the driver
681 		 * or the kernel.
682 		 */
683 		ic->ic_max_data_segment_length = ci->max_tx_pdu_len -
684 		    ISCSI_BHS_SIZE;
685 
686 		icc->ulp_submode = 0;
687 		if (ic->ic_header_crc32c) {
688 			icc->ulp_submode |= ULP_CRC_HEADER;
689 			ic->ic_max_data_segment_length -=
690 			    ISCSI_HEADER_DIGEST_SIZE;
691 		}
692 		if (ic->ic_data_crc32c) {
693 			icc->ulp_submode |= ULP_CRC_DATA;
694 			ic->ic_max_data_segment_length -=
695 			    ISCSI_DATA_DIGEST_SIZE;
696 		}
697 		so->so_options |= SO_NO_DDP;
698 		toep->params.ulp_mode = ULP_MODE_ISCSI;
699 		toep->ulpcb = icc;
700 
701 		send_iscsi_flowc_wr(icc->sc, toep, ci->max_tx_pdu_len);
702 		set_ulp_mode_iscsi(icc->sc, toep, icc->ulp_submode);
703 		error = 0;
704 	}
705 	INP_WUNLOCK(inp);
706 
707 	if (error == 0) {
708 		error = icl_cxgbei_setsockopt(ic, so, ci->max_tx_pdu_len,
709 		    ci->max_rx_pdu_len);
710 	}
711 
712 	return (error);
713 }
714 
715 void
716 icl_cxgbei_conn_close(struct icl_conn *ic)
717 {
718 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
719 	struct icl_pdu *ip;
720 	struct socket *so;
721 	struct sockbuf *sb;
722 	struct inpcb *inp;
723 	struct toepcb *toep = icc->toep;
724 
725 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
726 	ICL_CONN_LOCK_ASSERT_NOT(ic);
727 
728 	ICL_CONN_LOCK(ic);
729 	so = ic->ic_socket;
730 	if (ic->ic_disconnecting || so == NULL) {
731 		CTR4(KTR_CXGBE, "%s: icc %p (disconnecting = %d), so %p",
732 		    __func__, icc, ic->ic_disconnecting, so);
733 		ICL_CONN_UNLOCK(ic);
734 		return;
735 	}
736 	ic->ic_disconnecting = true;
737 
738 	/* These are unused in this driver right now. */
739 	MPASS(STAILQ_EMPTY(&ic->ic_to_send));
740 	MPASS(ic->ic_receive_pdu == NULL);
741 
742 #ifdef DIAGNOSTIC
743 	KASSERT(ic->ic_outstanding_pdus == 0,
744 	    ("destroying session with %d outstanding PDUs",
745 	     ic->ic_outstanding_pdus));
746 #endif
747 	ICL_CONN_UNLOCK(ic);
748 
749 	CTR3(KTR_CXGBE, "%s: tid %d, icc %p", __func__, toep ? toep->tid : -1,
750 	    icc);
751 	inp = sotoinpcb(so);
752 	sb = &so->so_rcv;
753 	INP_WLOCK(inp);
754 	if (toep != NULL) {	/* NULL if connection was never offloaded. */
755 		toep->ulpcb = NULL;
756 		mbufq_drain(&toep->ulp_pduq);
757 		SOCKBUF_LOCK(sb);
758 		if (icc->rx_flags & RXF_ACTIVE) {
759 			volatile u_int *p = &icc->rx_flags;
760 
761 			SOCKBUF_UNLOCK(sb);
762 			INP_WUNLOCK(inp);
763 
764 			while (*p & RXF_ACTIVE)
765 				pause("conclo", 1);
766 
767 			INP_WLOCK(inp);
768 			SOCKBUF_LOCK(sb);
769 		}
770 
771 		while (!STAILQ_EMPTY(&icc->rcvd_pdus)) {
772 			ip = STAILQ_FIRST(&icc->rcvd_pdus);
773 			STAILQ_REMOVE_HEAD(&icc->rcvd_pdus, ip_next);
774 			icl_cxgbei_conn_pdu_free(ic, ip);
775 		}
776 		SOCKBUF_UNLOCK(sb);
777 	}
778 	INP_WUNLOCK(inp);
779 
780 	ICL_CONN_LOCK(ic);
781 	ic->ic_socket = NULL;
782 	ICL_CONN_UNLOCK(ic);
783 
784 	/*
785 	 * XXXNP: we should send RST instead of FIN when PDUs held in various
786 	 * queues were purged instead of delivered reliably but soabort isn't
787 	 * really general purpose and wouldn't do the right thing here.
