xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision 99429157e8615dc3b7f11afbe3ed92de7476a5db)
1 /*-
2  * Copyright (c) 2012 Chelsio Communications, Inc.
3  * All rights reserved.
4  * Written by: Navdeep Parhar <np@FreeBSD.org>
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_ratelimit.h"
34 
35 #include <sys/param.h>
36 #include <sys/types.h>
37 #include <sys/systm.h>
38 #include <sys/kernel.h>
39 #include <sys/ktr.h>
40 #include <sys/lock.h>
41 #include <sys/limits.h>
42 #include <sys/module.h>
43 #include <sys/protosw.h>
44 #include <sys/domain.h>
45 #include <sys/refcount.h>
46 #include <sys/rmlock.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/taskqueue.h>
50 #include <net/if.h>
51 #include <net/if_var.h>
52 #include <netinet/in.h>
53 #include <netinet/in_pcb.h>
54 #include <netinet/in_var.h>
55 #include <netinet/ip.h>
56 #include <netinet/ip6.h>
57 #include <netinet6/scope6_var.h>
58 #define TCPSTATES
59 #include <netinet/tcp_fsm.h>
60 #include <netinet/tcp_var.h>
61 #include <netinet/toecore.h>
62 
63 #ifdef TCP_OFFLOAD
64 #include "common/common.h"
65 #include "common/t4_msg.h"
66 #include "common/t4_regs.h"
67 #include "common/t4_regs_values.h"
68 #include "common/t4_tcb.h"
69 #include "tom/t4_tom_l2t.h"
70 #include "tom/t4_tom.h"
71 
72 static struct protosw toe_protosw;
73 static struct pr_usrreqs toe_usrreqs;
74 
75 static struct protosw toe6_protosw;
76 static struct pr_usrreqs toe6_usrreqs;
77 
78 /* Module ops */
79 static int t4_tom_mod_load(void);
80 static int t4_tom_mod_unload(void);
81 static int t4_tom_modevent(module_t, int, void *);
82 
83 /* ULD ops and helpers */
84 static int t4_tom_activate(struct adapter *);
85 static int t4_tom_deactivate(struct adapter *);
86 
87 static struct uld_info tom_uld_info = {
88 	.uld_id = ULD_TOM,
89 	.activate = t4_tom_activate,
90 	.deactivate = t4_tom_deactivate,
91 };
92 
93 static void queue_tid_release(struct adapter *, int);
94 static void release_offload_resources(struct toepcb *);
95 static int alloc_tid_tabs(struct tid_info *);
96 static void free_tid_tabs(struct tid_info *);
97 static int add_lip(struct adapter *, struct in6_addr *);
98 static int delete_lip(struct adapter *, struct in6_addr *);
99 static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *);
100 static void init_clip_table(struct adapter *, struct tom_data *);
101 static void update_clip(struct adapter *, void *);
102 static void t4_clip_task(void *, int);
103 static void update_clip_table(struct adapter *, struct tom_data *);
104 static void destroy_clip_table(struct adapter *, struct tom_data *);
105 static void free_tom_data(struct adapter *, struct tom_data *);
106 static void reclaim_wr_resources(void *, int);
107 
108 static int in6_ifaddr_gen;
109 static eventhandler_tag ifaddr_evhandler;
110 static struct timeout_task clip_task;
111 
112 struct toepcb *
113 alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags)
114 {
115 	struct port_info *pi = vi->pi;
116 	struct adapter *sc = pi->adapter;
117 	struct toepcb *toep;
118 	int tx_credits, txsd_total, len;
119 
120 	/*
121 	 * The firmware counts tx work request credits in units of 16 bytes
122 	 * each.  Reserve room for an ABORT_REQ so the driver never has to worry
123 	 * about tx credits if it wants to abort a connection.
124 	 */
125 	tx_credits = sc->params.ofldq_wr_cred;
126 	tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
127 
128 	/*
129 	 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
130 	 * immediate payload, and firmware counts tx work request credits in
131 	 * units of 16 byte.  Calculate the maximum work requests possible.
