xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision 3a0793336edfc21cb6d4c8c5c5d7f1665f3e6c5a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/ktr.h>
41 #include <sys/lock.h>
42 #include <sys/limits.h>
43 #include <sys/module.h>
44 #include <sys/protosw.h>
45 #include <sys/domain.h>
46 #include <sys/refcount.h>
47 #include <sys/rmlock.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.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_seq.h>
65 #include <netinet/tcp_timer.h>
66 #include <netinet/tcp_var.h>
67 #include <netinet/toecore.h>
68 #include <netinet/cc/cc.h>
69 
70 #ifdef TCP_OFFLOAD
71 #include "common/common.h"
72 #include "common/t4_msg.h"
73 #include "common/t4_regs.h"
74 #include "common/t4_regs_values.h"
75 #include "common/t4_tcb.h"
76 #include "t4_clip.h"
77 #include "tom/t4_tom_l2t.h"
78 #include "tom/t4_tom.h"
79 #include "tom/t4_tls.h"
80 
81 static struct protosw toe_protosw;
82 static struct protosw toe6_protosw;
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 void free_tom_data(struct adapter *, struct tom_data *);
103 static void reclaim_wr_resources(void *, int);
104 
105 struct toepcb *
106 alloc_toepcb(struct vi_info *vi, int flags)
107 {
108 	struct port_info *pi = vi->pi;
109 	struct adapter *sc = pi->adapter;
110 	struct toepcb *toep;
111 	int tx_credits, txsd_total, len;
112 
113 	/*
114 	 * The firmware counts tx work request credits in units of 16 bytes
115 	 * each.  Reserve room for an ABORT_REQ so the driver never has to worry
116 	 * about tx credits if it wants to abort a connection.
117 	 */
118 	tx_credits = sc->params.ofldq_wr_cred;
119 	tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
120 
121 	/*
122 	 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
123 	 * immediate payload, and firmware counts tx work request credits in
124 	 * units of 16 byte.  Calculate the maximum work requests possible.
125 	 */
126 	txsd_total = tx_credits /
127 	    howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
128 
129 	len = offsetof(struct toepcb, txsd) +
130 	    txsd_total * sizeof(struct ofld_tx_sdesc);
131 
132 	toep = malloc(len, M_CXGBE, M_ZERO | flags);
133 	if (toep == NULL)
134 		return (NULL);
135 
136 	refcount_init(&toep->refcount, 1);
137 	toep->td = sc->tom_softc;
138 	toep->vi = vi;
139 	toep->tid = -1;
140 	toep->tx_total = tx_credits;
141 	toep->tx_credits = tx_credits;
142 	mbufq_init(&toep->ulp_pduq, INT_MAX);
143 	mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
144 	toep->txsd_total = txsd_total;
145 	toep->txsd_avail = txsd_total;
146 	toep->txsd_pidx = 0;
147 	toep->txsd_cidx = 0;
148 	aiotx_init_toep(toep);
149 
150 	return (toep);
151 }
152 
153 /*
154  * Initialize a toepcb after its params have been filled out.
155  */
156 int
157 init_toepcb(struct vi_info *vi, struct toepcb *toep)
158 {
159 	struct conn_params *cp = &toep->params;
160 	struct port_info *pi = vi->pi;
161 	struct adapter *sc = pi->adapter;
162 	struct tx_cl_rl_params *tc;
163 
164 	if (cp->tc_idx >= 0 && cp->tc_idx < sc->params.nsched_cls) {
165 		tc = &pi->sched_params->cl_rl[cp->tc_idx];
166 		mtx_lock(&sc->tc_lock);
167 		if (tc->state != CS_HW_CONFIGURED) {
168 			CH_ERR(vi, "tid %d cannot be bound to traffic class %d "
169 			    "because it is not configured (its state is %d)\n",
170 			    toep->tid, cp->tc_idx, tc->state);
171 			cp->tc_idx = -1;
172 		} else {
173 			tc->refcount++;
174 		}
175 		mtx_unlock(&sc->tc_lock);
176 	}
177 	toep->ofld_txq = &sc->sge.ofld_txq[cp->txq_idx];
178 	toep->ofld_rxq = &sc->sge.ofld_rxq[cp->rxq_idx];
179 	toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
180 
181 	tls_init_toep(toep);
182 	MPASS(ulp_mode(toep) != ULP_MODE_TCPDDP);
183 
184 	toep->flags |= TPF_INITIALIZED;
185 
186 	return (0);
187 }
188 
189 struct toepcb *
190 hold_toepcb(struct toepcb *toep)
191 {
192 
193 	refcount_acquire(&toep->refcount);
194 	return (toep);
195 }
196 
197 void
198 free_toepcb(struct toepcb *toep)
199 {
200 
201 	if (refcount_release(&toep->refcount) == 0)
202 		return;
203 
204 	KASSERT(!(toep->flags & TPF_ATTACHED),
205 	    ("%s: attached to an inpcb", __func__));
206 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
207 	    ("%s: CPL pending", __func__));
208 
209 	if (toep->flags & TPF_INITIALIZED) {
210 		if (ulp_mode(toep) == ULP_MODE_TCPDDP)
211 			ddp_uninit_toep(toep);
212 		tls_uninit_toep(toep);
213 	}
214 	free(toep, M_CXGBE);
215 }
216 
217 /*
218  * Set up the socket for TCP offload.
