xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision f4b37ed0f8b307b1f3f0f630ca725d68f1dff30d)
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 
34 #include <sys/param.h>
35 #include <sys/types.h>
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/ktr.h>
39 #include <sys/module.h>
40 #include <sys/protosw.h>
41 #include <sys/domain.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/taskqueue.h>
45 #include <net/if.h>
46 #include <net/if_var.h>
47 #include <netinet/in.h>
48 #include <netinet/in_pcb.h>
49 #include <netinet/in_var.h>
50 #include <netinet/ip.h>
51 #include <netinet/ip6.h>
52 #include <netinet/tcp_var.h>
53 #include <netinet6/scope6_var.h>
54 #define TCPSTATES
55 #include <netinet/tcp_fsm.h>
56 #include <netinet/toecore.h>
57 
58 #ifdef TCP_OFFLOAD
59 #include "common/common.h"
60 #include "common/t4_msg.h"
61 #include "common/t4_regs.h"
62 #include "common/t4_regs_values.h"
63 #include "common/t4_tcb.h"
64 #include "tom/t4_tom_l2t.h"
65 #include "tom/t4_tom.h"
66 
67 static struct protosw ddp_protosw;
68 static struct pr_usrreqs ddp_usrreqs;
69 
70 static struct protosw ddp6_protosw;
71 static struct pr_usrreqs ddp6_usrreqs;
72 
73 /* Module ops */
74 static int t4_tom_mod_load(void);
75 static int t4_tom_mod_unload(void);
76 static int t4_tom_modevent(module_t, int, void *);
77 
78 /* ULD ops and helpers */
79 static int t4_tom_activate(struct adapter *);
80 static int t4_tom_deactivate(struct adapter *);
81 
82 static struct uld_info tom_uld_info = {
83 	.uld_id = ULD_TOM,
84 	.activate = t4_tom_activate,
85 	.deactivate = t4_tom_deactivate,
86 };
87 
88 static void queue_tid_release(struct adapter *, int);
89 static void release_offload_resources(struct toepcb *);
90 static int alloc_tid_tabs(struct tid_info *);
91 static void free_tid_tabs(struct tid_info *);
92 static int add_lip(struct adapter *, struct in6_addr *);
93 static int delete_lip(struct adapter *, struct in6_addr *);
94 static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *);
95 static void init_clip_table(struct adapter *, struct tom_data *);
96 static void update_clip(struct adapter *, void *);
97 static void t4_clip_task(void *, int);
98 static void update_clip_table(struct adapter *, struct tom_data *);
99 static void destroy_clip_table(struct adapter *, struct tom_data *);
100 static void free_tom_data(struct adapter *, struct tom_data *);
101 static void reclaim_wr_resources(void *, int);
102 
103 static int in6_ifaddr_gen;
104 static eventhandler_tag ifaddr_evhandler;
105 static struct timeout_task clip_task;
106 
107 struct toepcb *
108 alloc_toepcb(struct port_info *pi, int txqid, int rxqid, int flags)
109 {
110 	struct adapter *sc = pi->adapter;
111 	struct toepcb *toep;
112 	int tx_credits, txsd_total, len;
113 
114 	/*
115 	 * The firmware counts tx work request credits in units of 16 bytes
116 	 * each.  Reserve room for an ABORT_REQ so the driver never has to worry
117 	 * about tx credits if it wants to abort a connection.
118 	 */
119 	tx_credits = sc->params.ofldq_wr_cred;
120 	tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
121 
122 	/*
123 	 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
124 	 * immediate payload, and firmware counts tx work request credits in
125 	 * units of 16 byte.  Calculate the maximum work requests possible.
126 	 */
127 	txsd_total = tx_credits /
128 	    howmany((sizeof(struct fw_ofld_tx_data_wr) + 1), 16);
129 
130 	if (txqid < 0)
131 		txqid = (arc4random() % pi->nofldtxq) + pi->first_ofld_txq;
132 	KASSERT(txqid >= pi->first_ofld_txq &&
133 	    txqid < pi->first_ofld_txq + pi->nofldtxq,
134 	    ("%s: txqid %d for port %p (first %d, n %d)", __func__, txqid, pi,
135 		pi->first_ofld_txq, pi->nofldtxq));
136 
137 	if (rxqid < 0)
138 		rxqid = (arc4random() % pi->nofldrxq) + pi->first_ofld_rxq;
139 	KASSERT(rxqid >= pi->first_ofld_rxq &&
140 	    rxqid < pi->first_ofld_rxq + pi->nofldrxq,
141 	    ("%s: rxqid %d for port %p (first %d, n %d)", __func__, rxqid, pi,
142 		pi->first_ofld_rxq, pi->nofldrxq));
143 
144 	len = offsetof(struct toepcb, txsd) +
145 	    txsd_total * sizeof(struct ofld_tx_sdesc);
146 
147 	toep = malloc(len, M_CXGBE, M_ZERO | flags);
148 	if (toep == NULL)
149 		return (NULL);
150 
151 	toep->td = sc->tom_softc;
152 	toep->port = pi;
153 	toep->tx_total = tx_credits;
154 	toep->tx_credits = tx_credits;
155 	toep->ofld_txq = &sc->sge.ofld_txq[txqid];
156 	toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid];
157 	toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
158 	toep->txsd_total = txsd_total;
159 	toep->txsd_avail = txsd_total;
160 	toep->txsd_pidx = 0;
161 	toep->txsd_cidx = 0;
162 
163 	return (toep);
164 }
165 
166 void
167 free_toepcb(struct toepcb *toep)
168 {
169 
170 	KASSERT(!(toep->flags & TPF_ATTACHED),
171 	    ("%s: attached to an inpcb", __func__));
172 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
173 	    ("%s: CPL pending", __func__));
174 
175 	free(toep, M_CXGBE);
176 }
177 
178 /*
179  * Set up the socket for TCP offload.