132 	 */
133 	txsd_total = tx_credits /
134 	    howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
135 
136 	if (txqid < 0)
137 		txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq;
138 	KASSERT(txqid >= vi->first_ofld_txq &&
139 	    txqid < vi->first_ofld_txq + vi->nofldtxq,
140 	    ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi,
141 		vi->first_ofld_txq, vi->nofldtxq));
142 
143 	if (rxqid < 0)
144 		rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq;
145 	KASSERT(rxqid >= vi->first_ofld_rxq &&
146 	    rxqid < vi->first_ofld_rxq + vi->nofldrxq,
147 	    ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi,
148 		vi->first_ofld_rxq, vi->nofldrxq));
149 
150 	len = offsetof(struct toepcb, txsd) +
151 	    txsd_total * sizeof(struct ofld_tx_sdesc);
152 
153 	toep = malloc(len, M_CXGBE, M_ZERO | flags);
154 	if (toep == NULL)
155 		return (NULL);
156 
157 	refcount_init(&toep->refcount, 1);
158 	toep->td = sc->tom_softc;
159 	toep->vi = vi;
160 	toep->tc_idx = -1;
161 	toep->tx_total = tx_credits;
162 	toep->tx_credits = tx_credits;
163 	toep->ofld_txq = &sc->sge.ofld_txq[txqid];
164 	toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid];
165 	toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
166 	mbufq_init(&toep->ulp_pduq, INT_MAX);
167 	mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
168 	toep->txsd_total = txsd_total;
169 	toep->txsd_avail = txsd_total;
170 	toep->txsd_pidx = 0;
171 	toep->txsd_cidx = 0;
172 	aiotx_init_toep(toep);
173 	ddp_init_toep(toep);
174 
175 	return (toep);
176 }
177 
178 struct toepcb *
179 hold_toepcb(struct toepcb *toep)
180 {
181 
182 	refcount_acquire(&toep->refcount);
183 	return (toep);
184 }
185 
186 void
187 free_toepcb(struct toepcb *toep)
188 {
189 
190 	if (refcount_release(&toep->refcount) == 0)
191 		return;
192 
193 	KASSERT(!(toep->flags & TPF_ATTACHED),
194 	    ("%s: attached to an inpcb", __func__));
195 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
196 	    ("%s: CPL pending", __func__));
197 
198 	ddp_uninit_toep(toep);
199 	free(toep, M_CXGBE);
200 }
201 
202 /*
203  * Set up the socket for TCP offload.
204  */
205 void
206 offload_socket(struct socket *so, struct toepcb *toep)
207 {
208 	struct tom_data *td = toep->td;
209 	struct inpcb *inp = sotoinpcb(so);
210 	struct tcpcb *tp = intotcpcb(inp);
211 	struct sockbuf *sb;
212 
213 	INP_WLOCK_ASSERT(inp);
214 
215 	/* Update socket */
216 	sb = &so->so_snd;
217 	SOCKBUF_LOCK(sb);
218 	sb->sb_flags |= SB_NOCOALESCE;
219 	SOCKBUF_UNLOCK(sb);
220 	sb = &so->so_rcv;
221 	SOCKBUF_LOCK(sb);
222 	sb->sb_flags |= SB_NOCOALESCE;
223 	if (inp->inp_vflag & INP_IPV6)
224 		so->so_proto = &toe6_protosw;
225 	else
226 		so->so_proto = &toe_protosw;
227 	SOCKBUF_UNLOCK(sb);
228 
229 	/* Update TCP PCB */
230 	tp->tod = &td->tod;
231 	tp->t_toe = toep;
232 	tp->t_flags |= TF_TOE;
233 
234 	/* Install an extra hold on inp */
235 	toep->inp = inp;
236 	toep->flags |= TPF_ATTACHED;
237 	in_pcbref(inp);
238 
239 	/* Add the TOE PCB to the active list */
240 	mtx_lock(&td->toep_list_lock);
241 	TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
242 	mtx_unlock(&td->toep_list_lock);
243 }
244 
245 /* This is _not_ the normal way to "unoffload" a socket. */
246 void
247 undo_offload_socket(struct socket *so)
248 {
249 	struct inpcb *inp = sotoinpcb(so);
250 	struct tcpcb *tp = intotcpcb(inp);
251 	struct toepcb *toep = tp->t_toe;
252 	struct tom_data *td = toep->td;
253 	struct sockbuf *sb;
254 
255 	INP_WLOCK_ASSERT(inp);
256 
257 	sb = &so->so_snd;
258 	SOCKBUF_LOCK(sb);
259 	sb->sb_flags &= ~SB_NOCOALESCE;
260 	SOCKBUF_UNLOCK(sb);
261 	sb = &so->so_rcv;
262 	SOCKBUF_LOCK(sb);
263 	sb->sb_flags &= ~SB_NOCOALESCE;
264 	SOCKBUF_UNLOCK(sb);
265 
266 	tp->tod = NULL;
267 	tp->t_toe = NULL;
268 	tp->t_flags &= ~TF_TOE;
269 
270 	toep->inp = NULL;
271 	toep->flags &= ~TPF_ATTACHED;
272 	if (in_pcbrele_wlocked(inp))
273 		panic("%s: inp freed.", __func__);
274 
275 	mtx_lock(&td->toep_list_lock);
276 	TAILQ_REMOVE(&td->toep_list, toep, link);
277 	mtx_unlock(&td->toep_list_lock);
278 }
279 
280 static void
281 release_offload_resources(struct toepcb *toep)
282 {
283 	struct tom_data *td = toep->td;
284 	struct adapter *sc = td_adapter(td);
285 	int tid = toep->tid;
286 
287 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
288 	    ("%s: %p has CPL pending.", __func__, toep));
289 	KASSERT(!(toep->flags & TPF_ATTACHED),
290 	    ("%s: %p is still attached.", __func__, toep));
291 
292 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
293 	    __func__, toep, tid, toep->l2te, toep->ce);
294 
295 	/*
296 	 * These queues should have been emptied at approximately the same time
297 	 * that a normal connection's socket's so_snd would have been purged or
298 	 * drained.  Do _not_ clean up here.