219  */
220 void
221 offload_socket(struct socket *so, struct toepcb *toep)
222 {
223 	struct tom_data *td = toep->td;
224 	struct inpcb *inp = sotoinpcb(so);
225 	struct tcpcb *tp = intotcpcb(inp);
226 	struct sockbuf *sb;
227 
228 	INP_WLOCK_ASSERT(inp);
229 
230 	/* Update socket */
231 	sb = &so->so_snd;
232 	SOCKBUF_LOCK(sb);
233 	sb->sb_flags |= SB_NOCOALESCE;
234 	SOCKBUF_UNLOCK(sb);
235 	sb = &so->so_rcv;
236 	SOCKBUF_LOCK(sb);
237 	sb->sb_flags |= SB_NOCOALESCE;
238 	if (inp->inp_vflag & INP_IPV6)
239 		so->so_proto = &toe6_protosw;
240 	else
241 		so->so_proto = &toe_protosw;
242 	SOCKBUF_UNLOCK(sb);
243 
244 	/* Update TCP PCB */
245 	tp->tod = &td->tod;
246 	tp->t_toe = toep;
247 	tp->t_flags |= TF_TOE;
248 
249 	/* Install an extra hold on inp */
250 	toep->inp = inp;
251 	toep->flags |= TPF_ATTACHED;
252 	in_pcbref(inp);
253 
254 	/* Add the TOE PCB to the active list */
255 	mtx_lock(&td->toep_list_lock);
256 	TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
257 	mtx_unlock(&td->toep_list_lock);
258 }
259 
260 void
261 restore_so_proto(struct socket *so, bool v6)
262 {
263 	if (v6)
264 		so->so_proto = &tcp6_protosw;
265 	else
266 		so->so_proto = &tcp_protosw;
267 }
268 
269 /* This is _not_ the normal way to "unoffload" a socket. */
270 void
271 undo_offload_socket(struct socket *so)
272 {
273 	struct inpcb *inp = sotoinpcb(so);
274 	struct tcpcb *tp = intotcpcb(inp);
275 	struct toepcb *toep = tp->t_toe;
276 	struct tom_data *td = toep->td;
277 	struct sockbuf *sb;
278 
279 	INP_WLOCK_ASSERT(inp);
280 
281 	sb = &so->so_snd;
282 	SOCKBUF_LOCK(sb);
283 	sb->sb_flags &= ~SB_NOCOALESCE;
284 	SOCKBUF_UNLOCK(sb);
285 	sb = &so->so_rcv;
286 	SOCKBUF_LOCK(sb);
287 	sb->sb_flags &= ~SB_NOCOALESCE;
288 	restore_so_proto(so, inp->inp_vflag & INP_IPV6);
289 	SOCKBUF_UNLOCK(sb);
290 
291 	tp->tod = NULL;
292 	tp->t_toe = NULL;
293 	tp->t_flags &= ~TF_TOE;
294 
295 	toep->inp = NULL;
296 	toep->flags &= ~TPF_ATTACHED;
297 	if (in_pcbrele_wlocked(inp))
298 		panic("%s: inp freed.", __func__);
299 
300 	mtx_lock(&td->toep_list_lock);
301 	TAILQ_REMOVE(&td->toep_list, toep, link);
302 	mtx_unlock(&td->toep_list_lock);
303 }
304 
305 static void
306 release_offload_resources(struct toepcb *toep)
307 {
308 	struct tom_data *td = toep->td;
309 	struct adapter *sc = td_adapter(td);
310 	int tid = toep->tid;
311 
312 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
313 	    ("%s: %p has CPL pending.", __func__, toep));
314 	KASSERT(!(toep->flags & TPF_ATTACHED),
315 	    ("%s: %p is still attached.", __func__, toep));
316 
317 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
318 	    __func__, toep, tid, toep->l2te, toep->ce);
319 
320 	/*
321 	 * These queues should have been emptied at approximately the same time
322 	 * that a normal connection's socket's so_snd would have been purged or
323 	 * drained.  Do _not_ clean up here.
324 	 */
325 	MPASS(mbufq_empty(&toep->ulp_pduq));
326 	MPASS(mbufq_empty(&toep->ulp_pdu_reclaimq));
327 #ifdef INVARIANTS
328 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
329 		ddp_assert_empty(toep);
330 #endif
331 	MPASS(TAILQ_EMPTY(&toep->aiotx_jobq));
332 
333 	if (toep->l2te)
334 		t4_l2t_release(toep->l2te);
335 
336 	if (tid >= 0) {
337 		remove_tid(sc, tid, toep->ce ? 2 : 1);
338 		release_tid(sc, tid, toep->ctrlq);
339 	}
340 
341 	if (toep->ce)
342 		t4_release_clip_entry(sc, toep->ce);
343 
344 	if (toep->params.tc_idx != -1)
345 		t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx);
346 
347 	mtx_lock(&td->toep_list_lock);
348 	TAILQ_REMOVE(&td->toep_list, toep, link);
349 	mtx_unlock(&td->toep_list_lock);
350 
351 	free_toepcb(toep);
352 }
353 
354 /*
355  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
356  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
357  * pending CPL) then it is time to release all resources tied to the toepcb.
358  *
359  * Also gets called when an offloaded active open fails and the TOM wants the
360  * kernel to take the TCP PCB back.
361  */
362 static void
363 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
364 {
365 #if defined(KTR) || defined(INVARIANTS)
366 	struct inpcb *inp = tptoinpcb(tp);
367 #endif
368 	struct toepcb *toep = tp->t_toe;
369 
370 	INP_WLOCK_ASSERT(inp);
371 
372 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
373 	KASSERT(toep->flags & TPF_ATTACHED,
374 	    ("%s: not attached", __func__));
375 
376 #ifdef KTR
377 	if (tp->t_state == TCPS_SYN_SENT) {
378 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
379 		    __func__, toep->tid, toep, toep->flags, inp,
380 		    inp->inp_flags);
381 	} else {
382 		CTR6(KTR_CXGBE,
383 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
384 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
385 		    inp->inp_flags);
386 	}
387 #endif
388 
389 	tp->tod = NULL;
390 	tp->t_toe = NULL;
391 	tp->t_flags &= ~TF_TOE;
392 	toep->flags &= ~TPF_ATTACHED;
393 
394 	if (!(toep->flags & TPF_CPL_PENDING))
395 		release_offload_resources(toep);
396 }
397 
398 /*
399  * setsockopt handler.
400  */
401 static void
402 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
403 {
404 	struct adapter *sc = tod->tod_softc;
405 	struct toepcb *toep = tp->t_toe;
406 
407 	if (dir == SOPT_GET)
408 		return;
409 
410 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
411 
412 	switch (name) {
413 	case TCP_NODELAY:
414 		if (tp->t_state != TCPS_ESTABLISHED)
415 			break;
416 		toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
417 		t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
418 		    V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0);
419 		break;
420 	default:
421 		break;
422 	}
423 }
424 
425 static inline uint64_t
426 get_tcb_tflags(const uint64_t *tcb)
427 {
428 
429 	return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32));
430 }
431 
432 static inline uint32_t
433 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift)
434 {
435 #define LAST_WORD ((TCB_SIZE / 4) - 1)
436 	uint64_t t1, t2;
437 	int flit_idx;
438 
439 	MPASS(mask != 0);
440 	MPASS(word <= LAST_WORD);
441 	MPASS(shift < 32);
442 
443 	flit_idx = (LAST_WORD - word) / 2;
444 	if (word & 0x1)
445 		shift += 32;
446 	t1 = be64toh(tcb[flit_idx]) >> shift;
447 	t2 = 0;
448 	if (fls(mask) > 64 - shift) {
449 		/*
450 		 * Will spill over into the next logical flit, which is the flit
451 		 * before this one.  The flit_idx before this one must be valid.
452 		 */
453 		MPASS(flit_idx > 0);
454 		t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift);
455 	}
456 	return ((t2 | t1) & mask);
457 #undef LAST_WORD
458 }
459 #define GET_TCB_FIELD(tcb, F) \
460     get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F)
461 
462 /*
463  * Issues a CPL_GET_TCB to read the entire TCB for the tid.
464  */
465 static int
466 send_get_tcb(struct adapter *sc, u_int tid)
467 {
468 	struct cpl_get_tcb *cpl;
469 	struct wrq_cookie cookie;
470 
471 	MPASS(tid >= sc->tids.tid_base);
472 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
473 
474 	cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16),
475 	    &cookie);
476 	if (__predict_false(cpl == NULL))
477 		return (ENOMEM);
478 	bzero(cpl, sizeof(*cpl));
479 	INIT_TP_WR(cpl, tid);
480 	OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid));
481 	cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) |
482 	    V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id));
483 	cpl->cookie = 0xff;
484 	commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie);
485 
486 	return (0);
487 }
488 
489 static struct tcb_histent *
490 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags)
491 {
492 	struct tcb_histent *te;
493 
494 	MPASS(flags == M_NOWAIT || flags == M_WAITOK);
495 
496 	te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags);
497 	if (te == NULL)
498 		return (NULL);
499 	mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF);
500 	callout_init_mtx(&te->te_callout, &te->te_lock, 0);
501 	te->te_adapter = sc;
502 	te->te_tid = tid;
503 
504 	return (te);
505 }
506 
507 static void
508 free_tcb_histent(struct tcb_histent *te)
509 {
510 
511 	mtx_destroy(&te->te_lock);
512 	free(te, M_CXGBE);
513 }
514 
515 /*
516  * Start tracking the tid in the TCB history.
