xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision 895f86f15fbf6540071feb9328c3c50ed1f027b8)
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/lock.h>
40 #include <sys/module.h>
41 #include <sys/protosw.h>
42 #include <sys/domain.h>
43 #include <sys/rmlock.h>
44 #include <sys/socket.h>
45 #include <sys/socketvar.h>
46 #include <sys/taskqueue.h>
47 #include <net/if.h>
48 #include <net/if_var.h>
49 #include <netinet/in.h>
50 #include <netinet/in_pcb.h>
51 #include <netinet/in_var.h>
52 #include <netinet/ip.h>
53 #include <netinet/ip6.h>
54 #include <netinet/tcp_var.h>
55 #include <netinet6/scope6_var.h>
56 #define TCPSTATES
57 #include <netinet/tcp_fsm.h>
58 #include <netinet/toecore.h>
59 
60 #ifdef TCP_OFFLOAD
61 #include "common/common.h"
62 #include "common/t4_msg.h"
63 #include "common/t4_regs.h"
64 #include "common/t4_regs_values.h"
65 #include "common/t4_tcb.h"
66 #include "tom/t4_tom_l2t.h"
67 #include "tom/t4_tom.h"
68 
69 static struct protosw ddp_protosw;
70 static struct pr_usrreqs ddp_usrreqs;
71 
72 static struct protosw ddp6_protosw;
73 static struct pr_usrreqs ddp6_usrreqs;
74 
75 /* Module ops */
76 static int t4_tom_mod_load(void);
77 static int t4_tom_mod_unload(void);
78 static int t4_tom_modevent(module_t, int, void *);
79 
80 /* ULD ops and helpers */
81 static int t4_tom_activate(struct adapter *);
82 static int t4_tom_deactivate(struct adapter *);
83 
84 static struct uld_info tom_uld_info = {
85 	.uld_id = ULD_TOM,
86 	.activate = t4_tom_activate,
87 	.deactivate = t4_tom_deactivate,
88 };
89 
90 static void queue_tid_release(struct adapter *, int);
91 static void release_offload_resources(struct toepcb *);
92 static int alloc_tid_tabs(struct tid_info *);
93 static void free_tid_tabs(struct tid_info *);
94 static int add_lip(struct adapter *, struct in6_addr *);
95 static int delete_lip(struct adapter *, struct in6_addr *);
96 static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *);
97 static void init_clip_table(struct adapter *, struct tom_data *);
98 static void update_clip(struct adapter *, void *);
99 static void t4_clip_task(void *, int);
100 static void update_clip_table(struct adapter *, struct tom_data *);
101 static void destroy_clip_table(struct adapter *, struct tom_data *);
102 static void free_tom_data(struct adapter *, struct tom_data *);
103 static void reclaim_wr_resources(void *, int);
104 
105 static int in6_ifaddr_gen;
106 static eventhandler_tag ifaddr_evhandler;
107 static struct timeout_task clip_task;
108 
109 struct toepcb *
110 alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags)
111 {
112 	struct port_info *pi = vi->pi;
113 	struct adapter *sc = pi->adapter;
114 	struct toepcb *toep;
115 	int tx_credits, txsd_total, len;
116 
117 	/*
118 	 * The firmware counts tx work request credits in units of 16 bytes
119 	 * each.  Reserve room for an ABORT_REQ so the driver never has to worry
120 	 * about tx credits if it wants to abort a connection.
121 	 */
122 	tx_credits = sc->params.ofldq_wr_cred;
123 	tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
124 
125 	/*
126 	 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
127 	 * immediate payload, and firmware counts tx work request credits in
128 	 * units of 16 byte.  Calculate the maximum work requests possible.
129 	 */
130 	txsd_total = tx_credits /
131 	    howmany((sizeof(struct fw_ofld_tx_data_wr) + 1), 16);
132 
133 	if (txqid < 0)
134 		txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq;
135 	KASSERT(txqid >= vi->first_ofld_txq &&
136 	    txqid < vi->first_ofld_txq + vi->nofldtxq,
137 	    ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi,
138 		vi->first_ofld_txq, vi->nofldtxq));
139 
140 	if (rxqid < 0)
141 		rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq;
142 	KASSERT(rxqid >= vi->first_ofld_rxq &&
143 	    rxqid < vi->first_ofld_rxq + vi->nofldrxq,
144 	    ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi,
145 		vi->first_ofld_rxq, vi->nofldrxq));
146 
147 	len = offsetof(struct toepcb, txsd) +
148 	    txsd_total * sizeof(struct ofld_tx_sdesc);
149 
150 	toep = malloc(len, M_CXGBE, M_ZERO | flags);
151 	if (toep == NULL)
152 		return (NULL);
153 
154 	toep->td = sc->tom_softc;
155 	toep->vi = vi;
156 	toep->tx_total = tx_credits;
157 	toep->tx_credits = tx_credits;
158 	toep->ofld_txq = &sc->sge.ofld_txq[txqid];
159 	toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid];
160 	toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
161 	toep->txsd_total = txsd_total;
162 	toep->txsd_avail = txsd_total;
163 	toep->txsd_pidx = 0;
164 	toep->txsd_cidx = 0;
165 
166 	return (toep);
167 }
168 
169 void
170 free_toepcb(struct toepcb *toep)
171 {
172 
173 	KASSERT(!(toep->flags & TPF_ATTACHED),
174 	    ("%s: attached to an inpcb", __func__));
175 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
176 	    ("%s: CPL pending", __func__));
177 
178 	free(toep, M_CXGBE);
179 }
180 
181 /*
182  * Set up the socket for TCP offload.