180  */
181 void
182 offload_socket(struct socket *so, struct toepcb *toep)
183 {
184 	struct tom_data *td = toep->td;
185 	struct inpcb *inp = sotoinpcb(so);
186 	struct tcpcb *tp = intotcpcb(inp);
187 	struct sockbuf *sb;
188 
189 	INP_WLOCK_ASSERT(inp);
190 
191 	/* Update socket */
192 	sb = &so->so_snd;
193 	SOCKBUF_LOCK(sb);
194 	sb->sb_flags |= SB_NOCOALESCE;
195 	SOCKBUF_UNLOCK(sb);
196 	sb = &so->so_rcv;
197 	SOCKBUF_LOCK(sb);
198 	sb->sb_flags |= SB_NOCOALESCE;
199 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
200 		if (inp->inp_vflag & INP_IPV6)
201 			so->so_proto = &ddp6_protosw;
202 		else
203 			so->so_proto = &ddp_protosw;
204 	}
205 	SOCKBUF_UNLOCK(sb);
206 
207 	/* Update TCP PCB */
208 	tp->tod = &td->tod;
209 	tp->t_toe = toep;
210 	tp->t_flags |= TF_TOE;
211 
212 	/* Install an extra hold on inp */
213 	toep->inp = inp;
214 	toep->flags |= TPF_ATTACHED;
215 	in_pcbref(inp);
216 
217 	/* Add the TOE PCB to the active list */
218 	mtx_lock(&td->toep_list_lock);
219 	TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
220 	mtx_unlock(&td->toep_list_lock);
221 }
222 
223 /* This is _not_ the normal way to "unoffload" a socket. */
224 void
225 undo_offload_socket(struct socket *so)
226 {
227 	struct inpcb *inp = sotoinpcb(so);
228 	struct tcpcb *tp = intotcpcb(inp);
229 	struct toepcb *toep = tp->t_toe;
230 	struct tom_data *td = toep->td;
231 	struct sockbuf *sb;
232 
233 	INP_WLOCK_ASSERT(inp);
234 
235 	sb = &so->so_snd;
236 	SOCKBUF_LOCK(sb);
237 	sb->sb_flags &= ~SB_NOCOALESCE;
238 	SOCKBUF_UNLOCK(sb);
239 	sb = &so->so_rcv;
240 	SOCKBUF_LOCK(sb);
241 	sb->sb_flags &= ~SB_NOCOALESCE;
242 	SOCKBUF_UNLOCK(sb);
243 
244 	tp->tod = NULL;
245 	tp->t_toe = NULL;
246 	tp->t_flags &= ~TF_TOE;
247 
248 	toep->inp = NULL;
249 	toep->flags &= ~TPF_ATTACHED;
250 	if (in_pcbrele_wlocked(inp))
251 		panic("%s: inp freed.", __func__);
252 
253 	mtx_lock(&td->toep_list_lock);
254 	TAILQ_REMOVE(&td->toep_list, toep, link);
255 	mtx_unlock(&td->toep_list_lock);
256 }
257 
258 static void
259 release_offload_resources(struct toepcb *toep)
260 {
261 	struct tom_data *td = toep->td;
262 	struct adapter *sc = td_adapter(td);
263 	int tid = toep->tid;
264 
265 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
266 	    ("%s: %p has CPL pending.", __func__, toep));
267 	KASSERT(!(toep->flags & TPF_ATTACHED),
268 	    ("%s: %p is still attached.", __func__, toep));
269 
270 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
271 	    __func__, toep, tid, toep->l2te, toep->ce);
272 
273 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
274 		release_ddp_resources(toep);
275 
276 	if (toep->l2te)
277 		t4_l2t_release(toep->l2te);
278 
279 	if (tid >= 0) {
280 		remove_tid(sc, tid);
281 		release_tid(sc, tid, toep->ctrlq);
282 	}
283 
284 	if (toep->ce)
285 		release_lip(td, toep->ce);
286 
287 	mtx_lock(&td->toep_list_lock);
288 	TAILQ_REMOVE(&td->toep_list, toep, link);
289 	mtx_unlock(&td->toep_list_lock);
290 
291 	free_toepcb(toep);
292 }
293 
294 /*
295  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
296  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
297  * pending CPL) then it is time to release all resources tied to the toepcb.