517  */
518 int
519 add_tid_to_history(struct adapter *sc, u_int tid)
520 {
521 	struct tcb_histent *te = NULL;
522 	struct tom_data *td = sc->tom_softc;
523 	int rc;
524 
525 	MPASS(tid >= sc->tids.tid_base);
526 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
527 
528 	if (td->tcb_history == NULL)
529 		return (ENXIO);
530 
531 	rw_wlock(&td->tcb_history_lock);
532 	if (td->tcb_history[tid] != NULL) {
533 		rc = EEXIST;
534 		goto done;
535 	}
536 	te = alloc_tcb_histent(sc, tid, M_NOWAIT);
537 	if (te == NULL) {
538 		rc = ENOMEM;
539 		goto done;
540 	}
541 	mtx_lock(&te->te_lock);
542 	rc = send_get_tcb(sc, tid);
543 	if (rc == 0) {
544 		te->te_flags |= TE_RPL_PENDING;
545 		td->tcb_history[tid] = te;
546 	} else {
547 		free(te, M_CXGBE);
548 	}
549 	mtx_unlock(&te->te_lock);
550 done:
551 	rw_wunlock(&td->tcb_history_lock);
552 	return (rc);
553 }
554 
555 static void
556 remove_tcb_histent(struct tcb_histent *te)
557 {
558 	struct adapter *sc = te->te_adapter;
559 	struct tom_data *td = sc->tom_softc;
560 
561 	rw_assert(&td->tcb_history_lock, RA_WLOCKED);
562 	mtx_assert(&te->te_lock, MA_OWNED);
563 	MPASS(td->tcb_history[te->te_tid] == te);
564 
565 	td->tcb_history[te->te_tid] = NULL;
566 	free_tcb_histent(te);
567 	rw_wunlock(&td->tcb_history_lock);
568 }
569 
570 static inline struct tcb_histent *
571 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem)
572 {
573 	struct tcb_histent *te;
574 	struct tom_data *td = sc->tom_softc;
575 
576 	MPASS(tid >= sc->tids.tid_base);
577 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
578 
579 	if (td->tcb_history == NULL)
580 		return (NULL);
581 
582 	if (addrem)
583 		rw_wlock(&td->tcb_history_lock);
584 	else
585 		rw_rlock(&td->tcb_history_lock);
586 	te = td->tcb_history[tid];
587 	if (te != NULL) {
588 		mtx_lock(&te->te_lock);
589 		return (te);	/* with both locks held */
590 	}
591 	if (addrem)
592 		rw_wunlock(&td->tcb_history_lock);
593 	else
594 		rw_runlock(&td->tcb_history_lock);
595 
596 	return (te);
597 }
598 
599 static inline void
600 release_tcb_histent(struct tcb_histent *te)
601 {
602 	struct adapter *sc = te->te_adapter;
603 	struct tom_data *td = sc->tom_softc;
604 
605 	mtx_assert(&te->te_lock, MA_OWNED);
606 	mtx_unlock(&te->te_lock);
607 	rw_assert(&td->tcb_history_lock, RA_RLOCKED);
608 	rw_runlock(&td->tcb_history_lock);
609 }
610 
611 static void
612 request_tcb(void *arg)
613 {
614 	struct tcb_histent *te = arg;
615 
616 	mtx_assert(&te->te_lock, MA_OWNED);
617 
618 	/* Noone else is supposed to update the histent. */
619 	MPASS(!(te->te_flags & TE_RPL_PENDING));
620 	if (send_get_tcb(te->te_adapter, te->te_tid) == 0)
621 		te->te_flags |= TE_RPL_PENDING;
622 	else
623 		callout_schedule(&te->te_callout, hz / 100);
624 }
625 
626 static void
627 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb)
628 {
629 	struct tom_data *td = te->te_adapter->tom_softc;
630 	uint64_t tflags = get_tcb_tflags(tcb);
631 	uint8_t sample = 0;
632 
633 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) {
634 		if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0)
635 			sample |= TS_RTO;
636 		if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0)
637 			sample |= TS_DUPACKS;
638 		if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold)
639 			sample |= TS_FASTREXMT;
640 	}
641 
642 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) {
643 		uint32_t snd_wnd;
644 
645 		sample |= TS_SND_BACKLOGGED;	/* for whatever reason. */
646 
647 		snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
648 		if (tflags & V_TF_RECV_SCALE(1))
649 			snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE);
650 		if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd)
651 			sample |= TS_CWND_LIMITED;	/* maybe due to CWND */
652 	}
653 
654 	if (tflags & V_TF_CCTRL_ECN(1)) {
655 
656 		/*
657 		 * CE marker on incoming IP hdr, echoing ECE back in the TCP
658 		 * hdr.  Indicates congestion somewhere on the way from the peer
659 		 * to this node.
660 		 */
661 		if (tflags & V_TF_CCTRL_ECE(1))
662 			sample |= TS_ECN_ECE;
663 
664 		/*
665 		 * ECE seen and CWR sent (or about to be sent).  Might indicate
666 		 * congestion on the way to the peer.  This node is reducing its
667 		 * congestion window in response.
668 		 */
669 		if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1)))
670 			sample |= TS_ECN_CWR;
671 	}
672 
673 	te->te_sample[te->te_pidx] = sample;
674 	if (++te->te_pidx == nitems(te->te_sample))
675 		te->te_pidx = 0;
676 	memcpy(te->te_tcb, tcb, TCB_SIZE);
677 	te->te_flags |= TE_ACTIVE;
678 }
679 
680 static int
681 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
682 {
683 	struct adapter *sc = iq->adapter;
684 	const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *);
685 	const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1);
686 	struct tcb_histent *te;
687 	const u_int tid = GET_TID(cpl);
688 	bool remove;
689 
690 	remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED;
691 	te = lookup_tcb_histent(sc, tid, remove);
692 	if (te == NULL) {
693 		/* Not in the history.  Who issued the GET_TCB for this? */
694 		device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, "
695 		    "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid,
696 		    (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE),
697 		    GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE),
698 		    GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie);
699 		goto done;
700 	}
701 
702 	MPASS(te->te_flags & TE_RPL_PENDING);
703 	te->te_flags &= ~TE_RPL_PENDING;
704 	if (remove) {
705 		remove_tcb_histent(te);
706 	} else {
707 		update_tcb_histent(te, tcb);
708 		callout_reset(&te->te_callout, hz / 10, request_tcb, te);
709 		release_tcb_histent(te);
710 	}
711 done:
712 	m_freem(m);
713 	return (0);
714 }
715 
716 static void
717 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti)
718 {
719 	uint32_t v;
720 
721 	ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE);
722 
723 	v = GET_TCB_FIELD(tcb, T_SRTT);
724 	ti->tcpi_rtt = tcp_ticks_to_us(sc, v);
725 
726 	v = GET_TCB_FIELD(tcb, T_RTTVAR);
727 	ti->tcpi_rttvar = tcp_ticks_to_us(sc, v);
728 
729 	ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH);
730 	ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND);
731 	ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT);
732 	ti->tcpi_rcv_adv = GET_TCB_FIELD(tcb, RCV_ADV);
733 	ti->tcpi_dupacks = GET_TCB_FIELD(tcb, T_DUPACKS);
734 
735 	v = GET_TCB_FIELD(tcb, TX_MAX);
736 	ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW);
737 	ti->tcpi_snd_una = v - GET_TCB_FIELD(tcb, SND_UNA_RAW);
738 	ti->tcpi_snd_max = v - GET_TCB_FIELD(tcb, SND_MAX_RAW);
739 
740 	/* Receive window being advertised by us. */
741 	ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE);	/* Yes, SND. */
742 	ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND);
743 
744 	/* Send window */
745 	ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE);	/* Yes, RCV. */
746 	ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
747 	if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1))
748 		ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale;
749 	else
750 		ti->tcpi_snd_wscale = 0;
751 
752 }
753 
754 static void
755 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te,
756     struct tcp_info *ti)
757 {
758 
759 	fill_tcp_info_from_tcb(sc, te->te_tcb, ti);
760 }
761 
762 /*
763  * Reads the TCB for the given tid using a memory window and copies it to 'buf'
764  * in the same format as CPL_GET_TCB_RPL.