183  */
184 void
185 offload_socket(struct socket *so, struct toepcb *toep)
186 {
187 	struct tom_data *td = toep->td;
188 	struct inpcb *inp = sotoinpcb(so);
189 	struct tcpcb *tp = intotcpcb(inp);
190 	struct sockbuf *sb;
191 
192 	INP_WLOCK_ASSERT(inp);
193 
194 	/* Update socket */
195 	sb = &so->so_snd;
196 	SOCKBUF_LOCK(sb);
197 	sb->sb_flags |= SB_NOCOALESCE;
198 	SOCKBUF_UNLOCK(sb);
199 	sb = &so->so_rcv;
200 	SOCKBUF_LOCK(sb);
201 	sb->sb_flags |= SB_NOCOALESCE;
202 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
203 		if (inp->inp_vflag & INP_IPV6)
204 			so->so_proto = &ddp6_protosw;
205 		else
206 			so->so_proto = &ddp_protosw;
207 	}
208 	SOCKBUF_UNLOCK(sb);
209 
210 	/* Update TCP PCB */
211 	tp->tod = &td->tod;
212 	tp->t_toe = toep;
213 	tp->t_flags |= TF_TOE;
214 
215 	/* Install an extra hold on inp */
216 	toep->inp = inp;
217 	toep->flags |= TPF_ATTACHED;
218 	in_pcbref(inp);
219 
220 	/* Add the TOE PCB to the active list */
221 	mtx_lock(&td->toep_list_lock);
222 	TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
223 	mtx_unlock(&td->toep_list_lock);
224 }
225 
226 /* This is _not_ the normal way to "unoffload" a socket. */
227 void
228 undo_offload_socket(struct socket *so)
229 {
230 	struct inpcb *inp = sotoinpcb(so);
231 	struct tcpcb *tp = intotcpcb(inp);
232 	struct toepcb *toep = tp->t_toe;
233 	struct tom_data *td = toep->td;
234 	struct sockbuf *sb;
235 
236 	INP_WLOCK_ASSERT(inp);
237 
238 	sb = &so->so_snd;
239 	SOCKBUF_LOCK(sb);
240 	sb->sb_flags &= ~SB_NOCOALESCE;
241 	SOCKBUF_UNLOCK(sb);
242 	sb = &so->so_rcv;
243 	SOCKBUF_LOCK(sb);
244 	sb->sb_flags &= ~SB_NOCOALESCE;
245 	SOCKBUF_UNLOCK(sb);
246 
247 	tp->tod = NULL;
248 	tp->t_toe = NULL;
249 	tp->t_flags &= ~TF_TOE;
250 
251 	toep->inp = NULL;
252 	toep->flags &= ~TPF_ATTACHED;
253 	if (in_pcbrele_wlocked(inp))
254 		panic("%s: inp freed.", __func__);
255 
256 	mtx_lock(&td->toep_list_lock);
257 	TAILQ_REMOVE(&td->toep_list, toep, link);
258 	mtx_unlock(&td->toep_list_lock);
259 }
260 
261 static void
262 release_offload_resources(struct toepcb *toep)
263 {
264 	struct tom_data *td = toep->td;
265 	struct adapter *sc = td_adapter(td);
266 	int tid = toep->tid;
267 
268 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
269 	    ("%s: %p has CPL pending.", __func__, toep));
270 	KASSERT(!(toep->flags & TPF_ATTACHED),
271 	    ("%s: %p is still attached.", __func__, toep));
272 
273 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
274 	    __func__, toep, tid, toep->l2te, toep->ce);
275 
276 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
277 		release_ddp_resources(toep);
278 
279 	if (toep->l2te)
280 		t4_l2t_release(toep->l2te);
281 
282 	if (tid >= 0) {
283 		remove_tid(sc, tid);
284 		release_tid(sc, tid, toep->ctrlq);
285 	}
286 
287 	if (toep->ce)
288 		release_lip(td, toep->ce);
289 
290 	mtx_lock(&td->toep_list_lock);
291 	TAILQ_REMOVE(&td->toep_list, toep, link);
292 	mtx_unlock(&td->toep_list_lock);
293 
294 	free_toepcb(toep);
295 }
296 
297 /*
298  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
299  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
300  * pending CPL) then it is time to release all resources tied to the toepcb.
