xref: /freebsd/sys/dev/cxgbe/cxgbei/icl_cxgbei.c (revision 3ac125068a211377f0b3817c37cf1db95a87e8fb)
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/lock.h>
51 #include <sys/mbuf.h>
52 #include <sys/mutex.h>
53 #include <sys/module.h>
54 #include <sys/protosw.h>
55 #include <sys/socket.h>
56 #include <sys/socketvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
59 #include <sys/sx.h>
60 #include <sys/uio.h>
61 #include <machine/bus.h>
62 #include <vm/uma.h>
63 #include <vm/vm.h>
64 #include <vm/pmap.h>
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp.h>
68 #include <netinet/tcp_var.h>
69 #include <netinet/toecore.h>
70 
71 #include <dev/iscsi/icl.h>
72 #include <dev/iscsi/iscsi_proto.h>
73 #include <icl_conn_if.h>
74 
75 #include <cam/scsi/scsi_all.h>
76 #include <cam/scsi/scsi_da.h>
77 #include <cam/ctl/ctl_io.h>
78 #include <cam/ctl/ctl.h>
79 #include <cam/ctl/ctl_backend.h>
80 #include <cam/ctl/ctl_error.h>
81 #include <cam/ctl/ctl_frontend.h>
82 #include <cam/ctl/ctl_debug.h>
83 #include <cam/ctl/ctl_ha.h>
84 #include <cam/ctl/ctl_ioctl.h>
85 
86 #include <cam/cam.h>
87 #include <cam/cam_ccb.h>
88 #include <cam/cam_xpt.h>
89 #include <cam/cam_debug.h>
90 #include <cam/cam_sim.h>
91 #include <cam/cam_xpt_sim.h>
92 #include <cam/cam_xpt_periph.h>
93 #include <cam/cam_periph.h>
94 #include <cam/cam_compat.h>
95 #include <cam/scsi/scsi_message.h>
96 
97 #include "common/common.h"
98 #include "common/t4_tcb.h"
99 #include "tom/t4_tom.h"
100 #include "cxgbei.h"
101 
102 SYSCTL_NODE(_kern_icl, OID_AUTO, cxgbei, CTLFLAG_RD, 0, "Chelsio iSCSI offload");
103 static int coalesce = 1;
104 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, coalesce, CTLFLAG_RWTUN,
105 	&coalesce, 0, "Try to coalesce PDUs before sending");
106 static int partial_receive_len = 128 * 1024;
107 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, partial_receive_len, CTLFLAG_RWTUN,
108     &partial_receive_len, 0, "Minimum read size for partially received "
109     "data segment");
110 static int sendspace = 1048576;
111 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, sendspace, CTLFLAG_RWTUN,
112     &sendspace, 0, "Default send socket buffer size");
113 static int recvspace = 1048576;
114 SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, recvspace, CTLFLAG_RWTUN,
115     &recvspace, 0, "Default receive socket buffer size");
116 
117 static uma_zone_t prsv_zone;
118 static volatile u_int icl_cxgbei_ncons;
119 
120 #define ICL_CONN_LOCK(X)		mtx_lock(X->ic_lock)
121 #define ICL_CONN_UNLOCK(X)		mtx_unlock(X->ic_lock)
122 #define ICL_CONN_LOCK_ASSERT(X)		mtx_assert(X->ic_lock, MA_OWNED)
123 #define ICL_CONN_LOCK_ASSERT_NOT(X)	mtx_assert(X->ic_lock, MA_NOTOWNED)
124 
125 struct icl_pdu *icl_cxgbei_new_pdu(int);
126 void icl_cxgbei_new_pdu_set_conn(struct icl_pdu *, struct icl_conn *);
127 
128 static icl_conn_new_pdu_t	icl_cxgbei_conn_new_pdu;
129 icl_conn_pdu_free_t	icl_cxgbei_conn_pdu_free;
130 static icl_conn_pdu_data_segment_length_t
131 				    icl_cxgbei_conn_pdu_data_segment_length;
132 static icl_conn_pdu_append_data_t	icl_cxgbei_conn_pdu_append_data;
133 static icl_conn_pdu_get_data_t	icl_cxgbei_conn_pdu_get_data;
134 static icl_conn_pdu_queue_t	icl_cxgbei_conn_pdu_queue;
135 static icl_conn_handoff_t	icl_cxgbei_conn_handoff;
136 static icl_conn_free_t		icl_cxgbei_conn_free;
137 static icl_conn_close_t		icl_cxgbei_conn_close;
138 static icl_conn_task_setup_t	icl_cxgbei_conn_task_setup;
139 static icl_conn_task_done_t	icl_cxgbei_conn_task_done;
140 static icl_conn_transfer_setup_t	icl_cxgbei_conn_transfer_setup;
141 static icl_conn_transfer_done_t	icl_cxgbei_conn_transfer_done;
142 
143 static kobj_method_t icl_cxgbei_methods[] = {
144 	KOBJMETHOD(icl_conn_new_pdu, icl_cxgbei_conn_new_pdu),
145 	KOBJMETHOD(icl_conn_pdu_free, icl_cxgbei_conn_pdu_free),