298  *
299  * Also gets called when an offloaded active open fails and the TOM wants the
300  * kernel to take the TCP PCB back.
301  */
302 static void
303 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
304 {
305 #if defined(KTR) || defined(INVARIANTS)
306 	struct inpcb *inp = tp->t_inpcb;
307 #endif
308 	struct toepcb *toep = tp->t_toe;
309 
310 	INP_WLOCK_ASSERT(inp);
311 
312 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
313 	KASSERT(toep->flags & TPF_ATTACHED,
314 	    ("%s: not attached", __func__));
315 
316 #ifdef KTR
317 	if (tp->t_state == TCPS_SYN_SENT) {
318 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
319 		    __func__, toep->tid, toep, toep->flags, inp,
320 		    inp->inp_flags);
321 	} else {
322 		CTR6(KTR_CXGBE,
323 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
324 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
325 		    inp->inp_flags);
326 	}
327 #endif
328 
329 	tp->t_toe = NULL;
330 	tp->t_flags &= ~TF_TOE;
331 	toep->flags &= ~TPF_ATTACHED;
332 
333 	if (!(toep->flags & TPF_CPL_PENDING))
334 		release_offload_resources(toep);
335 }
336 
337 /*
338  * setsockopt handler.
339  */
340 static void
341 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
342 {
343 	struct adapter *sc = tod->tod_softc;
344 	struct toepcb *toep = tp->t_toe;
345 
346 	if (dir == SOPT_GET)
347 		return;
348 
349 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
350 
351 	switch (name) {
352 	case TCP_NODELAY:
353 		t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS, V_TF_NAGLE(1),
354 		    V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1));
355 		break;
356 	default:
357 		break;
358 	}
359 }
360 
361 /*
362  * The TOE driver will not receive any more CPLs for the tid associated with the
363  * toepcb; release the hold on the inpcb.
364  */
365 void
366 final_cpl_received(struct toepcb *toep)
367 {
368 	struct inpcb *inp = toep->inp;
369 
370 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
371 	INP_WLOCK_ASSERT(inp);
372 	KASSERT(toep->flags & TPF_CPL_PENDING,
373 	    ("%s: CPL not pending already?", __func__));
374 
375 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
376 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
377 
378 	toep->inp = NULL;
379 	toep->flags &= ~TPF_CPL_PENDING;
380 
381 	if (!(toep->flags & TPF_ATTACHED))
382 		release_offload_resources(toep);
383 
384 	if (!in_pcbrele_wlocked(inp))
385 		INP_WUNLOCK(inp);
386 }
387 
388 void
389 insert_tid(struct adapter *sc, int tid, void *ctx)
390 {
391 	struct tid_info *t = &sc->tids;
392 
393 	t->tid_tab[tid] = ctx;
394 	atomic_add_int(&t->tids_in_use, 1);
395 }
396 
397 void *
398 lookup_tid(struct adapter *sc, int tid)
399 {
400 	struct tid_info *t = &sc->tids;
401 
402 	return (t->tid_tab[tid]);
403 }
404 
405 void
406 update_tid(struct adapter *sc, int tid, void *ctx)
407 {
408 	struct tid_info *t = &sc->tids;
409 
410 	t->tid_tab[tid] = ctx;
411 }
412 
413 void
414 remove_tid(struct adapter *sc, int tid)
415 {
416 	struct tid_info *t = &sc->tids;
417 
418 	t->tid_tab[tid] = NULL;
419 	atomic_subtract_int(&t->tids_in_use, 1);
420 }
421 
422 void
423 release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq)
424 {
425 	struct wrqe *wr;
426 	struct cpl_tid_release *req;
427 
428 	wr = alloc_wrqe(sizeof(*req), ctrlq);
429 	if (wr == NULL) {
430 		queue_tid_release(sc, tid);	/* defer */
431 		return;
432 	}
433 	req = wrtod(wr);
434 
435 	INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid);
436 
437 	t4_wrq_tx(sc, wr);
438 }
439 
440 static void
441 queue_tid_release(struct adapter *sc, int tid)
442 {
443 
444 	CXGBE_UNIMPLEMENTED("deferred tid release");
445 }
446 
447 /*
448  * What mtu_idx to use, given a 4-tuple and/or an MSS cap
449  */
450 int
451 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss)
452 {
453 	unsigned short *mtus = &sc->params.mtus[0];
454 	int i, mss, n;
455 
456 	KASSERT(inc != NULL || pmss > 0,
457 	    ("%s: at least one of inc/pmss must be specified", __func__));
458 
459 	mss = inc ? tcp_mssopt(inc) : pmss;
460 	if (pmss > 0 && mss > pmss)
461 		mss = pmss;
462 
463 	if (inc->inc_flags & INC_ISIPV6)
464 		n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
465 	else
466 		n = sizeof(struct ip) + sizeof(struct tcphdr);
467 
468 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++)
469 		continue;
470 
471 	return (i);
472 }
473 
474 /*
475  * Determine the receive window size for a socket.