299 	 */
300 	MPASS(mbufq_len(&toep->ulp_pduq) == 0);
301 	MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0);
302 #ifdef INVARIANTS
303 	ddp_assert_empty(toep);
304 #endif
305 
306 	if (toep->l2te)
307 		t4_l2t_release(toep->l2te);
308 
309 	if (tid >= 0) {
310 		remove_tid(sc, tid, toep->ce ? 2 : 1);
311 		release_tid(sc, tid, toep->ctrlq);
312 	}
313 
314 	if (toep->ce)
315 		release_lip(td, toep->ce);
316 
317 #ifdef RATELIMIT
318 	if (toep->tc_idx != -1)
319 		t4_release_cl_rl_kbps(sc, toep->vi->pi->port_id, toep->tc_idx);
320 #endif
321 	mtx_lock(&td->toep_list_lock);
322 	TAILQ_REMOVE(&td->toep_list, toep, link);
323 	mtx_unlock(&td->toep_list_lock);
324 
325 	free_toepcb(toep);
326 }
327 
328 /*
329  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
330  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
331  * pending CPL) then it is time to release all resources tied to the toepcb.
332  *
333  * Also gets called when an offloaded active open fails and the TOM wants the
334  * kernel to take the TCP PCB back.
335  */
336 static void
337 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
338 {
339 #if defined(KTR) || defined(INVARIANTS)
340 	struct inpcb *inp = tp->t_inpcb;
341 #endif
342 	struct toepcb *toep = tp->t_toe;
343 
344 	INP_WLOCK_ASSERT(inp);
345 
346 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
347 	KASSERT(toep->flags & TPF_ATTACHED,
348 	    ("%s: not attached", __func__));
349 
350 #ifdef KTR
351 	if (tp->t_state == TCPS_SYN_SENT) {
352 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
353 		    __func__, toep->tid, toep, toep->flags, inp,
354 		    inp->inp_flags);
355 	} else {
356 		CTR6(KTR_CXGBE,
357 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
358 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
359 		    inp->inp_flags);
360 	}
361 #endif
362 
363 	tp->t_toe = NULL;
364 	tp->t_flags &= ~TF_TOE;
365 	toep->flags &= ~TPF_ATTACHED;
366 
367 	if (!(toep->flags & TPF_CPL_PENDING))
368 		release_offload_resources(toep);
369 }
370 
371 /*
372  * setsockopt handler.
373  */
374 static void
375 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
376 {
377 	struct adapter *sc = tod->tod_softc;
378 	struct toepcb *toep = tp->t_toe;
379 
380 	if (dir == SOPT_GET)
381 		return;
382 
383 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
384 
385 	switch (name) {
386 	case TCP_NODELAY:
387 		t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS,
388 		    V_TF_NAGLE(1), V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1),
389 		    0, 0, toep->ofld_rxq->iq.abs_id);
390 		break;
391 	default:
392 		break;
393 	}
394 }
395 
396 /*
397  * The TOE driver will not receive any more CPLs for the tid associated with the
398  * toepcb; release the hold on the inpcb.
399  */
400 void
401 final_cpl_received(struct toepcb *toep)
402 {
403 	struct inpcb *inp = toep->inp;
404 
405 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
406 	INP_WLOCK_ASSERT(inp);
407 	KASSERT(toep->flags & TPF_CPL_PENDING,
408 	    ("%s: CPL not pending already?", __func__));
409 
410 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
411 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
412 
413 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
414 		release_ddp_resources(toep);
415 	toep->inp = NULL;
416 	toep->flags &= ~TPF_CPL_PENDING;
417 	mbufq_drain(&toep->ulp_pdu_reclaimq);
418 
419 	if (!(toep->flags & TPF_ATTACHED))
420 		release_offload_resources(toep);
421 
422 	if (!in_pcbrele_wlocked(inp))
423 		INP_WUNLOCK(inp);
424 }
425 
426 void
427 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
428 {
429 	struct tid_info *t = &sc->tids;
430 
431 	t->tid_tab[tid] = ctx;
432 	atomic_add_int(&t->tids_in_use, ntids);
433 }
434 
435 void *
436 lookup_tid(struct adapter *sc, int tid)
437 {
438 	struct tid_info *t = &sc->tids;
439 
440 	return (t->tid_tab[tid]);
441 }
442 
443 void
444 update_tid(struct adapter *sc, int tid, void *ctx)
445 {
446 	struct tid_info *t = &sc->tids;
447 
448 	t->tid_tab[tid] = ctx;
449 }
450 
451 void
452 remove_tid(struct adapter *sc, int tid, int ntids)
453 {
454 	struct tid_info *t = &sc->tids;
455 
456 	t->tid_tab[tid] = NULL;
457 	atomic_subtract_int(&t->tids_in_use, ntids);
458 }
459 
460 void
461 release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq)
462 {
463 	struct wrqe *wr;
464 	struct cpl_tid_release *req;
465 
466 	wr = alloc_wrqe(sizeof(*req), ctrlq);
467 	if (wr == NULL) {
468 		queue_tid_release(sc, tid);	/* defer */
469 		return;
470 	}
471 	req = wrtod(wr);
472 
473 	INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid);
474 
475 	t4_wrq_tx(sc, wr);
476 }
477 
478 static void
479 queue_tid_release(struct adapter *sc, int tid)
480 {
481 
482 	CXGBE_UNIMPLEMENTED("deferred tid release");
483 }
484 
485 /*
486  * What mtu_idx to use, given a 4-tuple and/or an MSS cap
487  */
488 int
489 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss)
490 {
491 	unsigned short *mtus = &sc->params.mtus[0];
492 	int i, mss, n;
493 
494 	KASSERT(inc != NULL || pmss > 0,
495 	    ("%s: at least one of inc/pmss must be specified", __func__));
496 
497 	mss = inc ? tcp_mssopt(inc) : pmss;
498 	if (pmss > 0 && mss > pmss)
499 		mss = pmss;
500 
501 	if (inc->inc_flags & INC_ISIPV6)
502 		n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
503 	else
504 		n = sizeof(struct ip) + sizeof(struct tcphdr);
505 
506 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++)
507 		continue;
508 
509 	return (i);
510 }
511 
512 /*
513  * Determine the receive window size for a socket.