788 	 */
789 	soclose(so);
790 }
791 
792 int
793 icl_cxgbei_conn_task_setup(struct icl_conn *ic, struct icl_pdu *ip,
794     struct ccb_scsiio *csio, uint32_t *ittp, void **arg)
795 {
796 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
797 	struct toepcb *toep = icc->toep;
798 	struct adapter *sc = icc->sc;
799 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
800 	struct ppod_region *pr = &ci->pr;
801 	struct ppod_reservation *prsv;
802 	uint32_t itt;
803 	int rc = 0;
804 
805 	/* This is for the offload driver's state.  Must not be set already. */
806 	MPASS(arg != NULL);
807 	MPASS(*arg == NULL);
808 
809 	if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_IN ||
810 	    csio->dxfer_len < ci->ddp_threshold) {
811 no_ddp:
812 		/*
813 		 * No DDP for this I/O.  Allocate an ITT (based on the one
814 		 * passed in) that cannot be a valid hardware DDP tag in the
815 		 * iSCSI region.
816 		 */
817 		itt = *ittp & M_PPOD_TAG;
818 		itt = V_PPOD_TAG(itt) | pr->pr_invalid_bit;
819 		*ittp = htobe32(itt);
820 		MPASS(*arg == NULL);	/* State is maintained for DDP only. */
821 		if (rc != 0)
822 			counter_u64_add(ci->ddp_setup_error, 1);
823 		return (0);
824 	}
825 
826 	/*
827 	 * Reserve resources for DDP, update the itt that should be used in the
828 	 * PDU, and save DDP specific state for this I/O in *arg.
829 	 */
830 
831 	prsv = uma_zalloc(prsv_zone, M_NOWAIT);
832 	if (prsv == NULL) {
833 		rc = ENOMEM;
834 		goto no_ddp;
835 	}
836 
837 	/* XXX add support for all CAM_DATA_ types */
838 	MPASS((csio->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_VADDR);
839 	rc = t4_alloc_page_pods_for_buf(pr, (vm_offset_t)csio->data_ptr,
840 	    csio->dxfer_len, prsv);
841 	if (rc != 0) {
842 		uma_zfree(prsv_zone, prsv);
843 		goto no_ddp;
844 	}
845 
846 	rc = t4_write_page_pods_for_buf(sc, &toep->ofld_txq->wrq, toep->tid,
847 	    prsv, (vm_offset_t)csio->data_ptr, csio->dxfer_len);
848 	if (rc != 0) {
849 		t4_free_page_pods(prsv);
850 		uma_zfree(prsv_zone, prsv);
851 		goto no_ddp;
852 	}
853 
854 	*ittp = htobe32(prsv->prsv_tag);
855 	*arg = prsv;
856 	counter_u64_add(ci->ddp_setup_ok, 1);
857 	return (0);
858 }
859 
860 void
861 icl_cxgbei_conn_task_done(struct icl_conn *ic, void *arg)
862 {
863 
864 	if (arg != NULL) {
865 		struct ppod_reservation *prsv = arg;
866 
867 		t4_free_page_pods(prsv);
868 		uma_zfree(prsv_zone, prsv);
869 	}
870 }
871 
872 /* XXXNP: PDU should be passed in as parameter, like on the initiator. */
873 #define io_to_request_pdu(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND].ptr)
874 #define io_to_ppod_reservation(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND2].ptr)
875 
876 int
877 icl_cxgbei_conn_transfer_setup(struct icl_conn *ic, union ctl_io *io,
878     uint32_t *tttp, void **arg)
879 {
880 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
881 	struct toepcb *toep = icc->toep;
882 	struct ctl_scsiio *ctsio = &io->scsiio;
883 	struct adapter *sc = icc->sc;
884 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
885 	struct ppod_region *pr = &ci->pr;
886 	struct ppod_reservation *prsv;
887 	uint32_t ttt;
888 	int xferlen, rc = 0, alias;
889 
890 	/* This is for the offload driver's state.  Must not be set already. */
891 	MPASS(arg != NULL);
892 	MPASS(*arg == NULL);
893 
894 	if (ctsio->ext_data_filled == 0) {
895 		int first_burst;
896 		struct icl_pdu *ip = io_to_request_pdu(io);
897 		vm_offset_t buf;
898 #ifdef INVARIANTS
899 		struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
900 
901 		MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
902 		MPASS(ic == ip->ip_conn);
903 		MPASS(ip->ip_bhs_mbuf != NULL);
904 #endif
905 		first_burst = icl_pdu_data_segment_length(ip);
906 
907 		/*
908 		 * Note that ICL calls conn_transfer_setup even if the first
909 		 * burst had everything and there's nothing left to transfer.