476  */
477 u_long
478 select_rcv_wnd(struct socket *so)
479 {
480 	unsigned long wnd;
481 
482 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
483 
484 	wnd = sbspace(&so->so_rcv);
485 	if (wnd < MIN_RCV_WND)
486 		wnd = MIN_RCV_WND;
487 
488 	return min(wnd, MAX_RCV_WND);
489 }
490 
491 int
492 select_rcv_wscale(void)
493 {
494 	int wscale = 0;
495 	unsigned long space = sb_max;
496 
497 	if (space > MAX_RCV_WND)
498 		space = MAX_RCV_WND;
499 
500 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
501 		wscale++;
502 
503 	return (wscale);
504 }
505 
506 extern int always_keepalive;
507 #define VIID_SMACIDX(v)	(((unsigned int)(v) & 0x7f) << 1)
508 
509 /*
510  * socket so could be a listening socket too.
511  */
512 uint64_t
513 calc_opt0(struct socket *so, struct port_info *pi, struct l2t_entry *e,
514     int mtu_idx, int rscale, int rx_credits, int ulp_mode)
515 {
516 	uint64_t opt0;
517 
518 	KASSERT(rx_credits <= M_RCV_BUFSIZ,
519 	    ("%s: rcv_bufsiz too high", __func__));
520 
521 	opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) |
522 	    V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits);
523 
524 	if (so != NULL) {
525 		struct inpcb *inp = sotoinpcb(so);
526 		struct tcpcb *tp = intotcpcb(inp);
527 		int keepalive = always_keepalive ||
528 		    so_options_get(so) & SO_KEEPALIVE;
529 
530 		opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0);
531 		opt0 |= V_KEEP_ALIVE(keepalive != 0);
532 	}
533 
534 	if (e != NULL)
535 		opt0 |= V_L2T_IDX(e->idx);
536 
537 	if (pi != NULL) {
538 		opt0 |= V_SMAC_SEL(VIID_SMACIDX(pi->viid));
539 		opt0 |= V_TX_CHAN(pi->tx_chan);
540 	}
541 
542 	return htobe64(opt0);
543 }
544 
545 uint64_t
546 select_ntuple(struct port_info *pi, struct l2t_entry *e)
547 {
548 	struct adapter *sc = pi->adapter;
549 	struct tp_params *tp = &sc->params.tp;
550 	uint16_t viid = pi->viid;
551 	uint64_t ntuple = 0;
552 
553 	/*
554 	 * Initialize each of the fields which we care about which are present
555 	 * in the Compressed Filter Tuple.
556 	 */
557 	if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE)
558 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
559 
560 	if (tp->port_shift >= 0)
561 		ntuple |= (uint64_t)e->lport << tp->port_shift;
562 
563 	if (tp->protocol_shift >= 0)
564 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
565 
566 	if (tp->vnic_shift >= 0) {
567 		uint32_t vf = G_FW_VIID_VIN(viid);
568 		uint32_t pf = G_FW_VIID_PFN(viid);
569 		uint32_t vld = G_FW_VIID_VIVLD(viid);
570 
571 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) |
572 		    V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift;
573 	}
574 
575 	if (is_t4(sc))
576 		return (htobe32((uint32_t)ntuple));
577 	else
578 		return (htobe64(V_FILTER_TUPLE(ntuple)));
579 }
580 
581 void
582 set_tcpddp_ulp_mode(struct toepcb *toep)
583 {
584 
585 	toep->ulp_mode = ULP_MODE_TCPDDP;
586 	toep->ddp_flags = DDP_OK;
587 	toep->ddp_score = DDP_LOW_SCORE;
588 }
589 
590 int
591 negative_advice(int status)
592 {
593 
594 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
595 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
596 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
597 }
598 
599 static int
600 alloc_tid_tabs(struct tid_info *t)
601 {
602 	size_t size;
603 	unsigned int i;
604 
605 	size = t->ntids * sizeof(*t->tid_tab) +
606 	    t->natids * sizeof(*t->atid_tab) +
607 	    t->nstids * sizeof(*t->stid_tab);
608 
609 	t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT);
610 	if (t->tid_tab == NULL)
611 		return (ENOMEM);
612 
613 	mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF);
614 	t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids];
615 	t->afree = t->atid_tab;
616 	t->atids_in_use = 0;
617 	for (i = 1; i < t->natids; i++)
618 		t->atid_tab[i - 1].next = &t->atid_tab[i];
619 	t->atid_tab[t->natids - 1].next = NULL;
620 
621 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
622 	t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids];
623 	t->stids_in_use = 0;
624 	TAILQ_INIT(&t->stids);
625 	t->nstids_free_head = t->nstids;
626 
627 	atomic_store_rel_int(&t->tids_in_use, 0);
628 
629 	return (0);
630 }
631 
632 static void
633 free_tid_tabs(struct tid_info *t)
634 {
635 	KASSERT(t->tids_in_use == 0,
636 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
637 	KASSERT(t->atids_in_use == 0,