514  */
515 u_long
516 select_rcv_wnd(struct socket *so)
517 {
518 	unsigned long wnd;
519 
520 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
521 
522 	wnd = sbspace(&so->so_rcv);
523 	if (wnd < MIN_RCV_WND)
524 		wnd = MIN_RCV_WND;
525 
526 	return min(wnd, MAX_RCV_WND);
527 }
528 
529 int
530 select_rcv_wscale(void)
531 {
532 	int wscale = 0;
533 	unsigned long space = sb_max;
534 
535 	if (space > MAX_RCV_WND)
536 		space = MAX_RCV_WND;
537 
538 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
539 		wscale++;
540 
541 	return (wscale);
542 }
543 
544 extern int always_keepalive;
545 
546 /*
547  * socket so could be a listening socket too.
548  */
549 uint64_t
550 calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e,
551     int mtu_idx, int rscale, int rx_credits, int ulp_mode)
552 {
553 	uint64_t opt0;
554 
555 	KASSERT(rx_credits <= M_RCV_BUFSIZ,
556 	    ("%s: rcv_bufsiz too high", __func__));
557 
558 	opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) |
559 	    V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits);
560 
561 	if (so != NULL) {
562 		struct inpcb *inp = sotoinpcb(so);
563 		struct tcpcb *tp = intotcpcb(inp);
564 		int keepalive = always_keepalive ||
565 		    so_options_get(so) & SO_KEEPALIVE;
566 
567 		opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0);
568 		opt0 |= V_KEEP_ALIVE(keepalive != 0);
569 	}
570 
571 	if (e != NULL)
572 		opt0 |= V_L2T_IDX(e->idx);
573 
574 	if (vi != NULL) {
575 		opt0 |= V_SMAC_SEL(vi->smt_idx);
576 		opt0 |= V_TX_CHAN(vi->pi->tx_chan);
577 	}
578 
579 	return htobe64(opt0);
580 }
581 
582 uint64_t
583 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
584 {
585 	struct adapter *sc = vi->pi->adapter;
586 	struct tp_params *tp = &sc->params.tp;
587 	uint16_t viid = vi->viid;
588 	uint64_t ntuple = 0;
589 
590 	/*
591 	 * Initialize each of the fields which we care about which are present
592 	 * in the Compressed Filter Tuple.
593 	 */
594 	if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE)
595 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
596 
597 	if (tp->port_shift >= 0)
598 		ntuple |= (uint64_t)e->lport << tp->port_shift;
599 
600 	if (tp->protocol_shift >= 0)
601 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
602 
603 	if (tp->vnic_shift >= 0) {
604 		uint32_t vf = G_FW_VIID_VIN(viid);
605 		uint32_t pf = G_FW_VIID_PFN(viid);
606 		uint32_t vld = G_FW_VIID_VIVLD(viid);
607 
608 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) |
609 		    V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift;
610 	}
611 
612 	if (is_t4(sc))
613 		return (htobe32((uint32_t)ntuple));
614 	else
615 		return (htobe64(V_FILTER_TUPLE(ntuple)));
616 }
617 
618 void
619 set_tcpddp_ulp_mode(struct toepcb *toep)
620 {
621 
622 	toep->ulp_mode = ULP_MODE_TCPDDP;
623 	toep->ddp_flags = DDP_OK;
624 }
625 
626 int
627 negative_advice(int status)
628 {
629 
630 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
631 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
632 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
633 }
634 
635 static int
636 alloc_tid_tabs(struct tid_info *t)
637 {
638 	size_t size;
639 	unsigned int i;
640 
641 	size = t->ntids * sizeof(*t->tid_tab) +
642 	    t->natids * sizeof(*t->atid_tab) +
643 	    t->nstids * sizeof(*t->stid_tab);
644 
645 	t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT);
646 	if (t->tid_tab == NULL)
647 		return (ENOMEM);
648 
649 	mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF);
650 	t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids];
651 	t->afree = t->atid_tab;
652 	t->atids_in_use = 0;
653 	for (i = 1; i < t->natids; i++)
654 		t->atid_tab[i - 1].next = &t->atid_tab[i];
655 	t->atid_tab[t->natids - 1].next = NULL;
656 
657 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
658 	t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids];
659 	t->stids_in_use = 0;
660 	TAILQ_INIT(&t->stids);
661 	t->nstids_free_head = t->nstids;
662 
663 	atomic_store_rel_int(&t->tids_in_use, 0);
664 
665 	return (0);
666 }
667 
668 static void
669 free_tid_tabs(struct tid_info *t)
670 {
671 	KASSERT(t->tids_in_use == 0,
