xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision 2282756519d9337dec783509a432fd6ddfd06683)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2012 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ratelimit.h"
36 
37 #include <sys/param.h>
38 #include <sys/types.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/ktr.h>
42 #include <sys/lock.h>
43 #include <sys/limits.h>
44 #include <sys/module.h>
45 #include <sys/protosw.h>
46 #include <sys/domain.h>
47 #include <sys/refcount.h>
48 #include <sys/rmlock.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/taskqueue.h>
52 #include <net/if.h>
53 #include <net/if_var.h>
54 #include <net/if_types.h>
55 #include <net/if_vlan_var.h>
56 #include <netinet/in.h>
57 #include <netinet/in_pcb.h>
58 #include <netinet/in_var.h>
59 #include <netinet/ip.h>
60 #include <netinet/ip6.h>
61 #include <netinet6/scope6_var.h>
62 #define TCPSTATES
63 #include <netinet/tcp_fsm.h>
64 #include <netinet/tcp_timer.h>
65 #include <netinet/tcp_var.h>
66 #include <netinet/toecore.h>
67 
68 #ifdef TCP_OFFLOAD
69 #include "common/common.h"
70 #include "common/t4_msg.h"
71 #include "common/t4_regs.h"
72 #include "common/t4_regs_values.h"
73 #include "common/t4_tcb.h"
74 #include "tom/t4_tom_l2t.h"
75 #include "tom/t4_tom.h"
76 #include "tom/t4_tls.h"
77 
78 static struct protosw toe_protosw;
79 static struct pr_usrreqs toe_usrreqs;
80 
81 static struct protosw toe6_protosw;
82 static struct pr_usrreqs toe6_usrreqs;
83 
84 /* Module ops */
85 static int t4_tom_mod_load(void);
86 static int t4_tom_mod_unload(void);
87 static int t4_tom_modevent(module_t, int, void *);
88 
89 /* ULD ops and helpers */
90 static int t4_tom_activate(struct adapter *);
91 static int t4_tom_deactivate(struct adapter *);
92 
93 static struct uld_info tom_uld_info = {
94 	.uld_id = ULD_TOM,
95 	.activate = t4_tom_activate,
96 	.deactivate = t4_tom_deactivate,
97 };
98 
99 static void release_offload_resources(struct toepcb *);
100 static int alloc_tid_tabs(struct tid_info *);
101 static void free_tid_tabs(struct tid_info *);
102 static int add_lip(struct adapter *, struct in6_addr *);
103 static int delete_lip(struct adapter *, struct in6_addr *);
104 static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *);
105 static void init_clip_table(struct adapter *, struct tom_data *);
106 static void update_clip(struct adapter *, void *);
107 static void t4_clip_task(void *, int);
108 static void update_clip_table(struct adapter *, struct tom_data *);
109 static void destroy_clip_table(struct adapter *, struct tom_data *);
110 static void free_tom_data(struct adapter *, struct tom_data *);
111 static void reclaim_wr_resources(void *, int);
112 
113 static int in6_ifaddr_gen;
114 static eventhandler_tag ifaddr_evhandler;
115 static struct timeout_task clip_task;
116 
117 struct toepcb *
118 alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags)
119 {
120 	struct port_info *pi = vi->pi;
121 	struct adapter *sc = pi->adapter;
122 	struct toepcb *toep;
123 	int tx_credits, txsd_total, len;
124 
125 	/*
126 	 * The firmware counts tx work request credits in units of 16 bytes
127 	 * each.  Reserve room for an ABORT_REQ so the driver never has to worry
128 	 * about tx credits if it wants to abort a connection.
129 	 */
130 	tx_credits = sc->params.ofldq_wr_cred;
131 	tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
132 
133 	/*
134 	 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
135 	 * immediate payload, and firmware counts tx work request credits in
136 	 * units of 16 byte.  Calculate the maximum work requests possible.
137 	 */
138 	txsd_total = tx_credits /
139 	    howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
140 
141 	KASSERT(txqid >= vi->first_ofld_txq &&
142 	    txqid < vi->first_ofld_txq + vi->nofldtxq,
143 	    ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi,
144 		vi->first_ofld_txq, vi->nofldtxq));
145 
146 	KASSERT(rxqid >= vi->first_ofld_rxq &&
147 	    rxqid < vi->first_ofld_rxq + vi->nofldrxq,
148 	    ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi,
149 		vi->first_ofld_rxq, vi->nofldrxq));
150 
151 	len = offsetof(struct toepcb, txsd) +
152 	    txsd_total * sizeof(struct ofld_tx_sdesc);
153 
154 	toep = malloc(len, M_CXGBE, M_ZERO | flags);
155 	if (toep == NULL)
156 		return (NULL);
157 
158 	refcount_init(&toep->refcount, 1);
159 	toep->td = sc->tom_softc;
160 	toep->vi = vi;
161 	toep->tc_idx = -1;
162 	toep->tx_total = tx_credits;
163 	toep->tx_credits = tx_credits;
164 	toep->ofld_txq = &sc->sge.ofld_txq[txqid];
165 	toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid];
166 	toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
167 	mbufq_init(&toep->ulp_pduq, INT_MAX);
168 	mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
169 	toep->txsd_total = txsd_total;
170 	toep->txsd_avail = txsd_total;
171 	toep->txsd_pidx = 0;
172 	toep->txsd_cidx = 0;
173 	aiotx_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 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
199 		ddp_uninit_toep(toep);
200 	tls_uninit_toep(toep);
201 	free(toep, M_CXGBE);
202 }
203 
204 /*
205  * Set up the socket for TCP offload.