638 	    ("%s: %d atids still in use.", __func__, t->atids_in_use));
639 	KASSERT(t->stids_in_use == 0,
640 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
641 
642 	free(t->tid_tab, M_CXGBE);
643 	t->tid_tab = NULL;
644 
645 	if (mtx_initialized(&t->atid_lock))
646 		mtx_destroy(&t->atid_lock);
647 	if (mtx_initialized(&t->stid_lock))
648 		mtx_destroy(&t->stid_lock);
649 }
650 
651 static int
652 add_lip(struct adapter *sc, struct in6_addr *lip)
653 {
654         struct fw_clip_cmd c;
655 
656 	ASSERT_SYNCHRONIZED_OP(sc);
657 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
658 
659         memset(&c, 0, sizeof(c));
660 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
661 	    F_FW_CMD_WRITE);
662         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c));
663         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
664         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
665 
666 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
667 }
668 
669 static int
670 delete_lip(struct adapter *sc, struct in6_addr *lip)
671 {
672 	struct fw_clip_cmd c;
673 
674 	ASSERT_SYNCHRONIZED_OP(sc);
675 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
676 
677 	memset(&c, 0, sizeof(c));
678 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
679 	    F_FW_CMD_READ);
680         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c));
681         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
682         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
683 
684 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
685 }
686 
687 static struct clip_entry *
688 search_lip(struct tom_data *td, struct in6_addr *lip)
689 {
690 	struct clip_entry *ce;
691 
692 	mtx_assert(&td->clip_table_lock, MA_OWNED);
693 
694 	TAILQ_FOREACH(ce, &td->clip_table, link) {
695 		if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
696 			return (ce);
697 	}
698 
699 	return (NULL);
700 }
701 
702 struct clip_entry *
703 hold_lip(struct tom_data *td, struct in6_addr *lip)
704 {
705 	struct clip_entry *ce;
706 
707 	mtx_lock(&td->clip_table_lock);
708 	ce = search_lip(td, lip);
709 	if (ce != NULL)
710 		ce->refcount++;
711 	mtx_unlock(&td->clip_table_lock);
712 
713 	return (ce);
714 }
715 
716 void
717 release_lip(struct tom_data *td, struct clip_entry *ce)
718 {
719 
720 	mtx_lock(&td->clip_table_lock);
721 	KASSERT(search_lip(td, &ce->lip) == ce,
722 	    ("%s: CLIP entry %p p not in CLIP table.", __func__, ce));
723 	KASSERT(ce->refcount > 0,
724 	    ("%s: CLIP entry %p has refcount 0", __func__, ce));
725 	--ce->refcount;
726 	mtx_unlock(&td->clip_table_lock);
727 }
728 
729 static void
730 init_clip_table(struct adapter *sc, struct tom_data *td)
731 {
732 
733 	ASSERT_SYNCHRONIZED_OP(sc);
734 
735 	mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF);
736 	TAILQ_INIT(&td->clip_table);
737 	td->clip_gen = -1;
738 
739 	update_clip_table(sc, td);
740 }
741 
742 static void
743 update_clip(struct adapter *sc, void *arg __unused)
744 {
745 
746 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc"))
747 		return;
748 
749 	if (uld_active(sc, ULD_TOM))
750 		update_clip_table(sc, sc->tom_softc);
751 
752 	end_synchronized_op(sc, LOCK_HELD);
753 }
754 
755 static void
756 t4_clip_task(void *arg, int count)
757 {
758 
759 	t4_iterate(update_clip, NULL);
760 }
761 
762 static void
763 update_clip_table(struct adapter *sc, struct tom_data *td)
764 {
765 	struct in6_ifaddr *ia;
766 	struct in6_addr *lip, tlip;
767 	struct clip_head stale;
768 	struct clip_entry *ce, *ce_temp;
769 	int rc, gen = atomic_load_acq_int(&in6_ifaddr_gen);
770 
771 	ASSERT_SYNCHRONIZED_OP(sc);
772 
773 	IN6_IFADDR_RLOCK();
774 	mtx_lock(&td->clip_table_lock);
775 
776 	if (gen == td->clip_gen)
777 		goto done;
778 
779 	TAILQ_INIT(&stale);
780 	TAILQ_CONCAT(&stale, &td->clip_table, link);
781 
782 	TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) {
783 		lip = &ia->ia_addr.sin6_addr;
784 
785 		KASSERT(!IN6_IS_ADDR_MULTICAST(lip),
786 		    ("%s: mcast address in in6_ifaddr list", __func__));
787 
788 		if (IN6_IS_ADDR_LOOPBACK(lip))
789 			continue;
790 		if (IN6_IS_SCOPE_EMBED(lip)) {
791 			/* Remove the embedded scope */
792 			tlip = *lip;
793 			lip = &tlip;
794 			in6_clearscope(lip);
795 		}
796 		/*
797 		 * XXX: how to weed out the link local address for the loopback
798 		 * interface?  It's fe80::1 usually (always?).