672 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
673 	KASSERT(t->atids_in_use == 0,
674 	    ("%s: %d atids still in use.", __func__, t->atids_in_use));
675 	KASSERT(t->stids_in_use == 0,
676 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
677 
678 	free(t->tid_tab, M_CXGBE);
679 	t->tid_tab = NULL;
680 
681 	if (mtx_initialized(&t->atid_lock))
682 		mtx_destroy(&t->atid_lock);
683 	if (mtx_initialized(&t->stid_lock))
684 		mtx_destroy(&t->stid_lock);
685 }
686 
687 static int
688 add_lip(struct adapter *sc, struct in6_addr *lip)
689 {
690         struct fw_clip_cmd c;
691 
692 	ASSERT_SYNCHRONIZED_OP(sc);
693 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
694 
695         memset(&c, 0, sizeof(c));
696 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
697 	    F_FW_CMD_WRITE);
698         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c));
699         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
700         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
701 
702 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
703 }
704 
705 static int
706 delete_lip(struct adapter *sc, struct in6_addr *lip)
707 {
708 	struct fw_clip_cmd c;
709 
710 	ASSERT_SYNCHRONIZED_OP(sc);
711 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
712 
713 	memset(&c, 0, sizeof(c));
714 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
715 	    F_FW_CMD_READ);
716         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c));
717         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
718         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
719 
720 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
721 }
722 
723 static struct clip_entry *
724 search_lip(struct tom_data *td, struct in6_addr *lip)
725 {
726 	struct clip_entry *ce;
727 
728 	mtx_assert(&td->clip_table_lock, MA_OWNED);
729 
730 	TAILQ_FOREACH(ce, &td->clip_table, link) {
731 		if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
732 			return (ce);
733 	}
734 
735 	return (NULL);
736 }
737 
738 struct clip_entry *
739 hold_lip(struct tom_data *td, struct in6_addr *lip, struct clip_entry *ce)
740 {
741 
742 	mtx_lock(&td->clip_table_lock);
743 	if (ce == NULL)
744 		ce = search_lip(td, lip);
745 	if (ce != NULL)
746 		ce->refcount++;
747 	mtx_unlock(&td->clip_table_lock);
748 
749 	return (ce);
750 }
751 
752 void
753 release_lip(struct tom_data *td, struct clip_entry *ce)
754 {
755 
756 	mtx_lock(&td->clip_table_lock);
757 	KASSERT(search_lip(td, &ce->lip) == ce,
758 	    ("%s: CLIP entry %p p not in CLIP table.", __func__, ce));
759 	KASSERT(ce->refcount > 0,
760 	    ("%s: CLIP entry %p has refcount 0", __func__, ce));
761 	--ce->refcount;
762 	mtx_unlock(&td->clip_table_lock);
763 }
764 
765 static void
766 init_clip_table(struct adapter *sc, struct tom_data *td)
767 {
768 
769 	ASSERT_SYNCHRONIZED_OP(sc);
770 
771 	mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF);
772 	TAILQ_INIT(&td->clip_table);
773 	td->clip_gen = -1;
774 
775 	update_clip_table(sc, td);
776 }
777 
778 static void
779 update_clip(struct adapter *sc, void *arg __unused)
780 {
781 
782 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc"))
783 		return;
784 
785 	if (uld_active(sc, ULD_TOM))
786 		update_clip_table(sc, sc->tom_softc);
787 
788 	end_synchronized_op(sc, LOCK_HELD);
789 }
790 
791 static void
792 t4_clip_task(void *arg, int count)
793 {
794 
795 	t4_iterate(update_clip, NULL);
796 }
797 
798 static void
799 update_clip_table(struct adapter *sc, struct tom_data *td)
800 {
801 	struct rm_priotracker in6_ifa_tracker;
802 	struct in6_ifaddr *ia;
803 	struct in6_addr *lip, tlip;
804 	struct clip_head stale;
805 	struct clip_entry *ce, *ce_temp;
806 	struct vi_info *vi;
807 	int rc, gen, i, j;
808 	uintptr_t last_vnet;
809 
810 	ASSERT_SYNCHRONIZED_OP(sc);
811 
812 	IN6_IFADDR_RLOCK(&in6_ifa_tracker);
813 	mtx_lock(&td->clip_table_lock);
814 
815 	gen = atomic_load_acq_int(&in6_ifaddr_gen);
816 	if (gen == td->clip_gen)
817 		goto done;
818 
819 	TAILQ_INIT(&stale);
820 	TAILQ_CONCAT(&stale, &td->clip_table, link);
821 
822 	/*
823 	 * last_vnet optimizes the common cases where all if_vnet = NULL (no
824 	 * VIMAGE) or all if_vnet = vnet0.