146 	KOBJMETHOD(icl_conn_pdu_data_segment_length,
147 	    icl_cxgbei_conn_pdu_data_segment_length),
148 	KOBJMETHOD(icl_conn_pdu_append_data, icl_cxgbei_conn_pdu_append_data),
149 	KOBJMETHOD(icl_conn_pdu_get_data, icl_cxgbei_conn_pdu_get_data),
150 	KOBJMETHOD(icl_conn_pdu_queue, icl_cxgbei_conn_pdu_queue),
151 	KOBJMETHOD(icl_conn_handoff, icl_cxgbei_conn_handoff),
152 	KOBJMETHOD(icl_conn_free, icl_cxgbei_conn_free),
153 	KOBJMETHOD(icl_conn_close, icl_cxgbei_conn_close),
154 	KOBJMETHOD(icl_conn_task_setup, icl_cxgbei_conn_task_setup),
155 	KOBJMETHOD(icl_conn_task_done, icl_cxgbei_conn_task_done),
156 	KOBJMETHOD(icl_conn_transfer_setup, icl_cxgbei_conn_transfer_setup),
157 	KOBJMETHOD(icl_conn_transfer_done, icl_cxgbei_conn_transfer_done),
158 	{ 0, 0 }
159 };
160 
161 DEFINE_CLASS(icl_cxgbei, icl_cxgbei_methods, sizeof(struct icl_cxgbei_conn));
162 
163 void
164 icl_cxgbei_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip)
165 {
166 #ifdef INVARIANTS
167 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
168 #endif
169 
170 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
171 	MPASS(ic == ip->ip_conn);
172 	MPASS(ip->ip_bhs_mbuf != NULL);
173 
174 	m_freem(ip->ip_ahs_mbuf);
175 	m_freem(ip->ip_data_mbuf);
176 	m_freem(ip->ip_bhs_mbuf);	/* storage for icl_cxgbei_pdu itself */
177 
178 #ifdef DIAGNOSTIC
179 	if (__predict_true(ic != NULL))
180 		refcount_release(&ic->ic_outstanding_pdus);
181 #endif
182 }
183 
184 struct icl_pdu *
185 icl_cxgbei_new_pdu(int flags)
186 {
187 	struct icl_cxgbei_pdu *icp;
188 	struct icl_pdu *ip;
189 	struct mbuf *m;
190 	uintptr_t a;
191 
192 	m = m_gethdr(flags, MT_DATA);
193 	if (__predict_false(m == NULL))
194 		return (NULL);
195 
196 	a = roundup2(mtod(m, uintptr_t), _Alignof(struct icl_cxgbei_pdu));
197 	icp = (struct icl_cxgbei_pdu *)a;
198 	bzero(icp, sizeof(*icp));
199 
200 	icp->icp_signature = CXGBEI_PDU_SIGNATURE;
201 	ip = &icp->ip;
202 	ip->ip_bhs_mbuf = m;
203 
204 	a = roundup2((uintptr_t)(icp + 1), _Alignof(struct iscsi_bhs *));
205 	ip->ip_bhs = (struct iscsi_bhs *)a;
206 #ifdef INVARIANTS
207 	/* Everything must fit entirely in the mbuf. */
208 	a = (uintptr_t)(ip->ip_bhs + 1);
209 	MPASS(a <= (uintptr_t)m + MSIZE);
210 #endif
211 	bzero(ip->ip_bhs, sizeof(*ip->ip_bhs));
212 
213 	m->m_data = (void *)ip->ip_bhs;
214 	m->m_len = sizeof(struct iscsi_bhs);
215 	m->m_pkthdr.len = m->m_len;
216 
217 	return (ip);
218 }
219 
220 void
221 icl_cxgbei_new_pdu_set_conn(struct icl_pdu *ip, struct icl_conn *ic)
222 {
223 
224 	ip->ip_conn = ic;
225 #ifdef DIAGNOSTIC
226 	refcount_acquire(&ic->ic_outstanding_pdus);
227 #endif
228 }
229 
230 /*
231  * Allocate icl_pdu with empty BHS to fill up by the caller.
232  */
233 static struct icl_pdu *
234 icl_cxgbei_conn_new_pdu(struct icl_conn *ic, int flags)
235 {
236 	struct icl_pdu *ip;
237 
238 	ip = icl_cxgbei_new_pdu(flags);
239 	if (__predict_false(ip == NULL))
240 		return (NULL);
241 	icl_cxgbei_new_pdu_set_conn(ip, ic);
242 
243 	return (ip);
244 }
245 
246 static size_t
247 icl_pdu_data_segment_length(const struct icl_pdu *request)
248 {
249 	uint32_t len = 0;
250 
251 	len += request->ip_bhs->bhs_data_segment_len[0];
252 	len <<= 8;
253 	len += request->ip_bhs->bhs_data_segment_len[1];
254 	len <<= 8;
255 	len += request->ip_bhs->bhs_data_segment_len[2];
256 
257 	return (len);
258 }
259 
260 size_t
261 icl_cxgbei_conn_pdu_data_segment_length(struct icl_conn *ic,
262     const struct icl_pdu *request)
263 {
264 
265 	return (icl_pdu_data_segment_length(request));
266 }
267 
268 static struct mbuf *
269 finalize_pdu(struct icl_cxgbei_conn *icc, struct icl_cxgbei_pdu *icp)
270 {
271 	struct icl_pdu *ip = &icp->ip;
272 	uint8_t ulp_submode, padding;
273 	struct mbuf *m, *last;
274 	struct iscsi_bhs *bhs;
275 
276 	/*
277 	 * Fix up the data segment mbuf first.