301  *
302  * Also gets called when an offloaded active open fails and the TOM wants the
303  * kernel to take the TCP PCB back.
304  */
305 static void
306 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
307 {
308 #if defined(KTR) || defined(INVARIANTS)
309 	struct inpcb *inp = tp->t_inpcb;
310 #endif
311 	struct toepcb *toep = tp->t_toe;
312 
313 	INP_WLOCK_ASSERT(inp);
314 
315 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
316 	KASSERT(toep->flags & TPF_ATTACHED,
317 	    ("%s: not attached", __func__));
318 
319 #ifdef KTR
320 	if (tp->t_state == TCPS_SYN_SENT) {
321 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
322 		    __func__, toep->tid, toep, toep->flags, inp,
323 		    inp->inp_flags);
324 	} else {
325 		CTR6(KTR_CXGBE,
326 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
327 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
328 		    inp->inp_flags);
329 	}
330 #endif
331 
332 	tp->t_toe = NULL;
333 	tp->t_flags &= ~TF_TOE;
334 	toep->flags &= ~TPF_ATTACHED;
335 
336 	if (!(toep->flags & TPF_CPL_PENDING))
337 		release_offload_resources(toep);
338 }
339 
340 /*
341  * setsockopt handler.
342  */
343 static void
344 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
345 {
346 	struct adapter *sc = tod->tod_softc;
347 	struct toepcb *toep = tp->t_toe;
348 
349 	if (dir == SOPT_GET)
350 		return;
351 
352 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
353 
354 	switch (name) {
355 	case TCP_NODELAY:
356 		t4_set_tcb_field(sc, toep, 1, W_TCB_T_FLAGS, V_TF_NAGLE(1),
357 		    V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1));
358 		break;
359 	default:
360 		break;
361 	}
362 }
363 
364 /*
365  * The TOE driver will not receive any more CPLs for the tid associated with the
366  * toepcb; release the hold on the inpcb.
367  */
368 void
369 final_cpl_received(struct toepcb *toep)
370 {
371 	struct inpcb *inp = toep->inp;
372 
373 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
374 	INP_WLOCK_ASSERT(inp);
375 	KASSERT(toep->flags & TPF_CPL_PENDING,
376 	    ("%s: CPL not pending already?", __func__));
377 
378 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
379 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
380 
381 	toep->inp = NULL;
382 	toep->flags &= ~TPF_CPL_PENDING;
383 
384 	if (!(toep->flags & TPF_ATTACHED))
385 		release_offload_resources(toep);
386 
387 	if (!in_pcbrele_wlocked(inp))
388 		INP_WUNLOCK(inp);
389 }
390 
391 void
392 insert_tid(struct adapter *sc, int tid, void *ctx)
393 {
394 	struct tid_info *t = &sc->tids;
395 
396 	t->tid_tab[tid] = ctx;
397 	atomic_add_int(&t->tids_in_use, 1);
398 }
399 
400 void *
401 lookup_tid(struct adapter *sc, int tid)
402 {
403 	struct tid_info *t = &sc->tids;
404 
405 	return (t->tid_tab[tid]);
406 }
407 
408 void
409 update_tid(struct adapter *sc, int tid, void *ctx)
410 {
411 	struct tid_info *t = &sc->tids;
412 
413 	t->tid_tab[tid] = ctx;
414 }
415 
416 void
417 remove_tid(struct adapter *sc, int tid)
418 {
419 	struct tid_info *t = &sc->tids;
420 
421 	t->tid_tab[tid] = NULL;
422 	atomic_subtract_int(&t->tids_in_use, 1);
423 }
424 
425 void
426 release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq)
427 {
428 	struct wrqe *wr;
429 	struct cpl_tid_release *req;
430 
431 	wr = alloc_wrqe(sizeof(*req), ctrlq);
432 	if (wr == NULL) {
433 		queue_tid_release(sc, tid);	/* defer */
434 		return;
435 	}
436 	req = wrtod(wr);
437 
438 	INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid);
439 
440 	t4_wrq_tx(sc, wr);
441 }
442 
443 static void
444 queue_tid_release(struct adapter *sc, int tid)
445 {
446 
447 	CXGBE_UNIMPLEMENTED("deferred tid release");
448 }
449 
450 /*
451  * What mtu_idx to use, given a 4-tuple and/or an MSS cap
452  */
453 int
454 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss)
455 {
456 	unsigned short *mtus = &sc->params.mtus[0];
457 	int i, mss, n;
458 
459 	KASSERT(inc != NULL || pmss > 0,
460 	    ("%s: at least one of inc/pmss must be specified", __func__));
461 
462 	mss = inc ? tcp_mssopt(inc) : pmss;
463 	if (pmss > 0 && mss > pmss)
464 		mss = pmss;
465 
466 	if (inc->inc_flags & INC_ISIPV6)
467 		n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
468 	else
469 		n = sizeof(struct ip) + sizeof(struct tcphdr);
470 
471 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++)
472 		continue;
473 
474 	return (i);
475 }
476 
477 /*
478  * Determine the receive window size for a socket.