825 	 */
826 	last_vnet = (uintptr_t)(-1);
827 	for_each_port(sc, i)
828 	for_each_vi(sc->port[i], j, vi) {
829 		if (last_vnet == (uintptr_t)vi->ifp->if_vnet)
830 			continue;
831 
832 		/* XXX: races with if_vmove */
833 		CURVNET_SET(vi->ifp->if_vnet);
834 		TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) {
835 			lip = &ia->ia_addr.sin6_addr;
836 
837 			KASSERT(!IN6_IS_ADDR_MULTICAST(lip),
838 			    ("%s: mcast address in in6_ifaddr list", __func__));
839 
840 			if (IN6_IS_ADDR_LOOPBACK(lip))
841 				continue;
842 			if (IN6_IS_SCOPE_EMBED(lip)) {
843 				/* Remove the embedded scope */
844 				tlip = *lip;
845 				lip = &tlip;
846 				in6_clearscope(lip);
847 			}
848 			/*
849 			 * XXX: how to weed out the link local address for the
850 			 * loopback interface?  It's fe80::1 usually (always?).
851 			 */
852 
853 			/*
854 			 * If it's in the main list then we already know it's
855 			 * not stale.
856 			 */
857 			TAILQ_FOREACH(ce, &td->clip_table, link) {
858 				if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
859 					goto next;
860 			}
861 
862 			/*
863 			 * If it's in the stale list we should move it to the
864 			 * main list.
865 			 */
866 			TAILQ_FOREACH(ce, &stale, link) {
867 				if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) {
868 					TAILQ_REMOVE(&stale, ce, link);
869 					TAILQ_INSERT_TAIL(&td->clip_table, ce,
870 					    link);
871 					goto next;
872 				}
873 			}
874 
875 			/* A new IP6 address; add it to the CLIP table */
876 			ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT);
877 			memcpy(&ce->lip, lip, sizeof(ce->lip));
878 			ce->refcount = 0;
879 			rc = add_lip(sc, lip);
880 			if (rc == 0)
881 				TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
882 			else {
883 				char ip[INET6_ADDRSTRLEN];
884 
885 				inet_ntop(AF_INET6, &ce->lip, &ip[0],
886 				    sizeof(ip));
887 				log(LOG_ERR, "%s: could not add %s (%d)\n",
888 				    __func__, ip, rc);
889 				free(ce, M_CXGBE);
890 			}
891 next:
892 			continue;
893 		}
894 		CURVNET_RESTORE();
895 		last_vnet = (uintptr_t)vi->ifp->if_vnet;
896 	}
897 
898 	/*
899 	 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are
900 	 * no longer referenced by the driver.
901 	 */
902 	TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) {
903 		if (ce->refcount == 0) {
904 			rc = delete_lip(sc, &ce->lip);
905 			if (rc == 0) {
906 				TAILQ_REMOVE(&stale, ce, link);
907 				free(ce, M_CXGBE);
908 			} else {
909 				char ip[INET6_ADDRSTRLEN];
910 
911 				inet_ntop(AF_INET6, &ce->lip, &ip[0],
912 				    sizeof(ip));
913 				log(LOG_ERR, "%s: could not delete %s (%d)\n",
914 				    __func__, ip, rc);
915 			}
916 		}
917 	}
918 	/* The ones that are still referenced need to stay in the CLIP table */
919 	TAILQ_CONCAT(&td->clip_table, &stale, link);
920 
921 	td->clip_gen = gen;
922 done:
923 	mtx_unlock(&td->clip_table_lock);
924 	IN6_IFADDR_RUNLOCK(&in6_ifa_tracker);
925 }
926 
927 static void
928 destroy_clip_table(struct adapter *sc, struct tom_data *td)
929 {
930 	struct clip_entry *ce, *ce_temp;
931 
932 	if (mtx_initialized(&td->clip_table_lock)) {
933 		mtx_lock(&td->clip_table_lock);
934 		TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) {
935 			KASSERT(ce->refcount == 0,
936 			    ("%s: CLIP entry %p still in use (%d)", __func__,
937 			    ce, ce->refcount));
938 			TAILQ_REMOVE(&td->clip_table, ce, link);
939 			delete_lip(sc, &ce->lip);
940 			free(ce, M_CXGBE);
941 		}
942 		mtx_unlock(&td->clip_table_lock);
943 		mtx_destroy(&td->clip_table_lock);
944 	}
945 }
946 
947 static void
948 free_tom_data(struct adapter *sc, struct tom_data *td)
949 {
950 
951 	ASSERT_SYNCHRONIZED_OP(sc);
952 
953 	KASSERT(TAILQ_EMPTY(&td->toep_list),
954 	    ("%s: TOE PCB list is not empty.", __func__));
955 	KASSERT(td->lctx_count == 0,
956 	    ("%s: lctx hash table is not empty.", __func__));
957 
958 	t4_free_ppod_region(&td->pr);
959 	destroy_clip_table(sc, td);
960 
961 	if (td->listen_mask != 0)
962 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
963 
964 	if (mtx_initialized(&td->unsent_wr_lock))
965 		mtx_destroy(&td->unsent_wr_lock);
966 	if (mtx_initialized(&td->lctx_hash_lock))
967 		mtx_destroy(&td->lctx_hash_lock);
968 	if (mtx_initialized(&td->toep_list_lock))
969 		mtx_destroy(&td->toep_list_lock);
970 
971 	free_tid_tabs(&sc->tids);