206  */
207 void
208 offload_socket(struct socket *so, struct toepcb *toep)
209 {
210 	struct tom_data *td = toep->td;
211 	struct inpcb *inp = sotoinpcb(so);
212 	struct tcpcb *tp = intotcpcb(inp);
213 	struct sockbuf *sb;
214 
215 	INP_WLOCK_ASSERT(inp);
216 
217 	/* Update socket */
218 	sb = &so->so_snd;
219 	SOCKBUF_LOCK(sb);
220 	sb->sb_flags |= SB_NOCOALESCE;
221 	SOCKBUF_UNLOCK(sb);
222 	sb = &so->so_rcv;
223 	SOCKBUF_LOCK(sb);
224 	sb->sb_flags |= SB_NOCOALESCE;
225 	if (inp->inp_vflag & INP_IPV6)
226 		so->so_proto = &toe6_protosw;
227 	else
228 		so->so_proto = &toe_protosw;
229 	SOCKBUF_UNLOCK(sb);
230 
231 	/* Update TCP PCB */
232 	tp->tod = &td->tod;
233 	tp->t_toe = toep;
234 	tp->t_flags |= TF_TOE;
235 
236 	/* Install an extra hold on inp */
237 	toep->inp = inp;
238 	toep->flags |= TPF_ATTACHED;
239 	in_pcbref(inp);
240 
241 	/* Add the TOE PCB to the active list */
242 	mtx_lock(&td->toep_list_lock);
243 	TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
244 	mtx_unlock(&td->toep_list_lock);
245 }
246 
247 /* This is _not_ the normal way to "unoffload" a socket. */
248 void
249 undo_offload_socket(struct socket *so)
250 {
251 	struct inpcb *inp = sotoinpcb(so);
252 	struct tcpcb *tp = intotcpcb(inp);
253 	struct toepcb *toep = tp->t_toe;
254 	struct tom_data *td = toep->td;
255 	struct sockbuf *sb;
256 
257 	INP_WLOCK_ASSERT(inp);
258 
259 	sb = &so->so_snd;
260 	SOCKBUF_LOCK(sb);
261 	sb->sb_flags &= ~SB_NOCOALESCE;
262 	SOCKBUF_UNLOCK(sb);
263 	sb = &so->so_rcv;
264 	SOCKBUF_LOCK(sb);
265 	sb->sb_flags &= ~SB_NOCOALESCE;
266 	SOCKBUF_UNLOCK(sb);
267 
268 	tp->tod = NULL;
269 	tp->t_toe = NULL;
270 	tp->t_flags &= ~TF_TOE;
271 
272 	toep->inp = NULL;
273 	toep->flags &= ~TPF_ATTACHED;
274 	if (in_pcbrele_wlocked(inp))
275 		panic("%s: inp freed.", __func__);
276 
277 	mtx_lock(&td->toep_list_lock);
278 	TAILQ_REMOVE(&td->toep_list, toep, link);
279 	mtx_unlock(&td->toep_list_lock);
280 }
281 
282 static void
283 release_offload_resources(struct toepcb *toep)
284 {
285 	struct tom_data *td = toep->td;
286 	struct adapter *sc = td_adapter(td);
287 	int tid = toep->tid;
288 
289 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
290 	    ("%s: %p has CPL pending.", __func__, toep));
291 	KASSERT(!(toep->flags & TPF_ATTACHED),
292 	    ("%s: %p is still attached.", __func__, toep));
293 
294 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
295 	    __func__, toep, tid, toep->l2te, toep->ce);
296 
297 	/*
298 	 * These queues should have been emptied at approximately the same time
299 	 * that a normal connection's socket's so_snd would have been purged or
300 	 * drained.  Do _not_ clean up here.
301 	 */
302 	MPASS(mbufq_len(&toep->ulp_pduq) == 0);
303 	MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0);
304 #ifdef INVARIANTS
305 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
306 		ddp_assert_empty(toep);
307 #endif
308 
309 	if (toep->l2te)
310 		t4_l2t_release(toep->l2te);
311 
312 	if (tid >= 0) {
313 		remove_tid(sc, tid, toep->ce ? 2 : 1);
314 		release_tid(sc, tid, toep->ctrlq);
315 	}
316 
317 	if (toep->ce)
318 		release_lip(td, toep->ce);
319 
320 	if (toep->tc_idx != -1)
321 		t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->tc_idx);
322 
323 	mtx_lock(&td->toep_list_lock);
324 	TAILQ_REMOVE(&td->toep_list, toep, link);
325 	mtx_unlock(&td->toep_list_lock);
326 
327 	free_toepcb(toep);
328 }
329 
330 /*
331  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
332  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
333  * pending CPL) then it is time to release all resources tied to the toepcb.
334  *
335  * Also gets called when an offloaded active open fails and the TOM wants the
336  * kernel to take the TCP PCB back.
337  */
338 static void
339 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
340 {
341 #if defined(KTR) || defined(INVARIANTS)
342 	struct inpcb *inp = tp->t_inpcb;
343 #endif
344 	struct toepcb *toep = tp->t_toe;
345 
346 	INP_WLOCK_ASSERT(inp);
347 
348 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
349 	KASSERT(toep->flags & TPF_ATTACHED,
350 	    ("%s: not attached", __func__));
351 
352 #ifdef KTR
353 	if (tp->t_state == TCPS_SYN_SENT) {
354 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
355 		    __func__, toep->tid, toep, toep->flags, inp,
356 		    inp->inp_flags);
357 	} else {
358 		CTR6(KTR_CXGBE,
359 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
360 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
361 		    inp->inp_flags);
362 	}
363 #endif
364 
365 	tp->t_toe = NULL;
366 	tp->t_flags &= ~TF_TOE;
367 	toep->flags &= ~TPF_ATTACHED;
368 
369 	if (!(toep->flags & TPF_CPL_PENDING))
370 		release_offload_resources(toep);
371 }
372 
373 /*
374  * setsockopt handler.
375  */
376 static void
377 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
378 {
379 	struct adapter *sc = tod->tod_softc;
380 	struct toepcb *toep = tp->t_toe;
381 
382 	if (dir == SOPT_GET)
383 		return;
384 
385 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
386 
387 	switch (name) {
388 	case TCP_NODELAY:
389 		if (tp->t_state != TCPS_ESTABLISHED)
390 			break;
391 		t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
392 		    V_TF_NAGLE(1), V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1),
393 		    0, 0);
394 		break;
395 	default:
396 		break;
397 	}
398 }
399 
400 static inline int
401 get_tcb_bit(u_char *tcb, int bit)
402 {
403 	int ix, shift;
404 
405 	ix = 127 - (bit >> 3);
406 	shift = bit & 0x7;
407 
408 	return ((tcb[ix] >> shift) & 1);
409 }
410 
411 static inline uint64_t
412 get_tcb_bits(u_char *tcb, int hi, int lo)
413 {
414 	uint64_t rc = 0;
415 
416 	while (hi >= lo) {
417 		rc = (rc << 1) | get_tcb_bit(tcb, hi);
418 		--hi;
419 	}
420 
421 	return (rc);
422 }
423 
424 /*
425  * Called by the kernel to allow the TOE driver to "refine" values filled up in
426  * the tcp_info for an offloaded connection.