479  */
480 u_long
481 select_rcv_wnd(struct socket *so)
482 {
483 	unsigned long wnd;
484 
485 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
486 
487 	wnd = sbspace(&so->so_rcv);
488 	if (wnd < MIN_RCV_WND)
489 		wnd = MIN_RCV_WND;
490 
491 	return min(wnd, MAX_RCV_WND);
492 }
493 
494 int
495 select_rcv_wscale(void)
496 {
497 	int wscale = 0;
498 	unsigned long space = sb_max;
499 
500 	if (space > MAX_RCV_WND)
501 		space = MAX_RCV_WND;
502 
503 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
504 		wscale++;
505 
506 	return (wscale);
507 }
508 
509 extern int always_keepalive;
510 #define VIID_SMACIDX(v)	(((unsigned int)(v) & 0x7f) << 1)
511 
512 /*
513  * socket so could be a listening socket too.
514  */
515 uint64_t
516 calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e,
517     int mtu_idx, int rscale, int rx_credits, int ulp_mode)
518 {
519 	uint64_t opt0;
520 
521 	KASSERT(rx_credits <= M_RCV_BUFSIZ,
522 	    ("%s: rcv_bufsiz too high", __func__));
523 
524 	opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) |
525 	    V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits);
526 
527 	if (so != NULL) {
528 		struct inpcb *inp = sotoinpcb(so);
529 		struct tcpcb *tp = intotcpcb(inp);
530 		int keepalive = always_keepalive ||
531 		    so_options_get(so) & SO_KEEPALIVE;
532 
533 		opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0);
534 		opt0 |= V_KEEP_ALIVE(keepalive != 0);
535 	}
536 
537 	if (e != NULL)
538 		opt0 |= V_L2T_IDX(e->idx);
539 
540 	if (vi != NULL) {
541 		opt0 |= V_SMAC_SEL(VIID_SMACIDX(vi->viid));
542 		opt0 |= V_TX_CHAN(vi->pi->tx_chan);
543 	}
544 
545 	return htobe64(opt0);
546 }
547 
548 uint64_t
549 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
550 {
551 	struct adapter *sc = vi->pi->adapter;
552 	struct tp_params *tp = &sc->params.tp;
553 	uint16_t viid = vi->viid;
554 	uint64_t ntuple = 0;
555 
556 	/*
557 	 * Initialize each of the fields which we care about which are present
558 	 * in the Compressed Filter Tuple.
559 	 */
560 	if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE)
561 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
562 
563 	if (tp->port_shift >= 0)
564 		ntuple |= (uint64_t)e->lport << tp->port_shift;
565 
566 	if (tp->protocol_shift >= 0)
567 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
568 
569 	if (tp->vnic_shift >= 0) {
570 		uint32_t vf = G_FW_VIID_VIN(viid);
571 		uint32_t pf = G_FW_VIID_PFN(viid);
572 		uint32_t vld = G_FW_VIID_VIVLD(viid);
573 
574 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) |
575 		    V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift;
576 	}
577 
578 	if (is_t4(sc))
579 		return (htobe32((uint32_t)ntuple));
580 	else
581 		return (htobe64(V_FILTER_TUPLE(ntuple)));
582 }
583 
584 void
585 set_tcpddp_ulp_mode(struct toepcb *toep)
586 {
587 
588 	toep->ulp_mode = ULP_MODE_TCPDDP;
589 	toep->ddp_flags = DDP_OK;
590 	toep->ddp_score = DDP_LOW_SCORE;
591 }
592 
593 int
594 negative_advice(int status)
595 {
596 
597 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
598 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
599 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
600 }
601 
602 static int
603 alloc_tid_tabs(struct tid_info *t)
604 {
605 	size_t size;
606 	unsigned int i;
607 
608 	size = t->ntids * sizeof(*t->tid_tab) +
609 	    t->natids * sizeof(*t->atid_tab) +
610 	    t->nstids * sizeof(*t->stid_tab);
611 
612 	t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT);
613 	if (t->tid_tab == NULL)
614 		return (ENOMEM);
615 
616 	mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF);
617 	t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids];
618 	t->afree = t->atid_tab;
619 	t->atids_in_use = 0;
620 	for (i = 1; i < t->natids; i++)
621 		t->atid_tab[i - 1].next = &t->atid_tab[i];
622 	t->atid_tab[t->natids - 1].next = NULL;
623 
624 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
625 	t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids];
626 	t->stids_in_use = 0;
627 	TAILQ_INIT(&t->stids);
628 	t->nstids_free_head = t->nstids;
629 
630 	atomic_store_rel_int(&t->tids_in_use, 0);
631 
632 	return (0);
633 }
634 
635 static void
636 free_tid_tabs(struct tid_info *t)
637 {
638 	KASSERT(t->tids_in_use == 0,
639 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
640 	KASSERT(t->atids_in_use == 0,
641 	    ("%s: %d atids still in use.", __func__, t->atids_in_use));