799 		 */
800 
801 		/*
802 		 * If it's in the main list then we already know it's not stale.
803 		 */
804 		TAILQ_FOREACH(ce, &td->clip_table, link) {
805 			if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
806 				goto next;
807 		}
808 
809 		/*
810 		 * If it's in the stale list we should move it to the main list.
811 		 */
812 		TAILQ_FOREACH(ce, &stale, link) {
813 			if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) {
814 				TAILQ_REMOVE(&stale, ce, link);
815 				TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
816 				goto next;
817 			}
818 		}
819 
820 		/* A new IP6 address; add it to the CLIP table */
821 		ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT);
822 		memcpy(&ce->lip, lip, sizeof(ce->lip));
823 		ce->refcount = 0;
824 		rc = add_lip(sc, lip);
825 		if (rc == 0)
826 			TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
827 		else {
828 			char ip[INET6_ADDRSTRLEN];
829 
830 			inet_ntop(AF_INET6, &ce->lip, &ip[0], sizeof(ip));
831 			log(LOG_ERR, "%s: could not add %s (%d)\n",
832 			    __func__, ip, rc);
833 			free(ce, M_CXGBE);
834 		}
835 next:
836 		continue;
837 	}
838 
839 	/*
840 	 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are
841 	 * no longer referenced by the driver.
842 	 */
843 	TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) {
844 		if (ce->refcount == 0) {
845 			rc = delete_lip(sc, &ce->lip);
846 			if (rc == 0) {
847 				TAILQ_REMOVE(&stale, ce, link);
848 				free(ce, M_CXGBE);
849 			} else {
850 				char ip[INET6_ADDRSTRLEN];
851 
852 				inet_ntop(AF_INET6, &ce->lip, &ip[0],
853 				    sizeof(ip));
854 				log(LOG_ERR, "%s: could not delete %s (%d)\n",
855 				    __func__, ip, rc);
856 			}
857 		}
858 	}
859 	/* The ones that are still referenced need to stay in the CLIP table */
860 	TAILQ_CONCAT(&td->clip_table, &stale, link);
861 
862 	td->clip_gen = gen;
863 done:
864 	mtx_unlock(&td->clip_table_lock);
865 	IN6_IFADDR_RUNLOCK();
866 }
867 
868 static void
869 destroy_clip_table(struct adapter *sc, struct tom_data *td)
870 {
871 	struct clip_entry *ce, *ce_temp;
872 
873 	if (mtx_initialized(&td->clip_table_lock)) {
874 		mtx_lock(&td->clip_table_lock);
875 		TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) {
876 			KASSERT(ce->refcount == 0,
877 			    ("%s: CLIP entry %p still in use (%d)", __func__,
878 			    ce, ce->refcount));
879 			TAILQ_REMOVE(&td->clip_table, ce, link);
880 			delete_lip(sc, &ce->lip);
881 			free(ce, M_CXGBE);
882 		}
883 		mtx_unlock(&td->clip_table_lock);
884 		mtx_destroy(&td->clip_table_lock);
885 	}
886 }
887 
888 static void
889 free_tom_data(struct adapter *sc, struct tom_data *td)
890 {
891 
892 	ASSERT_SYNCHRONIZED_OP(sc);
893 
894 	KASSERT(TAILQ_EMPTY(&td->toep_list),
895 	    ("%s: TOE PCB list is not empty.", __func__));
896 	KASSERT(td->lctx_count == 0,
897 	    ("%s: lctx hash table is not empty.", __func__));
898 
899 	t4_uninit_l2t_cpl_handlers(sc);
900 	t4_uninit_cpl_io_handlers(sc);
901 	t4_uninit_ddp(sc, td);
902 	destroy_clip_table(sc, td);
903 
904 	if (td->listen_mask != 0)
905 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
906 
907 	if (mtx_initialized(&td->unsent_wr_lock))
908 		mtx_destroy(&td->unsent_wr_lock);
909 	if (mtx_initialized(&td->lctx_hash_lock))
910 		mtx_destroy(&td->lctx_hash_lock);
911 	if (mtx_initialized(&td->toep_list_lock))
912 		mtx_destroy(&td->toep_list_lock);
913 
914 	free_tid_tabs(&sc->tids);
915 	free(td, M_CXGBE);
916 }
917 
918 static void
919 reclaim_wr_resources(void *arg, int count)
920 {
921 	struct tom_data *td = arg;
922 	STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
923 	struct cpl_act_open_req *cpl;