427  */
428 static void
429 t4_tcp_info(struct toedev *tod, struct tcpcb *tp, struct tcp_info *ti)
430 {
431 	int i, j, k, rc;
432 	struct adapter *sc = tod->tod_softc;
433 	struct toepcb *toep = tp->t_toe;
434 	uint32_t addr, v;
435 	uint32_t buf[TCB_SIZE / sizeof(uint32_t)];
436 	u_char *tcb, tmp;
437 
438 	INP_WLOCK_ASSERT(tp->t_inpcb);
439 	MPASS(ti != NULL);
440 
441 	addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + toep->tid * TCB_SIZE;
442 	rc = read_via_memwin(sc, 2, addr, &buf[0], TCB_SIZE);
443 	if (rc != 0)
444 		return;
445 
446 	tcb = (u_char *)&buf[0];
447 	for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) {
448 		for (k = 0; k < 16; k++) {
449 			tmp = tcb[i + k];
450 			tcb[i + k] = tcb[j + k];
451 			tcb[j + k] = tmp;
452 		}
453 	}
454 
455 	ti->tcpi_state = get_tcb_bits(tcb, 115, 112);
456 
457 	v = get_tcb_bits(tcb, 271, 256);
458 	ti->tcpi_rtt = tcp_ticks_to_us(sc, v);
459 
460 	v = get_tcb_bits(tcb, 287, 272);
461 	ti->tcpi_rttvar = tcp_ticks_to_us(sc, v);
462 
463 	ti->tcpi_snd_ssthresh = get_tcb_bits(tcb, 487, 460);
464 	ti->tcpi_snd_cwnd = get_tcb_bits(tcb, 459, 432);
465 	ti->tcpi_rcv_nxt = get_tcb_bits(tcb, 553, 522);
466 
467 	ti->tcpi_snd_nxt = get_tcb_bits(tcb, 319, 288) -
468 	    get_tcb_bits(tcb, 375, 348);
469 
470 	/* Receive window being advertised by us. */
471 	ti->tcpi_rcv_space = get_tcb_bits(tcb, 581, 554);
472 
473 	/* Send window ceiling. */
474 	v = get_tcb_bits(tcb, 159, 144) << get_tcb_bits(tcb, 131, 128);
475 	ti->tcpi_snd_wnd = min(v, ti->tcpi_snd_cwnd);
476 }
477 
478 /*
479  * The TOE driver will not receive any more CPLs for the tid associated with the
480  * toepcb; release the hold on the inpcb.
481  */
482 void
483 final_cpl_received(struct toepcb *toep)
484 {
485 	struct inpcb *inp = toep->inp;
486 
487 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
488 	INP_WLOCK_ASSERT(inp);
489 	KASSERT(toep->flags & TPF_CPL_PENDING,
490 	    ("%s: CPL not pending already?", __func__));
491 
492 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
493 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
494 
495 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
496 		release_ddp_resources(toep);
497 	toep->inp = NULL;
498 	toep->flags &= ~TPF_CPL_PENDING;
499 	mbufq_drain(&toep->ulp_pdu_reclaimq);
500 
501 	if (!(toep->flags & TPF_ATTACHED))
502 		release_offload_resources(toep);
503 
504 	if (!in_pcbrele_wlocked(inp))
505 		INP_WUNLOCK(inp);
506 }
507 
508 void
509 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
510 {
511 	struct tid_info *t = &sc->tids;
512 
513 	MPASS(tid >= t->tid_base);
514 	MPASS(tid - t->tid_base < t->ntids);
515 
516 	t->tid_tab[tid - t->tid_base] = ctx;
517 	atomic_add_int(&t->tids_in_use, ntids);
518 }
519 
520 void *
521 lookup_tid(struct adapter *sc, int tid)
522 {
523 	struct tid_info *t = &sc->tids;
524 
525 	return (t->tid_tab[tid - t->tid_base]);
526 }
527 
528 void
529 update_tid(struct adapter *sc, int tid, void *ctx)
530 {
531 	struct tid_info *t = &sc->tids;
532 
533 	t->tid_tab[tid - t->tid_base] = ctx;
534 }
535 
536 void
537 remove_tid(struct adapter *sc, int tid, int ntids)
538 {
539 	struct tid_info *t = &sc->tids;
540 
541 	t->tid_tab[tid - t->tid_base] = NULL;
542 	atomic_subtract_int(&t->tids_in_use, ntids);
543 }
544 
545 /*
546  * What mtu_idx to use, given a 4-tuple.  Note that both s->mss and tcp_mssopt
547  * have the MSS that we should advertise in our SYN.  Advertised MSS doesn't
548  * account for any TCP options so the effective MSS (only payload, no headers or
549  * options) could be different.  We fill up tp->t_maxseg with the effective MSS
550  * at the end of the 3-way handshake.
551  */
552 int
553 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc,
554     struct offload_settings *s)
555 {
556 	unsigned short *mtus = &sc->params.mtus[0];
557 	int i, mss, mtu;
558 
559 	MPASS(inc != NULL);
560 
561 	mss = s->mss > 0 ? s->mss : tcp_mssopt(inc);
562 	if (inc->inc_flags & INC_ISIPV6)
563 		mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
564 	else
565 		mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr);
566 
567 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++)
568 		continue;
569 
570 	return (i);
571 }
572 
573 /*
574  * Determine the receive window size for a socket.
575  */
576 u_long
577 select_rcv_wnd(struct socket *so)
578 {
579 	unsigned long wnd;
580 
581 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
582 
583 	wnd = sbspace(&so->so_rcv);
584 	if (wnd < MIN_RCV_WND)
585 		wnd = MIN_RCV_WND;
586 
587 	return min(wnd, MAX_RCV_WND);
588 }
589 
590 int
591 select_rcv_wscale(void)
592 {
593 	int wscale = 0;
594 	unsigned long space = sb_max;
595 
596 	if (space > MAX_RCV_WND)
597 		space = MAX_RCV_WND;
598 
599 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
600 		wscale++;
601 
602 	return (wscale);
603 }
604 
605 /*
606  * socket so could be a listening socket too.
607  */
608 uint64_t
609 calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e,
610     int mtu_idx, int rscale, int rx_credits, int ulp_mode,
611     struct offload_settings *s)
612 {
613 	int keepalive;
614 	uint64_t opt0;
615 
616 	MPASS(so != NULL);
617 	MPASS(vi != NULL);
618 	KASSERT(rx_credits <= M_RCV_BUFSIZ,
619 	    ("%s: rcv_bufsiz too high", __func__));
620 
621 	opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) |
622 	    V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits) |
623 	    V_L2T_IDX(e->idx) | V_SMAC_SEL(vi->smt_idx) |
624 	    V_TX_CHAN(vi->pi->tx_chan);
625 
626 	keepalive = tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE;
627 	opt0 |= V_KEEP_ALIVE(keepalive != 0);
628 
629 	if (s->nagle < 0) {
630 		struct inpcb *inp = sotoinpcb(so);
631 		struct tcpcb *tp = intotcpcb(inp);
632 
633 		opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0);
634 	} else
635 		opt0 |= V_NAGLE(s->nagle != 0);
636 
637 	return htobe64(opt0);
638 }
639 
640 uint64_t
641 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
642 {
643 	struct adapter *sc = vi->pi->adapter;
644 	struct tp_params *tp = &sc->params.tp;
645 	uint16_t viid = vi->viid;
646 	uint64_t ntuple = 0;
647 
648 	/*
649 	 * Initialize each of the fields which we care about which are present
650 	 * in the Compressed Filter Tuple.