642 	KASSERT(t->stids_in_use == 0,
643 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
644 
645 	free(t->tid_tab, M_CXGBE);
646 	t->tid_tab = NULL;
647 
648 	if (mtx_initialized(&t->atid_lock))
649 		mtx_destroy(&t->atid_lock);
650 	if (mtx_initialized(&t->stid_lock))
651 		mtx_destroy(&t->stid_lock);
652 }
653 
654 static int
655 add_lip(struct adapter *sc, struct in6_addr *lip)
656 {
657         struct fw_clip_cmd c;
658 
659 	ASSERT_SYNCHRONIZED_OP(sc);
660 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
661 
662         memset(&c, 0, sizeof(c));
663 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
664 	    F_FW_CMD_WRITE);
665         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c));
666         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
667         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
668 
669 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
670 }
671 
672 static int
673 delete_lip(struct adapter *sc, struct in6_addr *lip)
674 {
675 	struct fw_clip_cmd c;
676 
677 	ASSERT_SYNCHRONIZED_OP(sc);
678 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
679 
680 	memset(&c, 0, sizeof(c));
681 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
682 	    F_FW_CMD_READ);
683         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c));
684         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
685         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
686 
687 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
688 }
689 
690 static struct clip_entry *
691 search_lip(struct tom_data *td, struct in6_addr *lip)
692 {
693 	struct clip_entry *ce;
694 
695 	mtx_assert(&td->clip_table_lock, MA_OWNED);
696 
697 	TAILQ_FOREACH(ce, &td->clip_table, link) {
698 		if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
699 			return (ce);
700 	}
701 
702 	return (NULL);
703 }
704 
705 struct clip_entry *
706 hold_lip(struct tom_data *td, struct in6_addr *lip)
707 {
708 	struct clip_entry *ce;
709 
710 	mtx_lock(&td->clip_table_lock);
711 	ce = search_lip(td, lip);
712 	if (ce != NULL)
713 		ce->refcount++;
714 	mtx_unlock(&td->clip_table_lock);
715 
716 	return (ce);
717 }
718 
719 void
720 release_lip(struct tom_data *td, struct clip_entry *ce)
721 {
722 
723 	mtx_lock(&td->clip_table_lock);
724 	KASSERT(search_lip(td, &ce->lip) == ce,
725 	    ("%s: CLIP entry %p p not in CLIP table.", __func__, ce));
726 	KASSERT(ce->refcount > 0,
727 	    ("%s: CLIP entry %p has refcount 0", __func__, ce));
728 	--ce->refcount;
729 	mtx_unlock(&td->clip_table_lock);
730 }
731 
732 static void
733 init_clip_table(struct adapter *sc, struct tom_data *td)
734 {
735 
736 	ASSERT_SYNCHRONIZED_OP(sc);
737 
738 	mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF);
739 	TAILQ_INIT(&td->clip_table);
740 	td->clip_gen = -1;
741 
742 	update_clip_table(sc, td);
743 }
744 
745 static void
746 update_clip(struct adapter *sc, void *arg __unused)
747 {
748 
749 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc"))
750 		return;
751 
752 	if (uld_active(sc, ULD_TOM))
753 		update_clip_table(sc, sc->tom_softc);
754 
755 	end_synchronized_op(sc, LOCK_HELD);
756 }
757 
758 static void
759 t4_clip_task(void *arg, int count)
760 {
761 
762 	t4_iterate(update_clip, NULL);
763 }
764 
765 static void
766 update_clip_table(struct adapter *sc, struct tom_data *td)
767 {
768 	struct rm_priotracker in6_ifa_tracker;
769 	struct in6_ifaddr *ia;
770 	struct in6_addr *lip, tlip;
771 	struct clip_head stale;
772 	struct clip_entry *ce, *ce_temp;
773 	int rc, gen = atomic_load_acq_int(&in6_ifaddr_gen);
774 
775 	ASSERT_SYNCHRONIZED_OP(sc);
776 
777 	IN6_IFADDR_RLOCK(&in6_ifa_tracker);
778 	mtx_lock(&td->clip_table_lock);
779 
780 	if (gen == td->clip_gen)
781 		goto done;
782 
783 	TAILQ_INIT(&stale);
784 	TAILQ_CONCAT(&stale, &td->clip_table, link);
785 
786 	TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) {
787 		lip = &ia->ia_addr.sin6_addr;
788 
789 		KASSERT(!IN6_IS_ADDR_MULTICAST(lip),
790 		    ("%s: mcast address in in6_ifaddr list", __func__));
791 
792 		if (IN6_IS_ADDR_LOOPBACK(lip))
793 			continue;
794 		if (IN6_IS_SCOPE_EMBED(lip)) {
795 			/* Remove the embedded scope */
796 			tlip = *lip;
797 			lip = &tlip;
798 			in6_clearscope(lip);
799 		}
800 		/*
801 		 * XXX: how to weed out the link local address for the loopback
802 		 * interface?  It's fe80::1 usually (always?).