765  */
766 static void
767 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf)
768 {
769 	int i, j, k, rc;
770 	uint32_t addr;
771 	u_char *tcb, tmp;
772 
773 	MPASS(tid >= sc->tids.tid_base);
774 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
775 
776 	addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE;
777 	rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE);
778 	if (rc != 0)
779 		return;
780 
781 	tcb = (u_char *)buf;
782 	for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) {
783 		for (k = 0; k < 16; k++) {
784 			tmp = tcb[i + k];
785 			tcb[i + k] = tcb[j + k];
786 			tcb[j + k] = tmp;
787 		}
788 	}
789 }
790 
791 static void
792 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti)
793 {
794 	uint64_t tcb[TCB_SIZE / sizeof(uint64_t)];
795 	struct tcb_histent *te;
796 
797 	ti->tcpi_toe_tid = tid;
798 	te = lookup_tcb_histent(sc, tid, false);
799 	if (te != NULL) {
800 		fill_tcp_info_from_history(sc, te, ti);
801 		release_tcb_histent(te);
802 	} else {
803 		if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) {
804 			/* XXX: tell firmware to flush TCB cache. */
805 		}
806 		read_tcb_using_memwin(sc, tid, tcb);
807 		fill_tcp_info_from_tcb(sc, tcb, ti);
808 	}
809 }
810 
811 /*
812  * Called by the kernel to allow the TOE driver to "refine" values filled up in
813  * the tcp_info for an offloaded connection.
814  */
815 static void
816 t4_tcp_info(struct toedev *tod, const struct tcpcb *tp, struct tcp_info *ti)
817 {
818 	struct adapter *sc = tod->tod_softc;
819 	struct toepcb *toep = tp->t_toe;
820 
821 	INP_LOCK_ASSERT(tptoinpcb(tp));
822 	MPASS(ti != NULL);
823 
824 	fill_tcp_info(sc, toep->tid, ti);
825 }
826 
827 #ifdef KERN_TLS
828 static int
829 t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp,
830     struct ktls_session *tls, int direction)
831 {
832 	struct toepcb *toep = tp->t_toe;
833 
834 	INP_WLOCK_ASSERT(tptoinpcb(tp));
835 	MPASS(tls != NULL);
836 
837 	return (tls_alloc_ktls(toep, tls, direction));
838 }
839 #endif
840 
841 static void
842 send_mss_flowc_wr(struct adapter *sc, struct toepcb *toep)
843 {
844 	struct wrq_cookie cookie;
845 	struct fw_flowc_wr *flowc;
846 	struct ofld_tx_sdesc *txsd;
847 	const int flowclen = sizeof(*flowc) + sizeof(struct fw_flowc_mnemval);
848 	const int flowclen16 = howmany(flowclen, 16);
849 
850 	if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0) {
851 		CH_ERR(sc, "%s: tid %u out of tx credits (%d, %d).\n", __func__,
852 		    toep->tid, toep->tx_credits, toep->txsd_avail);
853 		return;
854 	}
855 
856 	flowc = start_wrq_wr(&toep->ofld_txq->wrq, flowclen16, &cookie);
857 	if (__predict_false(flowc == NULL)) {
858 		CH_ERR(sc, "ENOMEM in %s for tid %u.\n", __func__, toep->tid);
859 		return;
860 	}
861 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
862 	    V_FW_FLOWC_WR_NPARAMS(1));
863 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
864 	    V_FW_WR_FLOWID(toep->tid));
865 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_MSS;
866 	flowc->mnemval[0].val = htobe32(toep->params.emss);
867 
868 	txsd = &toep->txsd[toep->txsd_pidx];
869 	txsd->tx_credits = flowclen16;
870 	txsd->plen = 0;
871 	toep->tx_credits -= txsd->tx_credits;
872 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
873 		toep->txsd_pidx = 0;
874 	toep->txsd_avail--;
875 	commit_wrq_wr(&toep->ofld_txq->wrq, flowc, &cookie);
876 }
877 
878 static void
879 t4_pmtu_update(struct toedev *tod, struct tcpcb *tp, tcp_seq seq, int mtu)
880 {
881 	struct work_request_hdr *wrh;
882 	struct ulp_txpkt *ulpmc;
883 	int idx, len;
884 	struct wrq_cookie cookie;
885 	struct inpcb *inp = tptoinpcb(tp);
886 	struct toepcb *toep = tp->t_toe;
887 	struct adapter *sc = td_adapter(toep->td);
888 	unsigned short *mtus = &sc->params.mtus[0];
889 
890 	INP_WLOCK_ASSERT(inp);
891 	MPASS(mtu > 0);	/* kernel is supposed to provide something usable. */
892 
893 	/* tp->snd_una and snd_max are in host byte order too. */
894 	seq = be32toh(seq);
895 
896 	CTR6(KTR_CXGBE, "%s: tid %d, seq 0x%08x, mtu %u, mtu_idx %u (%d)",
897 	    __func__, toep->tid, seq, mtu, toep->params.mtu_idx,
898 	    mtus[toep->params.mtu_idx]);
899 
900 	if (ulp_mode(toep) == ULP_MODE_NONE &&	/* XXX: Read TCB otherwise? */
901 	    (SEQ_LT(seq, tp->snd_una) || SEQ_GEQ(seq, tp->snd_max))) {
902 		CTR5(KTR_CXGBE,
903 		    "%s: tid %d, seq 0x%08x not in range [0x%08x, 0x%08x).",
904 		    __func__, toep->tid, seq, tp->snd_una, tp->snd_max);
905 		return;
906 	}
907 
908 	/* Find the best mtu_idx for the suggested MTU. */
909 	for (idx = 0; idx < NMTUS - 1 && mtus[idx + 1] <= mtu; idx++)
910 		continue;
911 	if (idx >= toep->params.mtu_idx)
912 		return;	/* Never increase the PMTU (just like the kernel). */
913 
914 	/*
915 	 * We'll send a compound work request with 2 SET_TCB_FIELDs -- the first
916 	 * one updates the mtu_idx and the second one triggers a retransmit.