278 	 */
279 	m = ip->ip_data_mbuf;
280 	ulp_submode = icc->ulp_submode;
281 	if (m) {
282 		last = m_last(m);
283 
284 		/*
285 		 * Round up the data segment to a 4B boundary.  Pad with 0 if
286 		 * necessary.  There will definitely be room in the mbuf.
287 		 */
288 		padding = roundup2(ip->ip_data_len, 4) - ip->ip_data_len;
289 		if (padding) {
290 			bzero(mtod(last, uint8_t *) + last->m_len, padding);
291 			last->m_len += padding;
292 		}
293 	} else {
294 		MPASS(ip->ip_data_len == 0);
295 		ulp_submode &= ~ULP_CRC_DATA;
296 		padding = 0;
297 	}
298 
299 	/*
300 	 * Now the header mbuf that has the BHS.
301 	 */
302 	m = ip->ip_bhs_mbuf;
303 	MPASS(m->m_pkthdr.len == sizeof(struct iscsi_bhs));
304 	MPASS(m->m_len == sizeof(struct iscsi_bhs));
305 
306 	bhs = ip->ip_bhs;
307 	bhs->bhs_data_segment_len[2] = ip->ip_data_len;
308 	bhs->bhs_data_segment_len[1] = ip->ip_data_len >> 8;
309 	bhs->bhs_data_segment_len[0] = ip->ip_data_len >> 16;
310 
311 	/* "Convert" PDU to mbuf chain.  Do not use icp/ip after this. */
312 	m->m_pkthdr.len = sizeof(struct iscsi_bhs) + ip->ip_data_len + padding;
313 	m->m_next = ip->ip_data_mbuf;
314 	set_mbuf_ulp_submode(m, ulp_submode);
315 #ifdef INVARIANTS
316 	bzero(icp, sizeof(*icp));
317 #endif
318 #ifdef DIAGNOSTIC
319 	refcount_release(&icc->ic.ic_outstanding_pdus);
320 #endif
321 
322 	return (m);
323 }
324 
325 int
326 icl_cxgbei_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *ip,
327     const void *addr, size_t len, int flags)
328 {
329 	struct mbuf *m;
330 #ifdef INVARIANTS
331 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
332 #endif
333 
334 	MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
335 	MPASS(ic == ip->ip_conn);
336 	KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len));
337 
338 	m = ip->ip_data_mbuf;
339 	if (m == NULL) {
340 		m = m_getjcl(M_NOWAIT, MT_DATA, 0, MJUM16BYTES);
341 		if (__predict_false(m == NULL))
342 			return (ENOMEM);
343 
344 		ip->ip_data_mbuf = m;
345 	}
346 
347 	if (__predict_true(m_append(m, len, addr) != 0)) {
348 		ip->ip_data_len += len;
349 		MPASS(ip->ip_data_len <= ic->ic_max_data_segment_length);
350 		return (0);
351 	} else {
352 	    	if (flags & M_WAITOK) {
353 			CXGBE_UNIMPLEMENTED("fail safe append");
354 		}
355 		ip->ip_data_len = m_length(m, NULL);
356 		return (1);
357 	}
358 }
359 
360 void
361 icl_cxgbei_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip,
362     size_t off, void *addr, size_t len)
363 {
364 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
365 
366 	if (icp->icp_flags & ICPF_RX_DDP)
367 		return; /* data is DDP'ed, no need to copy */
368 	m_copydata(ip->ip_data_mbuf, off, len, addr);
369 }
370 
371 void
372 icl_cxgbei_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip)
373 {
374 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
375 	struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
376 	struct socket *so = ic->ic_socket;
377 	struct toepcb *toep = icc->toep;
378 	struct inpcb *inp;
379 	struct mbuf *m;
380 
381 	MPASS(ic == ip->ip_conn);
382 	MPASS(ip->ip_bhs_mbuf != NULL);
383 	/* The kernel doesn't generate PDUs with AHS. */
384 	MPASS(ip->ip_ahs_mbuf == NULL && ip->ip_ahs_len == 0);
385 
386 	ICL_CONN_LOCK_ASSERT(ic);
387 	/* NOTE: sowriteable without so_snd lock is a mostly harmless race. */
388 	if (ic->ic_disconnecting || so == NULL || !sowriteable(so)) {
389 		icl_cxgbei_conn_pdu_free(ic, ip);
390 		return;
391 	}
392 
393 	m = finalize_pdu(icc, icp);
394 	M_ASSERTPKTHDR(m);
395 	MPASS((m->m_pkthdr.len & 3) == 0);
396 
397 	/*
398 	 * Do not get inp from toep->inp as the toepcb might have detached
399 	 * already.