803 		 */
804 
805 		/*
806 		 * If it's in the main list then we already know it's not stale.
807 		 */
808 		TAILQ_FOREACH(ce, &td->clip_table, link) {
809 			if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
810 				goto next;
811 		}
812 
813 		/*
814 		 * If it's in the stale list we should move it to the main list.
815 		 */
816 		TAILQ_FOREACH(ce, &stale, link) {
817 			if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) {
818 				TAILQ_REMOVE(&stale, ce, link);
819 				TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
820 				goto next;
821 			}
822 		}
823 
824 		/* A new IP6 address; add it to the CLIP table */
825 		ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT);
826 		memcpy(&ce->lip, lip, sizeof(ce->lip));
827 		ce->refcount = 0;
828 		rc = add_lip(sc, lip);
829 		if (rc == 0)
830 			TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
831 		else {
832 			char ip[INET6_ADDRSTRLEN];
833 
834 			inet_ntop(AF_INET6, &ce->lip, &ip[0], sizeof(ip));
835 			log(LOG_ERR, "%s: could not add %s (%d)\n",
836 			    __func__, ip, rc);
837 			free(ce, M_CXGBE);
838 		}
839 next:
840 		continue;
841 	}
842 
843 	/*
844 	 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are
845 	 * no longer referenced by the driver.
846 	 */
847 	TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) {
848 		if (ce->refcount == 0) {
849 			rc = delete_lip(sc, &ce->lip);
850 			if (rc == 0) {
851 				TAILQ_REMOVE(&stale, ce, link);
852 				free(ce, M_CXGBE);
853 			} else {
854 				char ip[INET6_ADDRSTRLEN];
855 
856 				inet_ntop(AF_INET6, &ce->lip, &ip[0],
857 				    sizeof(ip));
858 				log(LOG_ERR, "%s: could not delete %s (%d)\n",
859 				    __func__, ip, rc);
860 			}
861 		}
862 	}
863 	/* The ones that are still referenced need to stay in the CLIP table */
864 	TAILQ_CONCAT(&td->clip_table, &stale, link);
865 
866 	td->clip_gen = gen;
867 done:
868 	mtx_unlock(&td->clip_table_lock);
869 	IN6_IFADDR_RUNLOCK(&in6_ifa_tracker);
870 }
871 
872 static void
873 destroy_clip_table(struct adapter *sc, struct tom_data *td)
874 {
875 	struct clip_entry *ce, *ce_temp;
876 
877 	if (mtx_initialized(&td->clip_table_lock)) {
878 		mtx_lock(&td->clip_table_lock);
879 		TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) {
880 			KASSERT(ce->refcount == 0,
881 			    ("%s: CLIP entry %p still in use (%d)", __func__,
882 			    ce, ce->refcount));
883 			TAILQ_REMOVE(&td->clip_table, ce, link);
884 			delete_lip(sc, &ce->lip);
885 			free(ce, M_CXGBE);
886 		}
887 		mtx_unlock(&td->clip_table_lock);
888 		mtx_destroy(&td->clip_table_lock);
889 	}
890 }
891 
892 static void
893 free_tom_data(struct adapter *sc, struct tom_data *td)
894 {
895 
896 	ASSERT_SYNCHRONIZED_OP(sc);
897 
898 	KASSERT(TAILQ_EMPTY(&td->toep_list),
899 	    ("%s: TOE PCB list is not empty.", __func__));
900 	KASSERT(td->lctx_count == 0,
901 	    ("%s: lctx hash table is not empty.", __func__));
902 
903 	t4_uninit_l2t_cpl_handlers(sc);
904 	t4_uninit_cpl_io_handlers(sc);
905 	t4_uninit_ddp(sc, td);
906 	destroy_clip_table(sc, td);
907 
908 	if (td->listen_mask != 0)
909 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
910 
911 	if (mtx_initialized(&td->unsent_wr_lock))
912 		mtx_destroy(&td->unsent_wr_lock);
913 	if (mtx_initialized(&td->lctx_hash_lock))
914 		mtx_destroy(&td->lctx_hash_lock);
915 	if (mtx_initialized(&td->toep_list_lock))
916 		mtx_destroy(&td->toep_list_lock);
917 
918 	free_tid_tabs(&sc->tids);
919 	free(td, M_CXGBE);
920 }
921 
922 static void
923 reclaim_wr_resources(void *arg, int count)
924 {
925 	struct tom_data *td = arg;
926 	STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
927 	struct cpl_act_open_req *cpl;
928 	u_int opcode, atid;