917 	 */
918 	len = sizeof(*wrh) + 2 * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
919 	wrh = start_wrq_wr(toep->ctrlq, howmany(len, 16), &cookie);
920 	if (wrh == NULL) {
921 		CH_ERR(sc, "failed to change mtu_idx of tid %d (%u -> %u).\n",
922 		    toep->tid, toep->params.mtu_idx, idx);
923 		return;
924 	}
925 	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
926 	ulpmc = (struct ulp_txpkt *)(wrh + 1);
927 	ulpmc = mk_set_tcb_field_ulp(sc, ulpmc, toep->tid, W_TCB_T_MAXSEG,
928 	    V_TCB_T_MAXSEG(M_TCB_T_MAXSEG), V_TCB_T_MAXSEG(idx));
929 	ulpmc = mk_set_tcb_field_ulp(sc, ulpmc, toep->tid, W_TCB_TIMESTAMP,
930 	    V_TCB_TIMESTAMP(0x7FFFFULL << 11), 0);
931 	commit_wrq_wr(toep->ctrlq, wrh, &cookie);
932 
933 	/* Update the software toepcb and tcpcb. */
934 	toep->params.mtu_idx = idx;
935 	tp->t_maxseg = mtus[toep->params.mtu_idx];
936 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
937 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
938 	else
939 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
940 	toep->params.emss = tp->t_maxseg;
941 	if (tp->t_flags & TF_RCVD_TSTMP)
942 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
943 
944 	/* Update the firmware flowc. */
945 	send_mss_flowc_wr(sc, toep);
946 
947 	/* Update the MTU in the kernel's hostcache. */
948 	if (sc->tt.update_hc_on_pmtu_change != 0) {
949 		struct in_conninfo inc = {0};
950 
951 		inc.inc_fibnum = inp->inp_inc.inc_fibnum;
952 		if (inp->inp_inc.inc_flags & INC_ISIPV6) {
953 			inc.inc_flags |= INC_ISIPV6;
954 			inc.inc6_faddr = inp->inp_inc.inc6_faddr;
955 		} else {
956 			inc.inc_faddr = inp->inp_inc.inc_faddr;
957 		}
958 		tcp_hc_updatemtu(&inc, mtu);
959 	}
960 
961 	CTR6(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u), t_maxseg %u, emss %u",
962 	    __func__, toep->tid, toep->params.mtu_idx,
963 	    mtus[toep->params.mtu_idx], tp->t_maxseg, toep->params.emss);
964 }
965 
966 /*
967  * The TOE driver will not receive any more CPLs for the tid associated with the
968  * toepcb; release the hold on the inpcb.
969  */
970 void
971 final_cpl_received(struct toepcb *toep)
972 {
973 	struct inpcb *inp = toep->inp;
974 	bool need_wakeup;
975 
976 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
977 	INP_WLOCK_ASSERT(inp);
978 	KASSERT(toep->flags & TPF_CPL_PENDING,
979 	    ("%s: CPL not pending already?", __func__));
980 
981 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
982 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
983 
984 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
985 		release_ddp_resources(toep);
986 	toep->inp = NULL;
987 	need_wakeup = (toep->flags & TPF_WAITING_FOR_FINAL) != 0;
988 	toep->flags &= ~(TPF_CPL_PENDING | TPF_WAITING_FOR_FINAL);
989 	mbufq_drain(&toep->ulp_pduq);
990 	mbufq_drain(&toep->ulp_pdu_reclaimq);
991 
992 	if (!(toep->flags & TPF_ATTACHED))
993 		release_offload_resources(toep);
994 
995 	if (!in_pcbrele_wlocked(inp))
996 		INP_WUNLOCK(inp);
997 
998 	if (need_wakeup) {
999 		struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep);
1000 
1001 		mtx_lock(lock);
1002 		wakeup(toep);
1003 		mtx_unlock(lock);
1004 	}
1005 }
1006 
1007 void
1008 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
1009 {
1010 	struct tid_info *t = &sc->tids;
1011 
1012 	MPASS(tid >= t->tid_base);
1013 	MPASS(tid - t->tid_base < t->ntids);
1014 
1015 	t->tid_tab[tid - t->tid_base] = ctx;
1016 	atomic_add_int(&t->tids_in_use, ntids);
1017 }
1018 
1019 void *
1020 lookup_tid(struct adapter *sc, int tid)
1021 {
1022 	struct tid_info *t = &sc->tids;
1023 
1024 	return (t->tid_tab[tid - t->tid_base]);
1025 }
1026 
1027 void
1028 update_tid(struct adapter *sc, int tid, void *ctx)
1029 {
1030 	struct tid_info *t = &sc->tids;
1031 
1032 	t->tid_tab[tid - t->tid_base] = ctx;
1033 }
1034 
1035 void
1036 remove_tid(struct adapter *sc, int tid, int ntids)
1037 {
1038 	struct tid_info *t = &sc->tids;
1039 
1040 	t->tid_tab[tid - t->tid_base] = NULL;
1041 	atomic_subtract_int(&t->tids_in_use, ntids);
1042 }
1043 
1044 /*
1045  * What mtu_idx to use, given a 4-tuple.  Note that both s->mss and tcp_mssopt
1046  * have the MSS that we should advertise in our SYN.  Advertised MSS doesn't
1047  * account for any TCP options so the effective MSS (only payload, no headers or
1048  * options) could be different.
1049  */
1050 static int
1051 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc,
1052     struct offload_settings *s)
1053 {
1054 	unsigned short *mtus = &sc->params.mtus[0];
1055 	int i, mss, mtu;
1056 
1057 	MPASS(inc != NULL);
1058 
1059 	mss = s->mss > 0 ? s->mss : tcp_mssopt(inc);
1060 	if (inc->inc_flags & INC_ISIPV6)
1061 		mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1062 	else
1063 		mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr);
1064 
1065 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++)
1066 		continue;
1067 
1068 	return (i);
1069 }
1070 
1071 /*
1072  * Determine the receive window size for a socket.