400 	 */
401 	inp = sotoinpcb(so);
402 	INP_WLOCK(inp);
403 	if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) ||
404 	    __predict_false((toep->flags & TPF_ATTACHED) == 0))
405 		m_freem(m);
406 	else {
407 		mbufq_enqueue(&toep->ulp_pduq, m);
408 		t4_push_pdus(icc->sc, toep, 0);
409 	}
410 	INP_WUNLOCK(inp);
411 }
412 
413 static struct icl_conn *
414 icl_cxgbei_new_conn(const char *name, struct mtx *lock)
415 {
416 	struct icl_cxgbei_conn *icc;
417 	struct icl_conn *ic;
418 
419 	refcount_acquire(&icl_cxgbei_ncons);
420 
421 	icc = (struct icl_cxgbei_conn *)kobj_create(&icl_cxgbei_class, M_CXGBE,
422 	    M_WAITOK | M_ZERO);
423 	icc->icc_signature = CXGBEI_CONN_SIGNATURE;
424 	STAILQ_INIT(&icc->rcvd_pdus);
425 
426 	ic = &icc->ic;
427 	ic->ic_lock = lock;
428 
429 	/* XXXNP: review.  Most of these icl_conn fields aren't really used */
430 	STAILQ_INIT(&ic->ic_to_send);
431 	cv_init(&ic->ic_send_cv, "icl_cxgbei_tx");
432 	cv_init(&ic->ic_receive_cv, "icl_cxgbei_rx");
433 #ifdef DIAGNOSTIC
434 	refcount_init(&ic->ic_outstanding_pdus, 0);
435 #endif
436 	/* This is a stop-gap value that will be corrected during handoff. */
437 	ic->ic_max_data_segment_length = 16384;
438 	ic->ic_name = name;
439 	ic->ic_offload = "cxgbei";
440 	ic->ic_unmapped = false;
441 
442 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
443 
444 	return (ic);
445 }
446 
447 void
448 icl_cxgbei_conn_free(struct icl_conn *ic)
449 {
450 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
451 
452 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
453 
454 	CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc);
455 
456 	cv_destroy(&ic->ic_send_cv);
457 	cv_destroy(&ic->ic_receive_cv);
458 
459 	kobj_delete((struct kobj *)icc, M_CXGBE);
460 	refcount_release(&icl_cxgbei_ncons);
461 }
462 
463 static int
464 icl_cxgbei_setsockopt(struct icl_conn *ic, struct socket *so, int sspace,
465     int rspace)
466 {
467 	struct sockopt opt;
468 	int error, one = 1, ss, rs;
469 
470 	ss = max(sendspace, sspace);
471 	rs = max(recvspace, rspace);
472 
473 	error = soreserve(so, ss, rs);
474 	if (error != 0) {
475 		icl_cxgbei_conn_close(ic);
476 		return (error);
477 	}
478 	SOCKBUF_LOCK(&so->so_snd);
479 	so->so_snd.sb_flags |= SB_AUTOSIZE;
480 	SOCKBUF_UNLOCK(&so->so_snd);
481 	SOCKBUF_LOCK(&so->so_rcv);
482 	so->so_rcv.sb_flags |= SB_AUTOSIZE;
483 	SOCKBUF_UNLOCK(&so->so_rcv);
484 
485 	/*
486 	 * Disable Nagle.
487 	 */
488 	bzero(&opt, sizeof(opt));
489 	opt.sopt_dir = SOPT_SET;
490 	opt.sopt_level = IPPROTO_TCP;
491 	opt.sopt_name = TCP_NODELAY;
492 	opt.sopt_val = &one;
493 	opt.sopt_valsize = sizeof(one);
494 	error = sosetopt(so, &opt);
495 	if (error != 0) {
496 		icl_cxgbei_conn_close(ic);
497 		return (error);
498 	}
499 
500 	return (0);
501 }
502 
503 /*
504  * Request/response structure used to find out the adapter offloading a socket.
505  */
506 struct find_ofld_adapter_rr {
507 	struct socket *so;
508 	struct adapter *sc;	/* result */
509 };
510 
511 static void
512 find_offload_adapter(struct adapter *sc, void *arg)
513 {
514 	struct find_ofld_adapter_rr *fa = arg;
515 	struct socket *so = fa->so;
516 	struct tom_data *td = sc->tom_softc;
517 	struct tcpcb *tp;
518 	struct inpcb *inp;
519 
520 	/* Non-TCP were filtered out earlier. */
521 	MPASS(so->so_proto->pr_protocol == IPPROTO_TCP);
522 
523 	if (fa->sc != NULL)
524 		return;	/* Found already. */
525 
526 	if (td == NULL)
527 		return;	/* TOE not enabled on this adapter. */
528 
529 	inp = sotoinpcb(so);
530 	INP_WLOCK(inp);
531 	if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) {
532 		tp = intotcpcb(inp);
533 		if (tp->t_flags & TF_TOE && tp->tod == &td->tod)
534 			fa->sc = sc;	/* Found. */
535 	}
536 	INP_WUNLOCK(inp);
537 }
538 
539 /* XXXNP: move this to t4_tom. */
540 static void
541 send_iscsi_flowc_wr(struct adapter *sc, struct toepcb *toep, int maxlen)
542 {
543 	struct wrqe *wr;
544 	struct fw_flowc_wr *flowc;
545 	const u_int nparams = 1;
546 	u_int flowclen;
547 	struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx];
548 
549 	flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval);
550 
551 	wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq);
552 	if (wr == NULL) {
553 		/* XXX */
554 		panic("%s: allocation failure.", __func__);
555 	}
556 	flowc = wrtod(wr);
557 	memset(flowc, 0, wr->wr_len);
558 
559 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
560 	    V_FW_FLOWC_WR_NPARAMS(nparams));
561 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) |
562 	    V_FW_WR_FLOWID(toep->tid));
563 
564 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_TXDATAPLEN_MAX;
565 	flowc->mnemval[0].val = htobe32(maxlen);
566 
567 	txsd->tx_credits = howmany(flowclen, 16);
568 	txsd->plen = 0;
569 	KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0,
570 	    ("%s: not enough credits (%d)", __func__, toep->tx_credits));
571 	toep->tx_credits -= txsd->tx_credits;
572 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
573 		toep->txsd_pidx = 0;
574 	toep->txsd_avail--;
575 
576         t4_wrq_tx(sc, wr);
577 }
578 
579 static void
580 set_ulp_mode_iscsi(struct adapter *sc, struct toepcb *toep, int hcrc, int dcrc)
581 {
582 	uint64_t val = ULP_MODE_ISCSI;
583 
584 	if (hcrc)
585 		val |= ULP_CRC_HEADER << 4;
586 	if (dcrc)
587 		val |= ULP_CRC_DATA << 4;
588 
589 	CTR4(KTR_CXGBE, "%s: tid %u, ULP_MODE_ISCSI, CRC hdr=%d data=%d",
590 	    __func__, toep->tid, hcrc, dcrc);
591 
592 	t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_ULP_TYPE,
593 	    V_TCB_ULP_TYPE(M_TCB_ULP_TYPE) | V_TCB_ULP_RAW(M_TCB_ULP_RAW), val,
594 	    0, 0, toep->ofld_rxq->iq.abs_id);
595 }
596 
597 /*
598  * XXXNP: Who is responsible for cleaning up the socket if this returns with an
599  * error?  Review all error paths.