929 	struct wrqe *wr;
930 	struct adapter *sc;
931 
932 	mtx_lock(&td->unsent_wr_lock);
933 	STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
934 	mtx_unlock(&td->unsent_wr_lock);
935 
936 	while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
937 		STAILQ_REMOVE_HEAD(&twr_list, link);
938 
939 		cpl = wrtod(wr);
940 		opcode = GET_OPCODE(cpl);
941 
942 		switch (opcode) {
943 		case CPL_ACT_OPEN_REQ:
944 		case CPL_ACT_OPEN_REQ6:
945 			atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
946 			sc = td_adapter(td);
947 
948 			CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
949 			act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
950 			free(wr, M_CXGBE);
951 			break;
952 		default:
953 			log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
954 			    "opcode %x\n", __func__, wr, wr->wr_len, opcode);
955 			/* WR not freed here; go look at it with a debugger.  */
956 		}
957 	}
958 }
959 
960 /*
961  * Ground control to Major TOM
962  * Commencing countdown, engines on
963  */
964 static int
965 t4_tom_activate(struct adapter *sc)
966 {
967 	struct tom_data *td;
968 	struct toedev *tod;
969 	struct vi_info *vi;
970 	int i, rc, v;
971 
972 	ASSERT_SYNCHRONIZED_OP(sc);
973 
974 	/* per-adapter softc for TOM */
975 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
976 	if (td == NULL)
977 		return (ENOMEM);
978 
979 	/* List of TOE PCBs and associated lock */
980 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
981 	TAILQ_INIT(&td->toep_list);
982 
983 	/* Listen context */
984 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
985 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
986 	    &td->listen_mask, HASH_NOWAIT);
987 
988 	/* List of WRs for which L2 resolution failed */
989 	mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
990 	STAILQ_INIT(&td->unsent_wr_list);
991 	TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
992 
993 	/* TID tables */
994 	rc = alloc_tid_tabs(&sc->tids);
995 	if (rc != 0)
996 		goto done;
997 
998 	/* DDP page pods and CPL handlers */
999 	t4_init_ddp(sc, td);
1000 
1001 	/* CLIP table for IPv6 offload */
1002 	init_clip_table(sc, td);
1003 
1004 	/* CPL handlers */
1005 	t4_init_connect_cpl_handlers(sc);
1006 	t4_init_l2t_cpl_handlers(sc);
1007 	t4_init_listen_cpl_handlers(sc);
1008 	t4_init_cpl_io_handlers(sc);
1009 
1010 	/* toedev ops */
1011 	tod = &td->tod;
1012 	init_toedev(tod);
1013 	tod->tod_softc = sc;
1014 	tod->tod_connect = t4_connect;
1015 	tod->tod_listen_start = t4_listen_start;
1016 	tod->tod_listen_stop = t4_listen_stop;
1017 	tod->tod_rcvd = t4_rcvd;
1018 	tod->tod_output = t4_tod_output;
1019 	tod->tod_send_rst = t4_send_rst;
1020 	tod->tod_send_fin = t4_send_fin;
1021 	tod->tod_pcb_detach = t4_pcb_detach;
1022 	tod->tod_l2_update = t4_l2_update;
1023 	tod->tod_syncache_added = t4_syncache_added;
1024 	tod->tod_syncache_removed = t4_syncache_removed;
1025 	tod->tod_syncache_respond = t4_syncache_respond;
1026 	tod->tod_offload_socket = t4_offload_socket;
1027 	tod->tod_ctloutput = t4_ctloutput;
1028 
1029 	for_each_port(sc, i) {
1030 		for_each_vi(sc->port[i], v, vi) {
1031 			TOEDEV(vi->ifp) = &td->tod;
1032 		}
1033 	}
1034 
1035 	sc->tom_softc = td;
1036 	register_toedev(sc->tom_softc);
1037 
1038 done:
1039 	if (rc != 0)
1040 		free_tom_data(sc, td);
1041 	return (rc);
1042 }
1043 
1044 static int
1045 t4_tom_deactivate(struct adapter *sc)
1046 {
1047 	int rc = 0;
1048 	struct tom_data *td = sc->tom_softc;
1049 
1050 	ASSERT_SYNCHRONIZED_OP(sc);
1051 
1052 	if (td == NULL)
1053 		return (0);	/* XXX. KASSERT? */
1054 
1055 	if (sc->offload_map != 0)
1056 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1057 