651 	 */
652 	if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE)
653 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
654 
655 	if (tp->port_shift >= 0)
656 		ntuple |= (uint64_t)e->lport << tp->port_shift;
657 
658 	if (tp->protocol_shift >= 0)
659 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
660 
661 	if (tp->vnic_shift >= 0 && tp->ingress_config & F_VNIC) {
662 		uint32_t vf = G_FW_VIID_VIN(viid);
663 		uint32_t pf = G_FW_VIID_PFN(viid);
664 		uint32_t vld = G_FW_VIID_VIVLD(viid);
665 
666 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) |
667 		    V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift;
668 	}
669 
670 	if (is_t4(sc))
671 		return (htobe32((uint32_t)ntuple));
672 	else
673 		return (htobe64(V_FILTER_TUPLE(ntuple)));
674 }
675 
676 static int
677 is_tls_sock(struct socket *so, struct adapter *sc)
678 {
679 	struct inpcb *inp = sotoinpcb(so);
680 	int i, rc;
681 
682 	/* XXX: Eventually add a SO_WANT_TLS socket option perhaps? */
683 	rc = 0;
684 	ADAPTER_LOCK(sc);
685 	for (i = 0; i < sc->tt.num_tls_rx_ports; i++) {
686 		if (inp->inp_lport == htons(sc->tt.tls_rx_ports[i]) ||
687 		    inp->inp_fport == htons(sc->tt.tls_rx_ports[i])) {
688 			rc = 1;
689 			break;
690 		}
691 	}
692 	ADAPTER_UNLOCK(sc);
693 	return (rc);
694 }
695 
696 int
697 select_ulp_mode(struct socket *so, struct adapter *sc,
698     struct offload_settings *s)
699 {
700 
701 	if (can_tls_offload(sc) &&
702 	    (s->tls > 0 || (s->tls < 0 && is_tls_sock(so, sc))))
703 		return (ULP_MODE_TLS);
704 	else if (s->ddp > 0 ||
705 	    (s->ddp < 0 && sc->tt.ddp && (so->so_options & SO_NO_DDP) == 0))
706 		return (ULP_MODE_TCPDDP);
707 	else
708 		return (ULP_MODE_NONE);
709 }
710 
711 void
712 set_ulp_mode(struct toepcb *toep, int ulp_mode)
713 {
714 
715 	CTR4(KTR_CXGBE, "%s: toep %p (tid %d) ulp_mode %d",
716 	    __func__, toep, toep->tid, ulp_mode);
717 	toep->ulp_mode = ulp_mode;
718 	tls_init_toep(toep);
719 	if (toep->ulp_mode == ULP_MODE_TCPDDP)
720 		ddp_init_toep(toep);
721 }
722 
723 int
724 negative_advice(int status)
725 {
726 
727 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
728 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
729 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
730 }
731 
732 static int
733 alloc_tid_tab(struct tid_info *t, int flags)
734 {
735 
736 	MPASS(t->ntids > 0);
737 	MPASS(t->tid_tab == NULL);
738 
739 	t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE,
740 	    M_ZERO | flags);
741 	if (t->tid_tab == NULL)
742 		return (ENOMEM);
743 	atomic_store_rel_int(&t->tids_in_use, 0);
744 
745 	return (0);
746 }
747 
748 static void
749 free_tid_tab(struct tid_info *t)
750 {
751 
752 	KASSERT(t->tids_in_use == 0,
753 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
754 
755 	free(t->tid_tab, M_CXGBE);
756 	t->tid_tab = NULL;
757 }
758 
759 static int
760 alloc_stid_tab(struct tid_info *t, int flags)
761 {
762 
763 	MPASS(t->nstids > 0);
764 	MPASS(t->stid_tab == NULL);
765 
766 	t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE,
767 	    M_ZERO | flags);
768 	if (t->stid_tab == NULL)
769 		return (ENOMEM);
770 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
771 	t->stids_in_use = 0;
772 	TAILQ_INIT(&t->stids);
773 	t->nstids_free_head = t->nstids;
774 
775 	return (0);
776 }
777 
778 static void
779 free_stid_tab(struct tid_info *t)
780 {
781 
782 	KASSERT(t->stids_in_use == 0,
783 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
784 
785 	if (mtx_initialized(&t->stid_lock))
786 		mtx_destroy(&t->stid_lock);
787 	free(t->stid_tab, M_CXGBE);
788 	t->stid_tab = NULL;
789 }
790 
791 static void
792 free_tid_tabs(struct tid_info *t)
793 {
794 
795 	free_tid_tab(t);
796 	free_atid_tab(t);
797 	free_stid_tab(t);
798 }
799 
800 static int
801 alloc_tid_tabs(struct tid_info *t)
802 {
803 	int rc;
804 
805 	rc = alloc_tid_tab(t, M_NOWAIT);
806 	if (rc != 0)
807 		goto failed;
808 
809 	rc = alloc_atid_tab(t, M_NOWAIT);
810 	if (rc != 0)
811 		goto failed;
812 
813 	rc = alloc_stid_tab(t, M_NOWAIT);
814 	if (rc != 0)
815 		goto failed;
816 
817 	return (0);
818 failed:
819 	free_tid_tabs(t);
820 	return (rc);
821 }
822 
823 static int
824 add_lip(struct adapter *sc, struct in6_addr *lip)
825 {
826         struct fw_clip_cmd c;
827 
828 	ASSERT_SYNCHRONIZED_OP(sc);
829 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
830 
831         memset(&c, 0, sizeof(c));
832 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
833 	    F_FW_CMD_WRITE);
834         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c));
835         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
836         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
837 
838 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
839 }
840 
841 static int
842 delete_lip(struct adapter *sc, struct in6_addr *lip)
843 {
844 	struct fw_clip_cmd c;
845 
846 	ASSERT_SYNCHRONIZED_OP(sc);
847 	/* mtx_assert(&td->clip_table_lock, MA_OWNED); */
848 
849 	memset(&c, 0, sizeof(c));
850 	c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST |
851 	    F_FW_CMD_READ);
852         c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c));
853         c.ip_hi = *(uint64_t *)&lip->s6_addr[0];
854         c.ip_lo = *(uint64_t *)&lip->s6_addr[8];
855 
856 	return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c));
857 }
858 
859 static struct clip_entry *
860 search_lip(struct tom_data *td, struct in6_addr *lip)
861 {
862 	struct clip_entry *ce;
863 
864 	mtx_assert(&td->clip_table_lock, MA_OWNED);
865 
866 	TAILQ_FOREACH(ce, &td->clip_table, link) {
867 		if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
868 			return (ce);
869 	}
870 
871 	return (NULL);
872 }
873 
874 struct clip_entry *
875 hold_lip(struct tom_data *td, struct in6_addr *lip, struct clip_entry *ce)
876 {
877 
878 	mtx_lock(&td->clip_table_lock);
879 	if (ce == NULL)
880 		ce = search_lip(td, lip);
881 	if (ce != NULL)
882 		ce->refcount++;
883 	mtx_unlock(&td->clip_table_lock);
884 
885 	return (ce);
886 }
887 
888 void
889 release_lip(struct tom_data *td, struct clip_entry *ce)
890 {
891 
892 	mtx_lock(&td->clip_table_lock);
893 	KASSERT(search_lip(td, &ce->lip) == ce,
894 	    ("%s: CLIP entry %p p not in CLIP table.", __func__, ce));
895 	KASSERT(ce->refcount > 0,
896 	    ("%s: CLIP entry %p has refcount 0", __func__, ce));
897 	--ce->refcount;
898 	mtx_unlock(&td->clip_table_lock);
899 }
900 
901 static void
902 init_clip_table(struct adapter *sc, struct tom_data *td)
903 {
904 
905 	ASSERT_SYNCHRONIZED_OP(sc);
906 
907 	mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF);
908 	TAILQ_INIT(&td->clip_table);
909 	td->clip_gen = -1;
910 
911 	update_clip_table(sc, td);
912 }
913 
914 static void
915 update_clip(struct adapter *sc, void *arg __unused)
916 {
917 
918 	if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc"))
919 		return;
920 
921 	if (uld_active(sc, ULD_TOM))
922 		update_clip_table(sc, sc->tom_softc);
923 
924 	end_synchronized_op(sc, LOCK_HELD);
925 }
926 
927 static void
928 t4_clip_task(void *arg, int count)
929 {
930 
931 	t4_iterate(update_clip, NULL);
932 }
933 
934 static void
935 update_clip_table(struct adapter *sc, struct tom_data *td)
936 {
937 	struct rm_priotracker in6_ifa_tracker;
938 	struct in6_ifaddr *ia;
939 	struct in6_addr *lip, tlip;
940 	struct clip_head stale;
941 	struct clip_entry *ce, *ce_temp;
942 	struct vi_info *vi;
943 	int rc, gen, i, j;
944 	uintptr_t last_vnet;
945 
946 	ASSERT_SYNCHRONIZED_OP(sc);
947 
948 	IN6_IFADDR_RLOCK(&in6_ifa_tracker);
949 	mtx_lock(&td->clip_table_lock);
950 
951 	gen = atomic_load_acq_int(&in6_ifaddr_gen);
952 	if (gen == td->clip_gen)
953 		goto done;
954 
955 	TAILQ_INIT(&stale);
956 	TAILQ_CONCAT(&stale, &td->clip_table, link);
957 
958 	/*
959 	 * last_vnet optimizes the common cases where all if_vnet = NULL (no
960 	 * VIMAGE) or all if_vnet = vnet0.
961 	 */
962 	last_vnet = (uintptr_t)(-1);
963 	for_each_port(sc, i)
964 	for_each_vi(sc->port[i], j, vi) {
965 		if (last_vnet == (uintptr_t)vi->ifp->if_vnet)
966 			continue;
967 
968 		/* XXX: races with if_vmove */
969 		CURVNET_SET(vi->ifp->if_vnet);
970 		CK_STAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) {
971 			lip = &ia->ia_addr.sin6_addr;
972 
973 			KASSERT(!IN6_IS_ADDR_MULTICAST(lip),
974 			    ("%s: mcast address in in6_ifaddr list", __func__));
975 
976 			if (IN6_IS_ADDR_LOOPBACK(lip))
977 				continue;
978 			if (IN6_IS_SCOPE_EMBED(lip)) {
979 				/* Remove the embedded scope */
980 				tlip = *lip;
981 				lip = &tlip;
982 				in6_clearscope(lip);
983 			}
984 			/*
985 			 * XXX: how to weed out the link local address for the
986 			 * loopback interface?  It's fe80::1 usually (always?).
987 			 */
988 
989 			/*
990 			 * If it's in the main list then we already know it's
991 			 * not stale.
992 			 */
993 			TAILQ_FOREACH(ce, &td->clip_table, link) {
994 				if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip))
995 					goto next;
996 			}
997 
998 			/*
999 			 * If it's in the stale list we should move it to the
1000 			 * main list.
1001 			 */
1002 			TAILQ_FOREACH(ce, &stale, link) {
1003 				if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) {
1004 					TAILQ_REMOVE(&stale, ce, link);
1005 					TAILQ_INSERT_TAIL(&td->clip_table, ce,
1006 					    link);
1007 					goto next;
1008 				}
1009 			}
1010 
1011 			/* A new IP6 address; add it to the CLIP table */
1012 			ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT);
1013 			memcpy(&ce->lip, lip, sizeof(ce->lip));
1014 			ce->refcount = 0;
1015 			rc = add_lip(sc, lip);
1016 			if (rc == 0)
1017 				TAILQ_INSERT_TAIL(&td->clip_table, ce, link);
1018 			else {
1019 				char ip[INET6_ADDRSTRLEN];
1020 
1021 				inet_ntop(AF_INET6, &ce->lip, &ip[0],
1022 				    sizeof(ip));
1023 				log(LOG_ERR, "%s: could not add %s (%d)\n",
1024 				    __func__, ip, rc);
1025 				free(ce, M_CXGBE);
1026 			}
1027 next:
1028 			continue;
1029 		}
1030 		CURVNET_RESTORE();
1031 		last_vnet = (uintptr_t)vi->ifp->if_vnet;
1032 	}
1033 
1034 	/*
1035 	 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are
1036 	 * no longer referenced by the driver.