1073  */
1074 u_long
1075 select_rcv_wnd(struct socket *so)
1076 {
1077 	unsigned long wnd;
1078 
1079 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1080 
1081 	wnd = sbspace(&so->so_rcv);
1082 	if (wnd < MIN_RCV_WND)
1083 		wnd = MIN_RCV_WND;
1084 
1085 	return min(wnd, MAX_RCV_WND);
1086 }
1087 
1088 int
1089 select_rcv_wscale(void)
1090 {
1091 	int wscale = 0;
1092 	unsigned long space = sb_max;
1093 
1094 	if (space > MAX_RCV_WND)
1095 		space = MAX_RCV_WND;
1096 
1097 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
1098 		wscale++;
1099 
1100 	return (wscale);
1101 }
1102 
1103 __be64
1104 calc_options0(struct vi_info *vi, struct conn_params *cp)
1105 {
1106 	uint64_t opt0 = 0;
1107 
1108 	opt0 |= F_TCAM_BYPASS;
1109 
1110 	MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE);
1111 	opt0 |= V_WND_SCALE(cp->wscale);
1112 
1113 	MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS);
1114 	opt0 |= V_MSS_IDX(cp->mtu_idx);
1115 
1116 	MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE);
1117 	opt0 |= V_ULP_MODE(cp->ulp_mode);
1118 
1119 	MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ);
1120 	opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize);
1121 
1122 	MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->adapter->vres.l2t.size);
1123 	opt0 |= V_L2T_IDX(cp->l2t_idx);
1124 
1125 	opt0 |= V_SMAC_SEL(vi->smt_idx);
1126 	opt0 |= V_TX_CHAN(vi->pi->tx_chan);
1127 
1128 	MPASS(cp->keepalive == 0 || cp->keepalive == 1);
1129 	opt0 |= V_KEEP_ALIVE(cp->keepalive);
1130 
1131 	MPASS(cp->nagle == 0 || cp->nagle == 1);
1132 	opt0 |= V_NAGLE(cp->nagle);
1133 
1134 	return (htobe64(opt0));
1135 }
1136 
1137 __be32
1138 calc_options2(struct vi_info *vi, struct conn_params *cp)
1139 {
1140 	uint32_t opt2 = 0;
1141 	struct port_info *pi = vi->pi;
1142 	struct adapter *sc = pi->adapter;
1143 
1144 	/*
1145 	 * rx flow control, rx coalesce, congestion control, and tx pace are all
1146 	 * explicitly set by the driver.  On T5+ the ISS is also set by the
1147 	 * driver to the value picked by the kernel.
1148 	 */
1149 	if (is_t4(sc)) {
1150 		opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID;
1151 		opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID;
1152 	} else {
1153 		opt2 |= F_T5_OPT_2_VALID;	/* all 4 valid */
1154 		opt2 |= F_T5_ISS;		/* ISS provided in CPL */
1155 	}
1156 
1157 	MPASS(cp->sack == 0 || cp->sack == 1);
1158 	opt2 |= V_SACK_EN(cp->sack);
1159 
1160 	MPASS(cp->tstamp == 0 || cp->tstamp == 1);
1161 	opt2 |= V_TSTAMPS_EN(cp->tstamp);
1162 
1163 	if (cp->wscale > 0)
1164 		opt2 |= F_WND_SCALE_EN;
1165 
1166 	MPASS(cp->ecn == 0 || cp->ecn == 1);
1167 	opt2 |= V_CCTRL_ECN(cp->ecn);
1168 
1169 	opt2 |= V_TX_QUEUE(TX_MODQ(pi->tx_chan));
1170 	opt2 |= V_PACE(0);
1171 	opt2 |= F_RSS_QUEUE_VALID;
1172 	opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id);
1173 	if (chip_id(sc) <= CHELSIO_T6) {
1174 		MPASS(pi->rx_chan == 0 || pi->rx_chan == 1);
1175 		opt2 |= V_RX_CHANNEL(pi->rx_chan);
1176 	}
1177 
1178 	MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL);
1179 	opt2 |= V_CONG_CNTRL(cp->cong_algo);
1180 
1181 	MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1);
1182 	if (cp->rx_coalesce == 1)
1183 		opt2 |= V_RX_COALESCE(M_RX_COALESCE);
1184 
1185 	opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
1186 	MPASS(cp->ulp_mode != ULP_MODE_TCPDDP);
1187 
1188 	return (htobe32(opt2));
1189 }
1190 
1191 uint64_t
1192 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
1193 {
1194 	struct adapter *sc = vi->adapter;
1195 	struct tp_params *tp = &sc->params.tp;
1196 	uint64_t ntuple = 0;
1197 
1198 	/*
1199 	 * Initialize each of the fields which we care about which are present
1200 	 * in the Compressed Filter Tuple.
1201 	 */
1202 	if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE)
1203 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
1204 
1205 	if (tp->port_shift >= 0)
1206 		ntuple |= (uint64_t)e->lport << tp->port_shift;
1207 
1208 	if (tp->protocol_shift >= 0)
1209 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
1210 
1211 	if (tp->vnic_shift >= 0 && tp->vnic_mode == FW_VNIC_MODE_PF_VF) {
1212 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) |
1213 		    V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) <<
1214 		    tp->vnic_shift;
1215 	}
1216 
1217 	if (is_t4(sc))
1218 		return (htobe32((uint32_t)ntuple));
1219 	else
1220 		return (htobe64(V_FILTER_TUPLE(ntuple)));
1221 }
1222 
1223 /*
1224  * Initialize various connection parameters.
1225  */
1226 void
1227 init_conn_params(struct vi_info *vi , struct offload_settings *s,
1228     struct in_conninfo *inc, struct socket *so,
1229     const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp)
1230 {
1231 	struct port_info *pi = vi->pi;
1232 	struct adapter *sc = pi->adapter;
1233 	struct tom_tunables *tt = &sc->tt;
1234 	struct inpcb *inp = sotoinpcb(so);
1235 	struct tcpcb *tp = intotcpcb(inp);
1236 	u_long wnd;
1237 	u_int q_idx;
1238 
1239 	MPASS(s->offload != 0);
1240 
1241 	/* Congestion control algorithm */
1242 	if (s->cong_algo >= 0)
1243 		cp->cong_algo = s->cong_algo & M_CONG_CNTRL;
1244 	else if (sc->tt.cong_algorithm >= 0)
1245 		cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL;
1246 	else {
1247 		struct cc_algo *cc = CC_ALGO(tp);
1248 
1249 		if (strcasecmp(cc->name, "reno") == 0)
1250 			cp->cong_algo = CONG_ALG_RENO;
1251 		else if (strcasecmp(cc->name, "tahoe") == 0)
1252 			cp->cong_algo = CONG_ALG_TAHOE;
1253 		if (strcasecmp(cc->name, "newreno") == 0)
1254 			cp->cong_algo = CONG_ALG_NEWRENO;
1255 		if (strcasecmp(cc->name, "highspeed") == 0)
1256 			cp->cong_algo = CONG_ALG_HIGHSPEED;
1257 		else {
1258 			/*
1259 			 * Use newreno in case the algorithm selected by the
1260 			 * host stack is not supported by the hardware.
1261 			 */
1262 			cp->cong_algo = CONG_ALG_NEWRENO;
1263 		}
1264 	}
1265 
1266 	/* Tx traffic scheduling class. */
1267 	if (s->sched_class >= 0 && s->sched_class < sc->params.nsched_cls)
1268 		cp->tc_idx = s->sched_class;
1269 	else
1270 		cp->tc_idx = -1;
1271 
1272 	/* Nagle's algorithm. */
1273 	if (s->nagle >= 0)
1274 		cp->nagle = s->nagle > 0 ? 1 : 0;
1275 	else
1276 		cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
1277 
1278 	/* TCP Keepalive. */
1279 	if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE)
1280 		cp->keepalive = 1;
1281 	else
1282 		cp->keepalive = 0;
1283 
1284 	/* Optimization that's specific to T5 @ 40G. */
1285 	if (tt->tx_align >= 0)
1286 		cp->tx_align =  tt->tx_align > 0 ? 1 : 0;
1287 	else if (chip_id(sc) == CHELSIO_T5 &&
1288 	    (port_top_speed(pi) > 10 || sc->params.nports > 2))
1289 		cp->tx_align = 1;
1290 	else
1291 		cp->tx_align = 0;
1292 
1293 	/* ULP mode. */
1294 	cp->ulp_mode = ULP_MODE_NONE;
1295 
1296 	/* Rx coalescing. */
1297 	if (s->rx_coalesce >= 0)
1298 		cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0;
1299 	else if (tt->rx_coalesce >= 0)
1300 		cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0;
1301 	else
1302 		cp->rx_coalesce = 1;	/* default */
1303 
1304 	/*
1305 	 * Index in the PMTU table.  This controls the MSS that we announce in
1306 	 * our SYN initially, but after ESTABLISHED it controls the MSS that we
1307 	 * use to send data.