600  *
601  * XXXNP: What happens to the socket's fd reference if the operation is
602  * successful, and how does that affect the socket's life cycle?
603  */
604 int
605 icl_cxgbei_conn_handoff(struct icl_conn *ic, int fd)
606 {
607 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
608 	struct cxgbei_data *ci;
609 	struct find_ofld_adapter_rr fa;
610 	struct file *fp;
611 	struct socket *so;
612 	struct inpcb *inp;
613 	struct tcpcb *tp;
614 	struct toepcb *toep;
615 	cap_rights_t rights;
616 	int error;
617 
618 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
619 	ICL_CONN_LOCK_ASSERT_NOT(ic);
620 
621 	/*
622 	 * Steal the socket from userland.
623 	 */
624 	error = fget(curthread, fd,
625 	    cap_rights_init(&rights, CAP_SOCK_CLIENT), &fp);
626 	if (error != 0)
627 		return (error);
628 	if (fp->f_type != DTYPE_SOCKET) {
629 		fdrop(fp, curthread);
630 		return (EINVAL);
631 	}
632 	so = fp->f_data;
633 	if (so->so_type != SOCK_STREAM ||
634 	    so->so_proto->pr_protocol != IPPROTO_TCP) {
635 		fdrop(fp, curthread);
636 		return (EINVAL);
637 	}
638 
639 	ICL_CONN_LOCK(ic);
640 	if (ic->ic_socket != NULL) {
641 		ICL_CONN_UNLOCK(ic);
642 		fdrop(fp, curthread);
643 		return (EBUSY);
644 	}
645 	ic->ic_disconnecting = false;
646 	ic->ic_socket = so;
647 	fp->f_ops = &badfileops;
648 	fp->f_data = NULL;
649 	fdrop(fp, curthread);
650 	ICL_CONN_UNLOCK(ic);
651 
652 	/* Find the adapter offloading this socket. */
653 	fa.sc = NULL;
654 	fa.so = so;
655 	t4_iterate(find_offload_adapter, &fa);
656 	if (fa.sc == NULL)
657 		return (EINVAL);
658 	icc->sc = fa.sc;
659 	ci = icc->sc->iscsi_ulp_softc;
660 
661 	inp = sotoinpcb(so);
662 	INP_WLOCK(inp);
663 	tp = intotcpcb(inp);
664 	if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))
665 		error = EBUSY;
666 	else {
667 		/*
668 		 * socket could not have been "unoffloaded" if here.
669 		 */
670 		MPASS(tp->t_flags & TF_TOE);
671 		MPASS(tp->tod != NULL);
672 		MPASS(tp->t_toe != NULL);
673 		toep = tp->t_toe;
674 		MPASS(toep->vi->pi->adapter == icc->sc);
675 		icc->toep = toep;
676 		icc->cwt = cxgbei_select_worker_thread(icc);
677 
678 		/*
679 		 * We maintain the _send_ DSL in this field just to have a
680 		 * convenient way to assert that the kernel never sends
681 		 * oversized PDUs.  This field is otherwise unused in the driver
682 		 * or the kernel.