924 	u_int opcode, atid;
925 	struct wrqe *wr;
926 	struct adapter *sc;
927 
928 	mtx_lock(&td->unsent_wr_lock);
929 	STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
930 	mtx_unlock(&td->unsent_wr_lock);
931 
932 	while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
933 		STAILQ_REMOVE_HEAD(&twr_list, link);
934 
935 		cpl = wrtod(wr);
936 		opcode = GET_OPCODE(cpl);
937 
938 		switch (opcode) {
939 		case CPL_ACT_OPEN_REQ:
940 		case CPL_ACT_OPEN_REQ6:
941 			atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
942 			sc = td_adapter(td);
943 
944 			CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
945 			act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
946 			free(wr, M_CXGBE);
947 			break;
948 		default:
949 			log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
950 			    "opcode %x\n", __func__, wr, wr->wr_len, opcode);
951 			/* WR not freed here; go look at it with a debugger.  */
952 		}
953 	}
954 }
955 
956 /*
957  * Ground control to Major TOM
958  * Commencing countdown, engines on
959  */
960 static int
961 t4_tom_activate(struct adapter *sc)
962 {
963 	struct tom_data *td;
964 	struct toedev *tod;
965 	int i, rc;
966 
967 	ASSERT_SYNCHRONIZED_OP(sc);
968 
969 	/* per-adapter softc for TOM */
970 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
971 	if (td == NULL)
972 		return (ENOMEM);
973 
974 	/* List of TOE PCBs and associated lock */
975 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
976 	TAILQ_INIT(&td->toep_list);
977 
978 	/* Listen context */
979 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
980 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
981 	    &td->listen_mask, HASH_NOWAIT);
982 
983 	/* List of WRs for which L2 resolution failed */
984 	mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
985 	STAILQ_INIT(&td->unsent_wr_list);
986 	TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
987 
988 	/* TID tables */
989 	rc = alloc_tid_tabs(&sc->tids);
990 	if (rc != 0)
991 		goto done;
992 
993 	/* DDP page pods and CPL handlers */
994 	t4_init_ddp(sc, td);
995 
996 	/* CLIP table for IPv6 offload */
997 	init_clip_table(sc, td);
998 
999 	/* CPL handlers */
1000 	t4_init_connect_cpl_handlers(sc);
1001 	t4_init_l2t_cpl_handlers(sc);
1002 	t4_init_listen_cpl_handlers(sc);
1003 	t4_init_cpl_io_handlers(sc);
1004 
1005 	/* toedev ops */
1006 	tod = &td->tod;
1007 	init_toedev(tod);
1008 	tod->tod_softc = sc;
1009 	tod->tod_connect = t4_connect;
1010 	tod->tod_listen_start = t4_listen_start;
1011 	tod->tod_listen_stop = t4_listen_stop;
1012 	tod->tod_rcvd = t4_rcvd;
1013 	tod->tod_output = t4_tod_output;
1014 	tod->tod_send_rst = t4_send_rst;
1015 	tod->tod_send_fin = t4_send_fin;
1016 	tod->tod_pcb_detach = t4_pcb_detach;
1017 	tod->tod_l2_update = t4_l2_update;
1018 	tod->tod_syncache_added = t4_syncache_added;
1019 	tod->tod_syncache_removed = t4_syncache_removed;
1020 	tod->tod_syncache_respond = t4_syncache_respond;
1021 	tod->tod_offload_socket = t4_offload_socket;
1022 	tod->tod_ctloutput = t4_ctloutput;
1023 
1024 	for_each_port(sc, i)
1025 		TOEDEV(sc->port[i]->ifp) = &td->tod;
1026 
1027 	sc->tom_softc = td;
1028 	register_toedev(sc->tom_softc);
1029 
1030 done:
1031 	if (rc != 0)
1032 		free_tom_data(sc, td);
1033 	return (rc);
1034 }
1035 
1036 static int
1037 t4_tom_deactivate(struct adapter *sc)
1038 {
1039 	int rc = 0;
1040 	struct tom_data *td = sc->tom_softc;
1041 
1042 	ASSERT_SYNCHRONIZED_OP(sc);
1043 
1044 	if (td == NULL)
1045 		return (0);	/* XXX. KASSERT? */
1046 
1047 	if (sc->offload_map != 0)
1048 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1049 