1308 	 */
1309 	cp->mtu_idx = find_best_mtu_idx(sc, inc, s);
1310 
1311 	/* Tx queue for this connection. */
1312 	if (s->txq == QUEUE_RANDOM)
1313 		q_idx = arc4random();
1314 	else if (s->txq == QUEUE_ROUNDROBIN)
1315 		q_idx = atomic_fetchadd_int(&vi->txq_rr, 1);
1316 	else
1317 		q_idx = s->txq;
1318 	cp->txq_idx = vi->first_ofld_txq + q_idx % vi->nofldtxq;
1319 
1320 	/* Rx queue for this connection. */
1321 	if (s->rxq == QUEUE_RANDOM)
1322 		q_idx = arc4random();
1323 	else if (s->rxq == QUEUE_ROUNDROBIN)
1324 		q_idx = atomic_fetchadd_int(&vi->rxq_rr, 1);
1325 	else
1326 		q_idx = s->rxq;
1327 	cp->rxq_idx = vi->first_ofld_rxq + q_idx % vi->nofldrxq;
1328 
1329 	if (SOLISTENING(so)) {
1330 		/* Passive open */
1331 		MPASS(tcpopt != NULL);
1332 
1333 		/* TCP timestamp option */
1334 		if (tcpopt->tstamp &&
1335 		    (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323)))
1336 			cp->tstamp = 1;
1337 		else
1338 			cp->tstamp = 0;
1339 
1340 		/* SACK */
1341 		if (tcpopt->sack &&
1342 		    (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack)))
1343 			cp->sack = 1;
1344 		else
1345 			cp->sack = 0;
1346 
1347 		/* Receive window scaling. */
1348 		if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323)
1349 			cp->wscale = select_rcv_wscale();
1350 		else
1351 			cp->wscale = 0;
1352 
1353 		/* ECN */
1354 		if (tcpopt->ecn &&	/* XXX: review. */
1355 		    (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn)))
1356 			cp->ecn = 1;
1357 		else
1358 			cp->ecn = 0;
1359 
1360 		wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND);
1361 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1362 
1363 		if (tt->sndbuf > 0)
1364 			cp->sndbuf = tt->sndbuf;
1365 		else if (so->sol_sbsnd_flags & SB_AUTOSIZE &&
1366 		    V_tcp_do_autosndbuf)
1367 			cp->sndbuf = 256 * 1024;
1368 		else
1369 			cp->sndbuf = so->sol_sbsnd_hiwat;
1370 	} else {
1371 		/* Active open */
1372 
1373 		/* TCP timestamp option */
1374 		if (s->tstamp > 0 ||
1375 		    (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP)))
1376 			cp->tstamp = 1;
1377 		else
1378 			cp->tstamp = 0;
1379 
1380 		/* SACK */
1381 		if (s->sack > 0 ||
1382 		    (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT)))
1383 			cp->sack = 1;
1384 		else
1385 			cp->sack = 0;
1386 
1387 		/* Receive window scaling */
1388 		if (tp->t_flags & TF_REQ_SCALE)
1389 			cp->wscale = select_rcv_wscale();
1390 		else
1391 			cp->wscale = 0;
1392 
1393 		/* ECN */
1394 		if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1))
1395 			cp->ecn = 1;
1396 		else
1397 			cp->ecn = 0;
1398 
1399 		SOCKBUF_LOCK(&so->so_rcv);
1400 		wnd = max(select_rcv_wnd(so), MIN_RCV_WND);
1401 		SOCKBUF_UNLOCK(&so->so_rcv);
1402 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1403 
1404 		if (tt->sndbuf > 0)
1405 			cp->sndbuf = tt->sndbuf;
1406 		else {
1407 			SOCKBUF_LOCK(&so->so_snd);
1408 			if (so->so_snd.sb_flags & SB_AUTOSIZE &&
1409 			    V_tcp_do_autosndbuf)
1410 				cp->sndbuf = 256 * 1024;
1411 			else
1412 				cp->sndbuf = so->so_snd.sb_hiwat;
1413 			SOCKBUF_UNLOCK(&so->so_snd);
1414 		}
1415 	}
1416 
1417 	cp->l2t_idx = l2t_idx;
1418 
1419 	/* This will be initialized on ESTABLISHED. */
1420 	cp->emss = 0;
1421 }
1422 
1423 int
1424 negative_advice(int status)
1425 {
1426 
1427 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
1428 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
1429 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
1430 }
1431 
1432 static int
1433 alloc_tid_tab(struct tid_info *t, int flags)
1434 {
1435 
1436 	MPASS(t->ntids > 0);
1437 	MPASS(t->tid_tab == NULL);
1438 
1439 	t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE,
1440 	    M_ZERO | flags);
1441 	if (t->tid_tab == NULL)
1442 		return (ENOMEM);
1443 	atomic_store_rel_int(&t->tids_in_use, 0);
1444 
1445 	return (0);
1446 }
1447 
1448 static void
1449 free_tid_tab(struct tid_info *t)
1450 {
1451 
1452 	KASSERT(t->tids_in_use == 0,
1453 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
1454 
1455 	free(t->tid_tab, M_CXGBE);
1456 	t->tid_tab = NULL;
1457 }
1458 
1459 static int
1460 alloc_stid_tab(struct tid_info *t, int flags)
1461 {
1462 
1463 	MPASS(t->nstids > 0);
1464 	MPASS(t->stid_tab == NULL);
1465 
1466 	t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE,
1467 	    M_ZERO | flags);
1468 	if (t->stid_tab == NULL)
1469 		return (ENOMEM);
1470 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
1471 	t->stids_in_use = 0;
1472 	TAILQ_INIT(&t->stids);
1473 	t->nstids_free_head = t->nstids;
1474 
1475 	return (0);
1476 }
1477 
1478 static void
1479 free_stid_tab(struct tid_info *t)
1480 {
1481 
1482 	KASSERT(t->stids_in_use == 0,
1483 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
1484 
1485 	if (mtx_initialized(&t->stid_lock))
1486 		mtx_destroy(&t->stid_lock);
1487 	free(t->stid_tab, M_CXGBE);
1488 	t->stid_tab = NULL;
1489 }
1490 
1491 static void
1492 free_tid_tabs(struct tid_info *t)
1493 {
1494 
1495 	free_tid_tab(t);
1496 	free_stid_tab(t);
1497 }
1498 
1499 static int
1500 alloc_tid_tabs(struct tid_info *t)
1501 {
1502 	int rc;
1503 
1504 	rc = alloc_tid_tab(t, M_NOWAIT);
1505 	if (rc != 0)
1506 		goto failed;
1507 
1508 	rc = alloc_stid_tab(t, M_NOWAIT);