683 		 */
684 		ic->ic_max_data_segment_length = ci->max_tx_pdu_len -
685 		    ISCSI_BHS_SIZE;
686 
687 		icc->ulp_submode = 0;
688 		if (ic->ic_header_crc32c) {
689 			icc->ulp_submode |= ULP_CRC_HEADER;
690 			ic->ic_max_data_segment_length -=
691 			    ISCSI_HEADER_DIGEST_SIZE;
692 		}
693 		if (ic->ic_data_crc32c) {
694 			icc->ulp_submode |= ULP_CRC_DATA;
695 			ic->ic_max_data_segment_length -=
696 			    ISCSI_DATA_DIGEST_SIZE;
697 		}
698 		so->so_options |= SO_NO_DDP;
699 		toep->ulp_mode = ULP_MODE_ISCSI;
700 		toep->ulpcb = icc;
701 
702 		send_iscsi_flowc_wr(icc->sc, toep, ci->max_tx_pdu_len);
703 		set_ulp_mode_iscsi(icc->sc, toep, ic->ic_header_crc32c,
704 		    ic->ic_data_crc32c);
705 		error = 0;
706 	}
707 	INP_WUNLOCK(inp);
708 
709 	if (error == 0) {
710 		error = icl_cxgbei_setsockopt(ic, so, ci->max_tx_pdu_len,
711 		    ci->max_rx_pdu_len);
712 	}
713 
714 	return (error);
715 }
716 
717 void
718 icl_cxgbei_conn_close(struct icl_conn *ic)
719 {
720 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
721 	struct icl_pdu *ip;
722 	struct socket *so;
723 	struct sockbuf *sb;
724 	struct inpcb *inp;
725 	struct toepcb *toep = icc->toep;
726 
727 	MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE);
728 	ICL_CONN_LOCK_ASSERT_NOT(ic);
729 
730 	ICL_CONN_LOCK(ic);
731 	so = ic->ic_socket;
732 	if (ic->ic_disconnecting || so == NULL) {
733 		CTR4(KTR_CXGBE, "%s: icc %p (disconnecting = %d), so %p",
734 		    __func__, icc, ic->ic_disconnecting, so);
735 		ICL_CONN_UNLOCK(ic);
736 		return;
737 	}
738 	ic->ic_disconnecting = true;
739 
740 	/* These are unused in this driver right now. */
741 	MPASS(STAILQ_EMPTY(&ic->ic_to_send));
742 	MPASS(ic->ic_receive_pdu == NULL);
743 
744 #ifdef DIAGNOSTIC
745 	KASSERT(ic->ic_outstanding_pdus == 0,
746 	    ("destroying session with %d outstanding PDUs",
747 	     ic->ic_outstanding_pdus));
748 #endif
749 	ICL_CONN_UNLOCK(ic);
750 
751 	CTR3(KTR_CXGBE, "%s: tid %d, icc %p", __func__, toep ? toep->tid : -1,
752 	    icc);
753 	inp = sotoinpcb(so);
754 	sb = &so->so_rcv;
755 	INP_WLOCK(inp);
756 	if (toep != NULL) {	/* NULL if connection was never offloaded. */
757 		toep->ulpcb = NULL;
758 		mbufq_drain(&toep->ulp_pduq);
759 		SOCKBUF_LOCK(sb);
760 		if (icc->rx_flags & RXF_ACTIVE) {
761 			volatile u_int *p = &icc->rx_flags;
762 
763 			SOCKBUF_UNLOCK(sb);
764 			INP_WUNLOCK(inp);
765 
766 			while (*p & RXF_ACTIVE)
767 				pause("conclo", 1);
768 
769 			INP_WLOCK(inp);
770 			SOCKBUF_LOCK(sb);
771 		}
772 
773 		while (!STAILQ_EMPTY(&icc->rcvd_pdus)) {
774 			ip = STAILQ_FIRST(&icc->rcvd_pdus);
775 			STAILQ_REMOVE_HEAD(&icc->rcvd_pdus, ip_next);
776 			icl_cxgbei_conn_pdu_free(ic, ip);
777 		}
778 		SOCKBUF_UNLOCK(sb);
779 	}
780 	INP_WUNLOCK(inp);
781 
782 	ICL_CONN_LOCK(ic);
783 	ic->ic_socket = NULL;
784 	ICL_CONN_UNLOCK(ic);
785 
786 	/*
787 	 * XXXNP: we should send RST instead of FIN when PDUs held in various
788 	 * queues were purged instead of delivered reliably but soabort isn't
789 	 * really general purpose and wouldn't do the right thing here.
790 	 */
791 	soclose(so);
792 }
793 
794 int
795 icl_cxgbei_conn_task_setup(struct icl_conn *ic, struct icl_pdu *ip,
796     struct ccb_scsiio *csio, uint32_t *ittp, void **arg)
797 {
798 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
799 	struct toepcb *toep = icc->toep;
800 	struct adapter *sc = icc->sc;
801 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
802 	struct ppod_region *pr = &ci->pr;
803 	struct ppod_reservation *prsv;
804 	uint32_t itt;
805 	int rc = 0;
806 
807 	/* This is for the offload driver's state.  Must not be set already. */
808 	MPASS(arg != NULL);
809 	MPASS(*arg == NULL);
810 
811 	if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_IN ||
812 	    csio->dxfer_len < ci->ddp_threshold) {
813 no_ddp:
814 		/*
815 		 * No DDP for this I/O.  Allocate an ITT (based on the one
816 		 * passed in) that cannot be a valid hardware DDP tag in the
817 		 * iSCSI region.
818 		 */
819 		itt = *ittp & M_PPOD_TAG;
820 		itt = V_PPOD_TAG(itt) | pr->pr_invalid_bit;
821 		*ittp = htobe32(itt);
822 		MPASS(*arg == NULL);	/* State is maintained for DDP only. */
823 		if (rc != 0)
824 			counter_u64_add(ci->ddp_setup_error, 1);
825 		return (0);
826 	}
827 
828 	/*
829 	 * Reserve resources for DDP, update the itt that should be used in the
830 	 * PDU, and save DDP specific state for this I/O in *arg.