1050 	if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1051 		return (EBUSY);	/* both iWARP and iSCSI rely on the TOE. */
1052 
1053 	mtx_lock(&td->toep_list_lock);
1054 	if (!TAILQ_EMPTY(&td->toep_list))
1055 		rc = EBUSY;
1056 	mtx_unlock(&td->toep_list_lock);
1057 
1058 	mtx_lock(&td->lctx_hash_lock);
1059 	if (td->lctx_count > 0)
1060 		rc = EBUSY;
1061 	mtx_unlock(&td->lctx_hash_lock);
1062 
1063 	taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1064 	mtx_lock(&td->unsent_wr_lock);
1065 	if (!STAILQ_EMPTY(&td->unsent_wr_list))
1066 		rc = EBUSY;
1067 	mtx_unlock(&td->unsent_wr_lock);
1068 
1069 	if (rc == 0) {
1070 		unregister_toedev(sc->tom_softc);
1071 		free_tom_data(sc, td);
1072 		sc->tom_softc = NULL;
1073 	}
1074 
1075 	return (rc);
1076 }
1077 
1078 static void
1079 t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp)
1080 {
1081 
1082 	atomic_add_rel_int(&in6_ifaddr_gen, 1);
1083 	taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4);
1084 }
1085 
1086 static int
1087 t4_tom_mod_load(void)
1088 {
1089 	int rc;
1090 	struct protosw *tcp_protosw, *tcp6_protosw;
1091 
1092 	tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM);
1093 	if (tcp_protosw == NULL)
1094 		return (ENOPROTOOPT);
1095 	bcopy(tcp_protosw, &ddp_protosw, sizeof(ddp_protosw));
1096 	bcopy(tcp_protosw->pr_usrreqs, &ddp_usrreqs, sizeof(ddp_usrreqs));
1097 	ddp_usrreqs.pru_soreceive = t4_soreceive_ddp;
1098 	ddp_protosw.pr_usrreqs = &ddp_usrreqs;
1099 
1100 	tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM);
1101 	if (tcp6_protosw == NULL)
1102 		return (ENOPROTOOPT);
1103 	bcopy(tcp6_protosw, &ddp6_protosw, sizeof(ddp6_protosw));
1104 	bcopy(tcp6_protosw->pr_usrreqs, &ddp6_usrreqs, sizeof(ddp6_usrreqs));
1105 	ddp6_usrreqs.pru_soreceive = t4_soreceive_ddp;
1106 	ddp6_protosw.pr_usrreqs = &ddp6_usrreqs;
1107 
1108 	TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL);
1109 	ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event,
1110 	    t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY);
1111 
1112 	rc = t4_register_uld(&tom_uld_info);
1113 	if (rc != 0)
1114 		t4_tom_mod_unload();
1115 
1116 	return (rc);
1117 }
1118 
1119 static void
1120 tom_uninit(struct adapter *sc, void *arg __unused)
1121 {
1122 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
1123 		return;
1124 
1125 	/* Try to free resources (works only if no port has IFCAP_TOE) */
1126 	if (uld_active(sc, ULD_TOM))
1127 		t4_deactivate_uld(sc, ULD_TOM);
1128 
1129 	end_synchronized_op(sc, 0);
1130 }
1131 
1132 static int
1133 t4_tom_mod_unload(void)
1134 {
1135 	t4_iterate(tom_uninit, NULL);
1136 
1137 	if (t4_unregister_uld(&tom_uld_info) == EBUSY)
1138 		return (EBUSY);
1139 
1140 	if (ifaddr_evhandler) {
1141 		EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler);
1142 		taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL);
1143 	}
1144 
1145 	return (0);
1146 }
1147 #endif	/* TCP_OFFLOAD */
1148 
1149 static int
1150 t4_tom_modevent(module_t mod, int cmd, void *arg)
1151 {
1152 	int rc = 0;
1153 
1154 #ifdef TCP_OFFLOAD
1155 	switch (cmd) {
1156 	case MOD_LOAD:
1157 		rc = t4_tom_mod_load();
1158 		break;
1159 
1160 	case MOD_UNLOAD:
1161 		rc = t4_tom_mod_unload();
1162 		break;
1163 
1164 	default:
1165 		rc = EINVAL;
1166 	}
1167 #else
1168 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
1169 	rc = EOPNOTSUPP;
1170 #endif
1171 	return (rc);
1172 }
1173 
1174 static moduledata_t t4_tom_moddata= {
1175 	"t4_tom",
1176 	t4_tom_modevent,
1177 	0
1178 };
1179 
1180 MODULE_VERSION(t4_tom, 1);
1181 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
1182 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
1183 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
1184