972 	free(td, M_CXGBE);
973 }
974 
975 static void
976 reclaim_wr_resources(void *arg, int count)
977 {
978 	struct tom_data *td = arg;
979 	STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
980 	struct cpl_act_open_req *cpl;
981 	u_int opcode, atid;
982 	struct wrqe *wr;
983 	struct adapter *sc;
984 
985 	mtx_lock(&td->unsent_wr_lock);
986 	STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
987 	mtx_unlock(&td->unsent_wr_lock);
988 
989 	while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
990 		STAILQ_REMOVE_HEAD(&twr_list, link);
991 
992 		cpl = wrtod(wr);
993 		opcode = GET_OPCODE(cpl);
994 
995 		switch (opcode) {
996 		case CPL_ACT_OPEN_REQ:
997 		case CPL_ACT_OPEN_REQ6:
998 			atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
999 			sc = td_adapter(td);
1000 
1001 			CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1002 			act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1003 			free(wr, M_CXGBE);
1004 			break;
1005 		default:
1006 			log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1007 			    "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1008 			/* WR not freed here; go look at it with a debugger.  */
1009 		}
1010 	}
1011 }
1012 
1013 /*
1014  * Ground control to Major TOM
1015  * Commencing countdown, engines on
1016  */
1017 static int
1018 t4_tom_activate(struct adapter *sc)
1019 {
1020 	struct tom_data *td;
1021 	struct toedev *tod;
1022 	struct vi_info *vi;
1023 	struct sge_ofld_rxq *ofld_rxq;
1024 	int i, j, rc, v;
1025 
1026 	ASSERT_SYNCHRONIZED_OP(sc);
1027 
1028 	/* per-adapter softc for TOM */
1029 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1030 	if (td == NULL)
1031 		return (ENOMEM);
1032 
1033 	/* List of TOE PCBs and associated lock */
1034 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1035 	TAILQ_INIT(&td->toep_list);
1036 
1037 	/* Listen context */
1038 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1039 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1040 	    &td->listen_mask, HASH_NOWAIT);
1041 
1042 	/* List of WRs for which L2 resolution failed */
1043 	mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1044 	STAILQ_INIT(&td->unsent_wr_list);
1045 	TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1046 
1047 	/* TID tables */
1048 	rc = alloc_tid_tabs(&sc->tids);
1049 	if (rc != 0)
1050 		goto done;
1051 
1052 	rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1053 	    t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1054 	if (rc != 0)
1055 		goto done;
1056 	t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1057 	    V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1058 
1059 	/* CLIP table for IPv6 offload */
1060 	init_clip_table(sc, td);
1061 
1062 	/* toedev ops */
1063 	tod = &td->tod;
1064 	init_toedev(tod);
1065 	tod->tod_softc = sc;
1066 	tod->tod_connect = t4_connect;
1067 	tod->tod_listen_start = t4_listen_start;
1068 	tod->tod_listen_stop = t4_listen_stop;
1069 	tod->tod_rcvd = t4_rcvd;
1070 	tod->tod_output = t4_tod_output;
1071 	tod->tod_send_rst = t4_send_rst;
1072 	tod->tod_send_fin = t4_send_fin;
1073 	tod->tod_pcb_detach = t4_pcb_detach;
1074 	tod->tod_l2_update = t4_l2_update;
1075 	tod->tod_syncache_added = t4_syncache_added;
1076 	tod->tod_syncache_removed = t4_syncache_removed;
1077 	tod->tod_syncache_respond = t4_syncache_respond;
1078 	tod->tod_offload_socket = t4_offload_socket;
1079 	tod->tod_ctloutput = t4_ctloutput;
1080 
1081 	for_each_port(sc, i) {
1082 		for_each_vi(sc->port[i], v, vi) {
1083 			TOEDEV(vi->ifp) = &td->tod;
1084 			for_each_ofld_rxq(vi, j, ofld_rxq) {
1085 				ofld_rxq->iq.set_tcb_rpl = do_set_tcb_rpl;
1086 				ofld_rxq->iq.l2t_write_rpl = do_l2t_write_rpl2;
1087 			}
1088 		}
1089 	}
1090 
1091 	sc->tom_softc = td;
1092 	register_toedev(sc->tom_softc);
1093 
1094 done:
1095 	if (rc != 0)
1096 		free_tom_data(sc, td);
1097 	return (rc);
1098 }
1099 
1100 static int
1101 t4_tom_deactivate(struct adapter *sc)
1102 {
1103 	int rc = 0;
1104 	struct tom_data *td = sc->tom_softc;
1105 
1106 	ASSERT_SYNCHRONIZED_OP(sc);
1107 
1108 	if (td == NULL)
1109 		return (0);	/* XXX. KASSERT? */
1110 
1111 	if (sc->offload_map != 0)
1112 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1113 
1114 	if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1115 		return (EBUSY);	/* both iWARP and iSCSI rely on the TOE. */
1116 