831 	 */
832 
833 	prsv = uma_zalloc(prsv_zone, M_NOWAIT);
834 	if (prsv == NULL) {
835 		rc = ENOMEM;
836 		goto no_ddp;
837 	}
838 
839 	/* XXX add support for all CAM_DATA_ types */
840 	MPASS((csio->ccb_h.flags & CAM_DATA_MASK) == CAM_DATA_VADDR);
841 	rc = t4_alloc_page_pods_for_buf(pr, (vm_offset_t)csio->data_ptr,
842 	    csio->dxfer_len, prsv);
843 	if (rc != 0) {
844 		uma_zfree(prsv_zone, prsv);
845 		goto no_ddp;
846 	}
847 
848 	rc = t4_write_page_pods_for_buf(sc, toep->ofld_txq, toep->tid, prsv,
849 	    (vm_offset_t)csio->data_ptr, csio->dxfer_len);
850 	if (rc != 0) {
851 		t4_free_page_pods(prsv);
852 		uma_zfree(prsv_zone, prsv);
853 		goto no_ddp;
854 	}
855 
856 	*ittp = htobe32(prsv->prsv_tag);
857 	*arg = prsv;
858 	counter_u64_add(ci->ddp_setup_ok, 1);
859 	return (0);
860 }
861 
862 void
863 icl_cxgbei_conn_task_done(struct icl_conn *ic, void *arg)
864 {
865 
866 	if (arg != NULL) {
867 		struct ppod_reservation *prsv = arg;
868 
869 		t4_free_page_pods(prsv);
870 		uma_zfree(prsv_zone, prsv);
871 	}
872 }
873 
874 /* XXXNP: PDU should be passed in as parameter, like on the initiator. */
875 #define io_to_request_pdu(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND].ptr)
876 #define io_to_ppod_reservation(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND2].ptr)
877 
878 int
879 icl_cxgbei_conn_transfer_setup(struct icl_conn *ic, union ctl_io *io,
880     uint32_t *tttp, void **arg)
881 {
882 	struct icl_cxgbei_conn *icc = ic_to_icc(ic);
883 	struct toepcb *toep = icc->toep;
884 	struct ctl_scsiio *ctsio = &io->scsiio;
885 	struct adapter *sc = icc->sc;
886 	struct cxgbei_data *ci = sc->iscsi_ulp_softc;
887 	struct ppod_region *pr = &ci->pr;
888 	struct ppod_reservation *prsv;
889 	uint32_t ttt;
890 	int xferlen, rc = 0, alias;
891 
892 	/* This is for the offload driver's state.  Must not be set already. */
893 	MPASS(arg != NULL);
894 	MPASS(*arg == NULL);
895 
896 	if (ctsio->ext_data_filled == 0) {
897 		int first_burst;
898 		struct icl_pdu *ip = io_to_request_pdu(io);
899 		vm_offset_t buf;
900 #ifdef INVARIANTS
901 		struct icl_cxgbei_pdu *icp = ip_to_icp(ip);
902 
903 		MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE);
904 		MPASS(ic == ip->ip_conn);
905 		MPASS(ip->ip_bhs_mbuf != NULL);
906 #endif
907 		first_burst = icl_pdu_data_segment_length(ip);
908 
909 		/*
910 		 * Note that ICL calls conn_transfer_setup even if the first
911 		 * burst had everything and there's nothing left to transfer.
912 		 */
913 		MPASS(ctsio->kern_data_len >= first_burst);
914 		xferlen = ctsio->kern_data_len;
915 		if (xferlen - first_burst < ci->ddp_threshold) {
916 no_ddp:
917 			/*
918 			 * No DDP for this transfer.  Allocate a TTT (based on
919 			 * the one passed in) that cannot be a valid hardware
920 			 * DDP tag in the iSCSI region.
921 			 */
922 			ttt = *tttp & M_PPOD_TAG;
923 			ttt = V_PPOD_TAG(ttt) | pr->pr_invalid_bit;
924 			*tttp = htobe32(ttt);
925 			MPASS(io_to_ppod_reservation(io) == NULL);
926 			if (rc != 0)
927 				counter_u64_add(ci->ddp_setup_error, 1);
928 			return (0);
929 		}
930 
931 		if (ctsio->kern_sg_entries == 0)
932 			buf = (vm_offset_t)ctsio->kern_data_ptr;
933 		else if (ctsio->kern_sg_entries == 1) {
934 			struct ctl_sg_entry *sgl = (void *)ctsio->kern_data_ptr;
935 
936 			MPASS(sgl->len == xferlen);
937 			buf = (vm_offset_t)sgl->addr;
938 		} else {
939 			rc = EAGAIN;	/* XXX implement */
940 			goto no_ddp;
941 		}
942 
943 
944 		/*
945 		 * Reserve resources for DDP, update the ttt that should be used
946 		 * in the PDU, and save DDP specific state for this I/O.