1037 	 */
1038 	TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) {
1039 		if (ce->refcount == 0) {
1040 			rc = delete_lip(sc, &ce->lip);
1041 			if (rc == 0) {
1042 				TAILQ_REMOVE(&stale, ce, link);
1043 				free(ce, M_CXGBE);
1044 			} else {
1045 				char ip[INET6_ADDRSTRLEN];
1046 
1047 				inet_ntop(AF_INET6, &ce->lip, &ip[0],
1048 				    sizeof(ip));
1049 				log(LOG_ERR, "%s: could not delete %s (%d)\n",
1050 				    __func__, ip, rc);
1051 			}
1052 		}
1053 	}
1054 	/* The ones that are still referenced need to stay in the CLIP table */
1055 	TAILQ_CONCAT(&td->clip_table, &stale, link);
1056 
1057 	td->clip_gen = gen;
1058 done:
1059 	mtx_unlock(&td->clip_table_lock);
1060 	IN6_IFADDR_RUNLOCK(&in6_ifa_tracker);
1061 }
1062 
1063 static void
1064 destroy_clip_table(struct adapter *sc, struct tom_data *td)
1065 {
1066 	struct clip_entry *ce, *ce_temp;
1067 
1068 	if (mtx_initialized(&td->clip_table_lock)) {
1069 		mtx_lock(&td->clip_table_lock);
1070 		TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) {
1071 			KASSERT(ce->refcount == 0,
1072 			    ("%s: CLIP entry %p still in use (%d)", __func__,
1073 			    ce, ce->refcount));
1074 			TAILQ_REMOVE(&td->clip_table, ce, link);
1075 			delete_lip(sc, &ce->lip);
1076 			free(ce, M_CXGBE);
1077 		}
1078 		mtx_unlock(&td->clip_table_lock);
1079 		mtx_destroy(&td->clip_table_lock);
1080 	}
1081 }
1082 
1083 static void
1084 free_tom_data(struct adapter *sc, struct tom_data *td)
1085 {
1086 
1087 	ASSERT_SYNCHRONIZED_OP(sc);
1088 
1089 	KASSERT(TAILQ_EMPTY(&td->toep_list),
1090 	    ("%s: TOE PCB list is not empty.", __func__));
1091 	KASSERT(td->lctx_count == 0,
1092 	    ("%s: lctx hash table is not empty.", __func__));
1093 
1094 	tls_free_kmap(td);
1095 	t4_free_ppod_region(&td->pr);
1096 	destroy_clip_table(sc, td);
1097 
1098 	if (td->listen_mask != 0)
1099 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
1100 
1101 	if (mtx_initialized(&td->unsent_wr_lock))
1102 		mtx_destroy(&td->unsent_wr_lock);
1103 	if (mtx_initialized(&td->lctx_hash_lock))
1104 		mtx_destroy(&td->lctx_hash_lock);
1105 	if (mtx_initialized(&td->toep_list_lock))
1106 		mtx_destroy(&td->toep_list_lock);
1107 
1108 	free_tid_tabs(&sc->tids);
1109 	free(td, M_CXGBE);
1110 }
1111 
1112 static char *
1113 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen,
1114     int *buflen)
1115 {
1116 	char *pkt;
1117 	struct tcphdr *th;
1118 	int ipv6, len;
1119 	const int maxlen =
1120 	    max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) +
1121 	    max(sizeof(struct ip), sizeof(struct ip6_hdr)) +
1122 	    sizeof(struct tcphdr);
1123 
1124 	MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN);
1125 
1126 	pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT);
1127 	if (pkt == NULL)
1128 		return (NULL);
1129 
1130 	ipv6 = inp->inp_vflag & INP_IPV6;
1131 	len = 0;
1132 
1133 	if (vtag == 0xffff) {
1134 		struct ether_header *eh = (void *)pkt;
1135 
1136 		if (ipv6)
1137 			eh->ether_type = htons(ETHERTYPE_IPV6);
1138 		else
1139 			eh->ether_type = htons(ETHERTYPE_IP);
1140 
1141 		len += sizeof(*eh);
1142 	} else {
1143 		struct ether_vlan_header *evh = (void *)pkt;
1144 
1145 		evh->evl_encap_proto = htons(ETHERTYPE_VLAN);
1146 		evh->evl_tag = htons(vtag);
1147 		if (ipv6)
1148 			evh->evl_proto = htons(ETHERTYPE_IPV6);
1149 		else
1150 			evh->evl_proto = htons(ETHERTYPE_IP);
1151 
1152 		len += sizeof(*evh);
1153 	}
1154 
1155 	if (ipv6) {
1156 		struct ip6_hdr *ip6 = (void *)&pkt[len];
1157 
1158 		ip6->ip6_vfc = IPV6_VERSION;
1159 		ip6->ip6_plen = htons(sizeof(struct tcphdr));
1160 		ip6->ip6_nxt = IPPROTO_TCP;
1161 		if (open_type == OPEN_TYPE_ACTIVE) {
1162 			ip6->ip6_src = inp->in6p_laddr;
1163 			ip6->ip6_dst = inp->in6p_faddr;
1164 		} else if (open_type == OPEN_TYPE_LISTEN) {
1165 			ip6->ip6_src = inp->in6p_laddr;
1166 			ip6->ip6_dst = ip6->ip6_src;
1167 		}
1168 
1169 		len += sizeof(*ip6);
1170 	} else {
1171 		struct ip *ip = (void *)&pkt[len];
1172 
1173 		ip->ip_v = IPVERSION;
1174 		ip->ip_hl = sizeof(*ip) >> 2;
1175 		ip->ip_tos = inp->inp_ip_tos;
1176 		ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr));
1177 		ip->ip_ttl = inp->inp_ip_ttl;
1178 		ip->ip_p = IPPROTO_TCP;
1179 		if (open_type == OPEN_TYPE_ACTIVE) {
1180 			ip->ip_src = inp->inp_laddr;
1181 			ip->ip_dst = inp->inp_faddr;
1182 		} else if (open_type == OPEN_TYPE_LISTEN) {
1183 			ip->ip_src = inp->inp_laddr;
1184 			ip->ip_dst = ip->ip_src;
1185 		}
1186 
1187 		len += sizeof(*ip);
1188 	}
1189 
1190 	th = (void *)&pkt[len];
1191 	if (open_type == OPEN_TYPE_ACTIVE) {
1192 		th->th_sport = inp->inp_lport;	/* network byte order already */
1193 		th->th_dport = inp->inp_fport;	/* ditto */
1194 	} else if (open_type == OPEN_TYPE_LISTEN) {
1195 		th->th_sport = inp->inp_lport;	/* network byte order already */
1196 		th->th_dport = th->th_sport;
1197 	}
1198 	len += sizeof(th);
1199 
1200 	*pktlen = *buflen = len;
1201 	return (pkt);
1202 }
1203 
1204 const struct offload_settings *
1205 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m,
1206     uint16_t vtag, struct inpcb *inp)
1207 {
1208 	const struct t4_offload_policy *op;
1209 	char *pkt;
1210 	struct offload_rule *r;
1211 	int i, matched, pktlen, buflen;
1212 	static const struct offload_settings allow_offloading_settings = {
1213 		.offload = 1,
1214 		.rx_coalesce = -1,
1215 		.cong_algo = -1,
1216 		.sched_class = -1,
1217 		.tstamp = -1,
1218 		.sack = -1,
1219 		.nagle = -1,
1220 		.ecn = -1,
1221 		.ddp = -1,
1222 		.tls = -1,
1223 		.txq = -1,
1224 		.rxq = -1,
1225 		.mss = -1,
1226 	};
1227 	static const struct offload_settings disallow_offloading_settings = {
1228 		.offload = 0,
1229 		/* rest is irrelevant when offload is off. */
1230 	};
1231 
1232 	rw_assert(&sc->policy_lock, RA_LOCKED);
1233 
1234 	/*
1235 	 * If there's no Connection Offloading Policy attached to the device
1236 	 * then we need to return a default static policy.  If
1237 	 * "cop_managed_offloading" is true, then we need to disallow
1238 	 * offloading until a COP is attached to the device.  Otherwise we
1239 	 * allow offloading ...