910 		 */
911 		MPASS(ctsio->kern_data_len >= first_burst);
912 		xferlen = ctsio->kern_data_len;
913 		if (xferlen - first_burst < ci->ddp_threshold) {
914 no_ddp:
915 			/*
916 			 * No DDP for this transfer.  Allocate a TTT (based on
917 			 * the one passed in) that cannot be a valid hardware
918 			 * DDP tag in the iSCSI region.
919 			 */
920 			ttt = *tttp & M_PPOD_TAG;
921 			ttt = V_PPOD_TAG(ttt) | pr->pr_invalid_bit;
922 			*tttp = htobe32(ttt);
923 			MPASS(io_to_ppod_reservation(io) == NULL);
924 			if (rc != 0)
925 				counter_u64_add(ci->ddp_setup_error, 1);
926 			return (0);
927 		}
928 
929 		if (ctsio->kern_sg_entries == 0)
930 			buf = (vm_offset_t)ctsio->kern_data_ptr;
931 		else if (ctsio->kern_sg_entries == 1) {
932 			struct ctl_sg_entry *sgl = (void *)ctsio->kern_data_ptr;
933 
934 			MPASS(sgl->len == xferlen);
935 			buf = (vm_offset_t)sgl->addr;
936 		} else {
937 			rc = EAGAIN;	/* XXX implement */
938 			goto no_ddp;
939 		}
940 
941 
942 		/*
943 		 * Reserve resources for DDP, update the ttt that should be used
944 		 * in the PDU, and save DDP specific state for this I/O.
945 		 */
946 
947 		MPASS(io_to_ppod_reservation(io) == NULL);
948 		prsv = uma_zalloc(prsv_zone, M_NOWAIT);
949 		if (prsv == NULL) {
950 			rc = ENOMEM;
951 			goto no_ddp;
952 		}
953 
954 		rc = t4_alloc_page_pods_for_buf(pr, buf, xferlen, prsv);
955 		if (rc != 0) {
956 			uma_zfree(prsv_zone, prsv);
957 			goto no_ddp;
958 		}
959 
960 		rc = t4_write_page_pods_for_buf(sc, &toep->ofld_txq->wrq,
961 		    toep->tid, prsv, buf, xferlen);
962 		if (rc != 0) {
963 			t4_free_page_pods(prsv);
964 			uma_zfree(prsv_zone, prsv);
965 			goto no_ddp;
966 		}
967 
968 		*tttp = htobe32(prsv->prsv_tag);
969 		io_to_ppod_reservation(io) = prsv;
970 		*arg = ctsio;
971 		counter_u64_add(ci->ddp_setup_ok, 1);
972 		return (0);
973 	}
974 
975 	/*
976 	 * In the middle of an I/O.  A non-NULL page pod reservation indicates
977 	 * that a DDP buffer is being used for the I/O.