947 		 */
948 
949 		MPASS(io_to_ppod_reservation(io) == NULL);
950 		prsv = uma_zalloc(prsv_zone, M_NOWAIT);
951 		if (prsv == NULL) {
952 			rc = ENOMEM;
953 			goto no_ddp;
954 		}
955 
956 		rc = t4_alloc_page_pods_for_buf(pr, buf, xferlen, prsv);
957 		if (rc != 0) {
958 			uma_zfree(prsv_zone, prsv);
959 			goto no_ddp;
960 		}
961 
962 		rc = t4_write_page_pods_for_buf(sc, toep->ofld_txq, toep->tid,
963 		    prsv, buf, xferlen);
964 		if (rc != 0) {
965 			t4_free_page_pods(prsv);
966 			uma_zfree(prsv_zone, prsv);
967 			goto no_ddp;
968 		}
969 
970 		*tttp = htobe32(prsv->prsv_tag);
971 		io_to_ppod_reservation(io) = prsv;
972 		*arg = ctsio;
973 		counter_u64_add(ci->ddp_setup_ok, 1);
974 		return (0);
975 	}
976 
977 	/*
978 	 * In the middle of an I/O.  A non-NULL page pod reservation indicates
979 	 * that a DDP buffer is being used for the I/O.
980 	 */
981 
982 	prsv = io_to_ppod_reservation(ctsio);
983 	if (prsv == NULL)
984 		goto no_ddp;
985 
986 	alias = (prsv->prsv_tag & pr->pr_alias_mask) >> pr->pr_alias_shift;
987 	alias++;
988 	prsv->prsv_tag &= ~pr->pr_alias_mask;
989 	prsv->prsv_tag |= alias << pr->pr_alias_shift & pr->pr_alias_mask;
990 
991 	*tttp = htobe32(prsv->prsv_tag);
992 	*arg = ctsio;
993 
994 	return (0);
995 }
996 
997 void
998 icl_cxgbei_conn_transfer_done(struct icl_conn *ic, void *arg)
999 {
1000 	struct ctl_scsiio *ctsio = arg;
1001 
1002 	if (ctsio != NULL && ctsio->kern_data_len == ctsio->ext_data_filled) {
1003 		struct ppod_reservation *prsv;
1004 
1005 		prsv = io_to_ppod_reservation(ctsio);
1006 		MPASS(prsv != NULL);
1007 
1008 		t4_free_page_pods(prsv);
1009 		uma_zfree(prsv_zone, prsv);
1010 	}
1011 }
1012 
1013 static void
1014 cxgbei_limits(struct adapter *sc, void *arg)
1015 {
1016 	struct icl_drv_limits *idl = arg;
1017 	struct cxgbei_data *ci;
1018 	int max_dsl;
1019 
1020 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4lims") != 0)
1021 		return;
1022 
1023 	if (uld_active(sc, ULD_ISCSI)) {
1024 		ci = sc->iscsi_ulp_softc;
1025 		MPASS(ci != NULL);
1026 
1027 		/*
1028 		 * AHS is not supported by the kernel so we'll not account for
1029 		 * it either in our PDU len -> data segment len conversions.
1030 		 */
1031 
1032 		max_dsl = ci->max_rx_pdu_len - ISCSI_BHS_SIZE -
1033 		    ISCSI_HEADER_DIGEST_SIZE - ISCSI_DATA_DIGEST_SIZE;
1034 		if (idl->idl_max_recv_data_segment_length > max_dsl)
1035 			idl->idl_max_recv_data_segment_length = max_dsl;
1036 
1037 		max_dsl = ci->max_tx_pdu_len - ISCSI_BHS_SIZE -
1038 		    ISCSI_HEADER_DIGEST_SIZE - ISCSI_DATA_DIGEST_SIZE;
1039 		if (idl->idl_max_send_data_segment_length > max_dsl)
1040 			idl->idl_max_send_data_segment_length = max_dsl;
1041 	}
1042 
1043 	end_synchronized_op(sc, LOCK_HELD);
1044 }
1045 
1046 static int
1047 icl_cxgbei_limits(struct icl_drv_limits *idl)
1048 {
1049 
1050 	/* Maximum allowed by the RFC.  cxgbei_limits will clip them. */
1051 	idl->idl_max_recv_data_segment_length = (1 << 24) - 1;
1052 	idl->idl_max_send_data_segment_length = (1 << 24) - 1;
1053 
1054 	/* These are somewhat arbitrary. */
1055 	idl->idl_max_burst_length = 2 * 1024 * 1024;
1056 	idl->idl_first_burst_length = 8192;
1057 
1058 	t4_iterate(cxgbei_limits, idl);
1059 
1060 	return (0);
1061 }
1062 
1063 int
1064 icl_cxgbei_mod_load(void)
1065 {
1066 	int rc;
1067 
1068 	/*
1069 	 * Space to track pagepod reservations.
1070 	 */
1071 	prsv_zone = uma_zcreate("Pagepod reservations",
1072 	    sizeof(struct ppod_reservation), NULL, NULL, NULL, NULL,
1073 	    CACHE_LINE_SIZE, 0);
1074 
1075 	refcount_init(&icl_cxgbei_ncons, 0);
1076 
1077 	rc = icl_register("cxgbei", false, -100, icl_cxgbei_limits,
1078 	    icl_cxgbei_new_conn);
1079 
1080 	return (rc);
1081 }
1082 
1083 int
1084 icl_cxgbei_mod_unload(void)
1085 {
1086 
1087 	if (icl_cxgbei_ncons != 0)
1088 		return (EBUSY);
1089 
1090 	icl_unregister("cxgbei", false);
1091 
1092 	uma_zdestroy(prsv_zone);
1093 
1094 	return (0);
1095 }
1096 #endif
1097