1058 	if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1059 		return (EBUSY);	/* both iWARP and iSCSI rely on the TOE. */
1060 
1061 	mtx_lock(&td->toep_list_lock);
1062 	if (!TAILQ_EMPTY(&td->toep_list))
1063 		rc = EBUSY;
1064 	mtx_unlock(&td->toep_list_lock);
1065 
1066 	mtx_lock(&td->lctx_hash_lock);
1067 	if (td->lctx_count > 0)
1068 		rc = EBUSY;
1069 	mtx_unlock(&td->lctx_hash_lock);
1070 
1071 	taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1072 	mtx_lock(&td->unsent_wr_lock);
1073 	if (!STAILQ_EMPTY(&td->unsent_wr_list))
1074 		rc = EBUSY;
1075 	mtx_unlock(&td->unsent_wr_lock);
1076 
1077 	if (rc == 0) {
1078 		unregister_toedev(sc->tom_softc);
1079 		free_tom_data(sc, td);
1080 		sc->tom_softc = NULL;
1081 	}
1082 
1083 	return (rc);
1084 }
1085 
1086 static void
1087 t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp)
1088 {
1089 
1090 	atomic_add_rel_int(&in6_ifaddr_gen, 1);
1091 	taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4);
1092 }
1093 
1094 static int
1095 t4_tom_mod_load(void)
1096 {
1097 	int rc;
1098 	struct protosw *tcp_protosw, *tcp6_protosw;
1099 
1100 	tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM);
1101 	if (tcp_protosw == NULL)
1102 		return (ENOPROTOOPT);
1103 	bcopy(tcp_protosw, &ddp_protosw, sizeof(ddp_protosw));
1104 	bcopy(tcp_protosw->pr_usrreqs, &ddp_usrreqs, sizeof(ddp_usrreqs));
1105 	ddp_usrreqs.pru_soreceive = t4_soreceive_ddp;
1106 	ddp_protosw.pr_usrreqs = &ddp_usrreqs;
1107 
1108 	tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM);
1109 	if (tcp6_protosw == NULL)
1110 		return (ENOPROTOOPT);
1111 	bcopy(tcp6_protosw, &ddp6_protosw, sizeof(ddp6_protosw));
1112 	bcopy(tcp6_protosw->pr_usrreqs, &ddp6_usrreqs, sizeof(ddp6_usrreqs));
1113 	ddp6_usrreqs.pru_soreceive = t4_soreceive_ddp;
1114 	ddp6_protosw.pr_usrreqs = &ddp6_usrreqs;
1115 
1116 	TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL);
1117 	ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event,
1118 	    t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY);
1119 
1120 	rc = t4_register_uld(&tom_uld_info);
1121 	if (rc != 0)
1122 		t4_tom_mod_unload();
1123 
1124 	return (rc);
1125 }
1126 
1127 static void
1128 tom_uninit(struct adapter *sc, void *arg __unused)
1129 {
1130 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
1131 		return;
1132 
1133 	/* Try to free resources (works only if no port has IFCAP_TOE) */
1134 	if (uld_active(sc, ULD_TOM))
1135 		t4_deactivate_uld(sc, ULD_TOM);
1136 
1137 	end_synchronized_op(sc, 0);
1138 }
1139 
1140 static int
1141 t4_tom_mod_unload(void)
1142 {
1143 	t4_iterate(tom_uninit, NULL);
1144 
1145 	if (t4_unregister_uld(&tom_uld_info) == EBUSY)
1146 		return (EBUSY);
1147 
1148 	if (ifaddr_evhandler) {
1149 		EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler);
1150 		taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL);
1151 	}
1152 
1153 	return (0);
1154 }
1155 #endif	/* TCP_OFFLOAD */
1156 
1157 static int
1158 t4_tom_modevent(module_t mod, int cmd, void *arg)
1159 {
1160 	int rc = 0;
1161 
1162 #ifdef TCP_OFFLOAD
1163 	switch (cmd) {
1164 	case MOD_LOAD:
1165 		rc = t4_tom_mod_load();
1166 		break;
1167 
1168 	case MOD_UNLOAD:
1169 		rc = t4_tom_mod_unload();
1170 		break;
1171 
1172 	default:
1173 		rc = EINVAL;
1174 	}
1175 #else
1176 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
1177 	rc = EOPNOTSUPP;
1178 #endif
1179 	return (rc);
1180 }
1181 
1182 static moduledata_t t4_tom_moddata= {
1183 	"t4_tom",
1184 	t4_tom_modevent,
1185 	0
1186 };
1187 
1188 MODULE_VERSION(t4_tom, 1);
1189 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
1190 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
1191 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
1192