1117 	mtx_lock(&td->toep_list_lock);
1118 	if (!TAILQ_EMPTY(&td->toep_list))
1119 		rc = EBUSY;
1120 	mtx_unlock(&td->toep_list_lock);
1121 
1122 	mtx_lock(&td->lctx_hash_lock);
1123 	if (td->lctx_count > 0)
1124 		rc = EBUSY;
1125 	mtx_unlock(&td->lctx_hash_lock);
1126 
1127 	taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1128 	mtx_lock(&td->unsent_wr_lock);
1129 	if (!STAILQ_EMPTY(&td->unsent_wr_list))
1130 		rc = EBUSY;
1131 	mtx_unlock(&td->unsent_wr_lock);
1132 
1133 	if (rc == 0) {
1134 		unregister_toedev(sc->tom_softc);
1135 		free_tom_data(sc, td);
1136 		sc->tom_softc = NULL;
1137 	}
1138 
1139 	return (rc);
1140 }
1141 
1142 static void
1143 t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp)
1144 {
1145 
1146 	atomic_add_rel_int(&in6_ifaddr_gen, 1);
1147 	taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4);
1148 }
1149 
1150 static int
1151 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1152 {
1153 	struct tcpcb *tp = so_sototcpcb(so);
1154 	struct toepcb *toep = tp->t_toe;
1155 	int error;
1156 
1157 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1158 		error = t4_aio_queue_ddp(so, job);
1159 		if (error != EOPNOTSUPP)
1160 			return (error);
1161 	}
1162 
1163 	return (t4_aio_queue_aiotx(so, job));
1164 }
1165 
1166 static int
1167 t4_tom_mod_load(void)
1168 {
1169 	int rc;
1170 	struct protosw *tcp_protosw, *tcp6_protosw;
1171 
1172 	/* CPL handlers */
1173 	t4_init_connect_cpl_handlers();
1174 	t4_init_listen_cpl_handlers();
1175 	t4_init_cpl_io_handlers();
1176 
1177 	rc = t4_ddp_mod_load();
1178 	if (rc != 0)
1179 		return (rc);
1180 
1181 	tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM);
1182 	if (tcp_protosw == NULL)
1183 		return (ENOPROTOOPT);
1184 	bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw));
1185 	bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs));
1186 	toe_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1187 	toe_protosw.pr_usrreqs = &toe_usrreqs;
1188 
1189 	tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM);
1190 	if (tcp6_protosw == NULL)
1191 		return (ENOPROTOOPT);
1192 	bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
1193 	bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs));
1194 	toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1195 	toe6_protosw.pr_usrreqs = &toe6_usrreqs;
1196 
1197 	TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL);
1198 	ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event,
1199 	    t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY);
1200 
1201 	rc = t4_register_uld(&tom_uld_info);
1202 	if (rc != 0)
1203 		t4_tom_mod_unload();
1204 
1205 	return (rc);
1206 }
1207 
1208 static void
1209 tom_uninit(struct adapter *sc, void *arg __unused)
1210 {
1211 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
1212 		return;
1213 
1214 	/* Try to free resources (works only if no port has IFCAP_TOE) */
1215 	if (uld_active(sc, ULD_TOM))
1216 		t4_deactivate_uld(sc, ULD_TOM);
1217 
1218 	end_synchronized_op(sc, 0);
1219 }
1220 
1221 static int
1222 t4_tom_mod_unload(void)
1223 {
1224 	t4_iterate(tom_uninit, NULL);
1225 
1226 	if (t4_unregister_uld(&tom_uld_info) == EBUSY)
1227 		return (EBUSY);
1228 
1229 	if (ifaddr_evhandler) {
1230 		EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler);
1231 		taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL);
1232 	}
1233 
1234 	t4_ddp_mod_unload();
1235 
1236 	t4_uninit_connect_cpl_handlers();
1237 	t4_uninit_listen_cpl_handlers();
1238 	t4_uninit_cpl_io_handlers();
1239 
1240 	return (0);
1241 }
1242 #endif	/* TCP_OFFLOAD */
1243 
1244 static int
1245 t4_tom_modevent(module_t mod, int cmd, void *arg)
1246 {
1247 	int rc = 0;
1248 
1249 #ifdef TCP_OFFLOAD
1250 	switch (cmd) {
1251 	case MOD_LOAD:
1252 		rc = t4_tom_mod_load();
1253 		break;
1254 
1255 	case MOD_UNLOAD:
1256 		rc = t4_tom_mod_unload();
1257 		break;
1258 
1259 	default:
1260 		rc = EINVAL;
1261 	}
1262 #else
1263 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
1264 	rc = EOPNOTSUPP;
1265 #endif
1266 	return (rc);
1267 }
1268 
1269 static moduledata_t t4_tom_moddata= {
1270 	"t4_tom",
1271 	t4_tom_modevent,
1272 	0
1273 };
1274 
1275 MODULE_VERSION(t4_tom, 1);
1276 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
1277 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
1278 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
1279