1240 	 */
1241 	op = sc->policy;
1242 	if (op == NULL) {
1243 		if (sc->tt.cop_managed_offloading)
1244 			return (&disallow_offloading_settings);
1245 		else
1246 			return (&allow_offloading_settings);
1247 	}
1248 
1249 	switch (open_type) {
1250 	case OPEN_TYPE_ACTIVE:
1251 	case OPEN_TYPE_LISTEN:
1252 		pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen);
1253 		break;
1254 	case OPEN_TYPE_PASSIVE:
1255 		MPASS(m != NULL);
1256 		pkt = mtod(m, char *);
1257 		MPASS(*pkt == CPL_PASS_ACCEPT_REQ);
1258 		pkt += sizeof(struct cpl_pass_accept_req);
1259 		pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req);
1260 		buflen = m->m_len - sizeof(struct cpl_pass_accept_req);
1261 		break;
1262 	default:
1263 		MPASS(0);
1264 		return (&disallow_offloading_settings);
1265 	}
1266 
1267 	if (pkt == NULL || pktlen == 0 || buflen == 0)
1268 		return (&disallow_offloading_settings);
1269 
1270 	r = &op->rule[0];
1271 	for (i = 0; i < op->nrules; i++, r++) {
1272 		if (r->open_type != open_type &&
1273 		    r->open_type != OPEN_TYPE_DONTCARE) {
1274 			continue;
1275 		}
1276 		matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen);
1277 		if (matched)
1278 			break;
1279 	}
1280 
1281 	if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN)
1282 		free(pkt, M_CXGBE);
1283 
1284 	return (matched ? &r->settings : &disallow_offloading_settings);
1285 }
1286 
1287 static void
1288 reclaim_wr_resources(void *arg, int count)
1289 {
1290 	struct tom_data *td = arg;
1291 	STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
1292 	struct cpl_act_open_req *cpl;
1293 	u_int opcode, atid;
1294 	struct wrqe *wr;
1295 	struct adapter *sc;
1296 
1297 	mtx_lock(&td->unsent_wr_lock);
1298 	STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
1299 	mtx_unlock(&td->unsent_wr_lock);
1300 
1301 	while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
1302 		STAILQ_REMOVE_HEAD(&twr_list, link);
1303 
1304 		cpl = wrtod(wr);
1305 		opcode = GET_OPCODE(cpl);
1306 
1307 		switch (opcode) {
1308 		case CPL_ACT_OPEN_REQ:
1309 		case CPL_ACT_OPEN_REQ6:
1310 			atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1311 			sc = td_adapter(td);
1312 
1313 			CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1314 			act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1315 			free(wr, M_CXGBE);
1316 			break;
1317 		default:
1318 			log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1319 			    "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1320 			/* WR not freed here; go look at it with a debugger.  */
1321 		}
1322 	}
1323 }
1324 
1325 /*
1326  * Ground control to Major TOM
1327  * Commencing countdown, engines on
1328  */
1329 static int
1330 t4_tom_activate(struct adapter *sc)
1331 {
1332 	struct tom_data *td;
1333 	struct toedev *tod;
1334 	struct vi_info *vi;
1335 	int i, rc, v;
1336 
1337 	ASSERT_SYNCHRONIZED_OP(sc);
1338 
1339 	/* per-adapter softc for TOM */
1340 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1341 	if (td == NULL)
1342 		return (ENOMEM);
1343 
1344 	/* List of TOE PCBs and associated lock */
1345 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1346 	TAILQ_INIT(&td->toep_list);
1347 
1348 	/* Listen context */
1349 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1350 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1351 	    &td->listen_mask, HASH_NOWAIT);
1352 
1353 	/* List of WRs for which L2 resolution failed */
1354 	mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1355 	STAILQ_INIT(&td->unsent_wr_list);
1356 	TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1357 
1358 	/* TID tables */
1359 	rc = alloc_tid_tabs(&sc->tids);
1360 	if (rc != 0)
1361 		goto done;
1362 
1363 	rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1364 	    t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1365 	if (rc != 0)
1366 		goto done;
1367 	t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1368 	    V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1369 
1370 	/* CLIP table for IPv6 offload */
1371 	init_clip_table(sc, td);
1372 
1373 	if (sc->vres.key.size != 0) {
1374 		rc = tls_init_kmap(sc, td);
1375 		if (rc != 0)
1376 			goto done;
1377 	}
1378 
1379 	/* toedev ops */
1380 	tod = &td->tod;
1381 	init_toedev(tod);
1382 	tod->tod_softc = sc;
1383 	tod->tod_connect = t4_connect;
1384 	tod->tod_listen_start = t4_listen_start;
1385 	tod->tod_listen_stop = t4_listen_stop;
1386 	tod->tod_rcvd = t4_rcvd;
1387 	tod->tod_output = t4_tod_output;
1388 	tod->tod_send_rst = t4_send_rst;
1389 	tod->tod_send_fin = t4_send_fin;
1390 	tod->tod_pcb_detach = t4_pcb_detach;
1391 	tod->tod_l2_update = t4_l2_update;
1392 	tod->tod_syncache_added = t4_syncache_added;
1393 	tod->tod_syncache_removed = t4_syncache_removed;
1394 	tod->tod_syncache_respond = t4_syncache_respond;
1395 	tod->tod_offload_socket = t4_offload_socket;
1396 	tod->tod_ctloutput = t4_ctloutput;
1397 	tod->tod_tcp_info = t4_tcp_info;
1398 
1399 	for_each_port(sc, i) {
1400 		for_each_vi(sc->port[i], v, vi) {
1401 			TOEDEV(vi->ifp) = &td->tod;
1402 		}
1403 	}
1404 
1405 	sc->tom_softc = td;
1406 	register_toedev(sc->tom_softc);
1407 
1408 done:
1409 	if (rc != 0)
1410 		free_tom_data(sc, td);
1411 	return (rc);
1412 }
1413 
1414 static int
1415 t4_tom_deactivate(struct adapter *sc)
1416 {
1417 	int rc = 0;
1418 	struct tom_data *td = sc->tom_softc;
1419 