978 	 */
979 
980 	prsv = io_to_ppod_reservation(ctsio);
981 	if (prsv == NULL)
982 		goto no_ddp;
983 
984 	alias = (prsv->prsv_tag & pr->pr_alias_mask) >> pr->pr_alias_shift;
985 	alias++;
986 	prsv->prsv_tag &= ~pr->pr_alias_mask;
987 	prsv->prsv_tag |= alias << pr->pr_alias_shift & pr->pr_alias_mask;
988 
989 	*tttp = htobe32(prsv->prsv_tag);
990 	*arg = ctsio;
991 
992 	return (0);
993 }
994 
995 void
996 icl_cxgbei_conn_transfer_done(struct icl_conn *ic, void *arg)
997 {
998 	struct ctl_scsiio *ctsio = arg;
999 
1000 	if (ctsio != NULL && ctsio->kern_data_len == ctsio->ext_data_filled) {
1001 		struct ppod_reservation *prsv;
1002 
1003 		prsv = io_to_ppod_reservation(ctsio);
1004 		MPASS(prsv != NULL);
1005 
1006 		t4_free_page_pods(prsv);
1007 		uma_zfree(prsv_zone, prsv);
1008 	}
1009 }
1010 
1011 static void
1012 cxgbei_limits(struct adapter *sc, void *arg)
1013 {
1014 	struct icl_drv_limits *idl = arg;
1015 	struct cxgbei_data *ci;
1016 	int max_dsl;
1017 
1018 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4lims") != 0)
1019 		return;
1020 
1021 	if (uld_active(sc, ULD_ISCSI)) {
1022 		ci = sc->iscsi_ulp_softc;
1023 		MPASS(ci != NULL);
1024 
1025 		/*
1026 		 * AHS is not supported by the kernel so we'll not account for
1027 		 * it either in our PDU len -> data segment len conversions.
1028 		 */
1029 
1030 		max_dsl = ci->max_rx_pdu_len - ISCSI_BHS_SIZE -
1031 		    ISCSI_HEADER_DIGEST_SIZE - ISCSI_DATA_DIGEST_SIZE;
1032 		if (idl->idl_max_recv_data_segment_length > max_dsl)
1033 			idl->idl_max_recv_data_segment_length = max_dsl;
1034 
1035 		max_dsl = ci->max_tx_pdu_len - ISCSI_BHS_SIZE -
1036 		    ISCSI_HEADER_DIGEST_SIZE - ISCSI_DATA_DIGEST_SIZE;
1037 		if (idl->idl_max_send_data_segment_length > max_dsl)
1038 			idl->idl_max_send_data_segment_length = max_dsl;
1039 	}
1040 
1041 	end_synchronized_op(sc, LOCK_HELD);
1042 }
1043 
1044 static int
1045 icl_cxgbei_limits(struct icl_drv_limits *idl)
1046 {
1047 
1048 	/* Maximum allowed by the RFC.  cxgbei_limits will clip them. */
1049 	idl->idl_max_recv_data_segment_length = (1 << 24) - 1;
1050 	idl->idl_max_send_data_segment_length = (1 << 24) - 1;
1051 
1052 	/* These are somewhat arbitrary. */
1053 	idl->idl_max_burst_length = 2 * 1024 * 1024;
1054 	idl->idl_first_burst_length = 8192;
1055 
1056 	t4_iterate(cxgbei_limits, idl);
1057 
1058 	return (0);
1059 }
1060 
1061 int
1062 icl_cxgbei_mod_load(void)
1063 {
1064 	int rc;
1065 
1066 	/*
1067 	 * Space to track pagepod reservations.
1068 	 */
1069 	prsv_zone = uma_zcreate("Pagepod reservations",
1070 	    sizeof(struct ppod_reservation), NULL, NULL, NULL, NULL,
1071 	    UMA_ALIGN_CACHE, 0);
1072 
1073 	refcount_init(&icl_cxgbei_ncons, 0);
1074 
1075 	rc = icl_register("cxgbei", false, -100, icl_cxgbei_limits,
1076 	    icl_cxgbei_new_conn);
1077 
1078 	return (rc);
1079 }
1080 
1081 int
1082 icl_cxgbei_mod_unload(void)
1083 {
1084 
1085 	if (icl_cxgbei_ncons != 0)
1086 		return (EBUSY);
1087 
1088 	icl_unregister("cxgbei", false);
1089 
1090 	uma_zdestroy(prsv_zone);
1091 
1092 	return (0);
1093 }
1094 #endif
1095