1420 	ASSERT_SYNCHRONIZED_OP(sc);
1421 
1422 	if (td == NULL)
1423 		return (0);	/* XXX. KASSERT? */
1424 
1425 	if (sc->offload_map != 0)
1426 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1427 
1428 	if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1429 		return (EBUSY);	/* both iWARP and iSCSI rely on the TOE. */
1430 
1431 	mtx_lock(&td->toep_list_lock);
1432 	if (!TAILQ_EMPTY(&td->toep_list))
1433 		rc = EBUSY;
1434 	mtx_unlock(&td->toep_list_lock);
1435 
1436 	mtx_lock(&td->lctx_hash_lock);
1437 	if (td->lctx_count > 0)
1438 		rc = EBUSY;
1439 	mtx_unlock(&td->lctx_hash_lock);
1440 
1441 	taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1442 	mtx_lock(&td->unsent_wr_lock);
1443 	if (!STAILQ_EMPTY(&td->unsent_wr_list))
1444 		rc = EBUSY;
1445 	mtx_unlock(&td->unsent_wr_lock);
1446 
1447 	if (rc == 0) {
1448 		unregister_toedev(sc->tom_softc);
1449 		free_tom_data(sc, td);
1450 		sc->tom_softc = NULL;
1451 	}
1452 
1453 	return (rc);
1454 }
1455 
1456 static void
1457 t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp)
1458 {
1459 
1460 	atomic_add_rel_int(&in6_ifaddr_gen, 1);
1461 	taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4);
1462 }
1463 
1464 static int
1465 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1466 {
1467 	struct tcpcb *tp = so_sototcpcb(so);
1468 	struct toepcb *toep = tp->t_toe;
1469 	int error;
1470 
1471 	if (toep->ulp_mode == ULP_MODE_TCPDDP) {
1472 		error = t4_aio_queue_ddp(so, job);
1473 		if (error != EOPNOTSUPP)
1474 			return (error);
1475 	}
1476 
1477 	return (t4_aio_queue_aiotx(so, job));
1478 }
1479 
1480 static int
1481 t4_ctloutput_tom(struct socket *so, struct sockopt *sopt)
1482 {
1483 
1484 	if (sopt->sopt_level != IPPROTO_TCP)
1485 		return (tcp_ctloutput(so, sopt));
1486 
1487 	switch (sopt->sopt_name) {
1488 	case TCP_TLSOM_SET_TLS_CONTEXT:
1489 	case TCP_TLSOM_GET_TLS_TOM:
1490 	case TCP_TLSOM_CLR_TLS_TOM:
1491 	case TCP_TLSOM_CLR_QUIES:
1492 		return (t4_ctloutput_tls(so, sopt));
1493 	default:
1494 		return (tcp_ctloutput(so, sopt));
1495 	}
1496 }
1497 
1498 static int
1499 t4_tom_mod_load(void)
1500 {
1501 	struct protosw *tcp_protosw, *tcp6_protosw;
1502 
1503 	/* CPL handlers */
1504 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2,
1505 	    CPL_COOKIE_TOM);
1506 	t4_init_connect_cpl_handlers();
1507 	t4_init_listen_cpl_handlers();
1508 	t4_init_cpl_io_handlers();
1509 
1510 	t4_ddp_mod_load();
1511 	t4_tls_mod_load();
1512 
1513 	tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM);
1514 	if (tcp_protosw == NULL)
1515 		return (ENOPROTOOPT);
1516 	bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw));
1517 	bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs));
1518 	toe_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1519 	toe_protosw.pr_ctloutput = t4_ctloutput_tom;
1520 	toe_protosw.pr_usrreqs = &toe_usrreqs;
1521 
1522 	tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM);
1523 	if (tcp6_protosw == NULL)
1524 		return (ENOPROTOOPT);
1525 	bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
1526 	bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs));
1527 	toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1528 	toe6_protosw.pr_ctloutput = t4_ctloutput_tom;
1529 	toe6_protosw.pr_usrreqs = &toe6_usrreqs;
1530 
1531 	TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL);
1532 	ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event,
1533 	    t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY);
1534 
1535 	return (t4_register_uld(&tom_uld_info));
1536 }
1537 
1538 static void
1539 tom_uninit(struct adapter *sc, void *arg __unused)
1540 {
1541 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
1542 		return;
1543 
1544 	/* Try to free resources (works only if no port has IFCAP_TOE) */
1545 	if (uld_active(sc, ULD_TOM))
1546 		t4_deactivate_uld(sc, ULD_TOM);
1547 
1548 	end_synchronized_op(sc, 0);
1549 }
1550 
1551 static int
1552 t4_tom_mod_unload(void)
1553 {
1554 	t4_iterate(tom_uninit, NULL);
1555 
1556 	if (t4_unregister_uld(&tom_uld_info) == EBUSY)
1557 		return (EBUSY);
1558 
1559 	if (ifaddr_evhandler) {
1560 		EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler);
1561 		taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL);
1562 	}
1563 
1564 	t4_tls_mod_unload();
1565 	t4_ddp_mod_unload();
1566 
1567 	t4_uninit_connect_cpl_handlers();
1568 	t4_uninit_listen_cpl_handlers();
1569 	t4_uninit_cpl_io_handlers();
1570 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM);
1571 
1572 	return (0);
1573 }
1574 #endif	/* TCP_OFFLOAD */
1575 
1576 static int
1577 t4_tom_modevent(module_t mod, int cmd, void *arg)
1578 {
1579 	int rc = 0;
1580 
1581 #ifdef TCP_OFFLOAD
1582 	switch (cmd) {
1583 	case MOD_LOAD:
1584 		rc = t4_tom_mod_load();
1585 		break;
1586 
1587 	case MOD_UNLOAD:
1588 		rc = t4_tom_mod_unload();
1589 		break;
1590 
1591 	default:
1592 		rc = EINVAL;
1593 	}
1594 #else
1595 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
1596 	rc = EOPNOTSUPP;
1597 #endif
1598 	return (rc);
1599 }
1600 
1601 static moduledata_t t4_tom_moddata= {
1602 	"t4_tom",
1603 	t4_tom_modevent,
1604 	0
1605 };
1606 
1607 MODULE_VERSION(t4_tom, 1);
1608 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
1609 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
1610 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
1611