1509 	if (rc != 0)
1510 		goto failed;
1511 
1512 	return (0);
1513 failed:
1514 	free_tid_tabs(t);
1515 	return (rc);
1516 }
1517 
1518 static inline void
1519 alloc_tcb_history(struct adapter *sc, struct tom_data *td)
1520 {
1521 
1522 	if (sc->tids.ntids == 0 || sc->tids.ntids > 1024)
1523 		return;
1524 	rw_init(&td->tcb_history_lock, "TCB history");
1525 	td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history),
1526 	    M_CXGBE, M_ZERO | M_NOWAIT);
1527 	td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0));
1528 }
1529 
1530 static inline void
1531 free_tcb_history(struct adapter *sc, struct tom_data *td)
1532 {
1533 #ifdef INVARIANTS
1534 	int i;
1535 
1536 	if (td->tcb_history != NULL) {
1537 		for (i = 0; i < sc->tids.ntids; i++) {
1538 			MPASS(td->tcb_history[i] == NULL);
1539 		}
1540 	}
1541 #endif
1542 	free(td->tcb_history, M_CXGBE);
1543 	if (rw_initialized(&td->tcb_history_lock))
1544 		rw_destroy(&td->tcb_history_lock);
1545 }
1546 
1547 static void
1548 free_tom_data(struct adapter *sc, struct tom_data *td)
1549 {
1550 
1551 	ASSERT_SYNCHRONIZED_OP(sc);
1552 
1553 	KASSERT(TAILQ_EMPTY(&td->toep_list),
1554 	    ("%s: TOE PCB list is not empty.", __func__));
1555 	KASSERT(td->lctx_count == 0,
1556 	    ("%s: lctx hash table is not empty.", __func__));
1557 
1558 	t4_free_ppod_region(&td->pr);
1559 
1560 	if (td->listen_mask != 0)
1561 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
1562 
1563 	if (mtx_initialized(&td->unsent_wr_lock))
1564 		mtx_destroy(&td->unsent_wr_lock);
1565 	if (mtx_initialized(&td->lctx_hash_lock))
1566 		mtx_destroy(&td->lctx_hash_lock);
1567 	if (mtx_initialized(&td->toep_list_lock))
1568 		mtx_destroy(&td->toep_list_lock);
1569 
1570 	free_tcb_history(sc, td);
1571 	free_tid_tabs(&sc->tids);
1572 	free(td, M_CXGBE);
1573 }
1574 
1575 static char *
1576 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen,
1577     int *buflen)
1578 {
1579 	char *pkt;
1580 	struct tcphdr *th;
1581 	int ipv6, len;
1582 	const int maxlen =
1583 	    max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) +
1584 	    max(sizeof(struct ip), sizeof(struct ip6_hdr)) +
1585 	    sizeof(struct tcphdr);
1586 
1587 	MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN);
1588 
1589 	pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT);
1590 	if (pkt == NULL)
1591 		return (NULL);
1592 
1593 	ipv6 = inp->inp_vflag & INP_IPV6;
1594 	len = 0;
1595 
1596 	if (EVL_VLANOFTAG(vtag) == 0xfff) {
1597 		struct ether_header *eh = (void *)pkt;
1598 
1599 		if (ipv6)
1600 			eh->ether_type = htons(ETHERTYPE_IPV6);
1601 		else
1602 			eh->ether_type = htons(ETHERTYPE_IP);
1603 
1604 		len += sizeof(*eh);
1605 	} else {
1606 		struct ether_vlan_header *evh = (void *)pkt;
1607 
1608 		evh->evl_encap_proto = htons(ETHERTYPE_VLAN);
1609 		evh->evl_tag = htons(vtag);
1610 		if (ipv6)
1611 			evh->evl_proto = htons(ETHERTYPE_IPV6);
1612 		else
1613 			evh->evl_proto = htons(ETHERTYPE_IP);
1614 
1615 		len += sizeof(*evh);
1616 	}
1617 
1618 	if (ipv6) {
1619 		struct ip6_hdr *ip6 = (void *)&pkt[len];
1620 
1621 		ip6->ip6_vfc = IPV6_VERSION;
1622 		ip6->ip6_plen = htons(sizeof(struct tcphdr));
1623 		ip6->ip6_nxt = IPPROTO_TCP;
1624 		if (open_type == OPEN_TYPE_ACTIVE) {
1625 			ip6->ip6_src = inp->in6p_laddr;
1626 			ip6->ip6_dst = inp->in6p_faddr;
1627 		} else if (open_type == OPEN_TYPE_LISTEN) {
1628 			ip6->ip6_src = inp->in6p_laddr;
1629 			ip6->ip6_dst = ip6->ip6_src;
1630 		}
1631 
1632 		len += sizeof(*ip6);
1633 	} else {
1634 		struct ip *ip = (void *)&pkt[len];
1635 
1636 		ip->ip_v = IPVERSION;
1637 		ip->ip_hl = sizeof(*ip) >> 2;
1638 		ip->ip_tos = inp->inp_ip_tos;
1639 		ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr));
1640 		ip->ip_ttl = inp->inp_ip_ttl;
1641 		ip->ip_p = IPPROTO_TCP;
1642 		if (open_type == OPEN_TYPE_ACTIVE) {
1643 			ip->ip_src = inp->inp_laddr;
1644 			ip->ip_dst = inp->inp_faddr;
1645 		} else if (open_type == OPEN_TYPE_LISTEN) {
1646 			ip->ip_src = inp->inp_laddr;
1647 			ip->ip_dst = ip->ip_src;
1648 		}
1649 
1650 		len += sizeof(*ip);
1651 	}
1652 
1653 	th = (void *)&pkt[len];
1654 	if (open_type == OPEN_TYPE_ACTIVE) {
1655 		th->th_sport = inp->inp_lport;	/* network byte order already */
1656 		th->th_dport = inp->inp_fport;	/* ditto */
1657 	} else if (open_type == OPEN_TYPE_LISTEN) {
1658 		th->th_sport = inp->inp_lport;	/* network byte order already */
1659 		th->th_dport = th->th_sport;
1660 	}
1661 	len += sizeof(th);
1662 
1663 	*pktlen = *buflen = len;
1664 	return (pkt);
1665 }
1666 
1667 const struct offload_settings *
1668 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m,
1669     uint16_t vtag, struct inpcb *inp)
1670 {
1671 	const struct t4_offload_policy *op;
1672 	char *pkt;
1673 	struct offload_rule *r;
1674 	int i, matched, pktlen, buflen;
1675 	static const struct offload_settings allow_offloading_settings = {
1676 		.offload = 1,
1677 		.rx_coalesce = -1,
1678 		.cong_algo = -1,
1679 		.sched_class = -1,
1680 		.tstamp = -1,
1681 		.sack = -1,
1682 		.nagle = -1,
1683 		.ecn = -1,
1684 		.ddp = -1,
1685 		.tls = -1,
1686 		.txq = QUEUE_RANDOM,
1687 		.rxq = QUEUE_RANDOM,
1688 		.mss = -1,
1689 	};
1690 	static const struct offload_settings disallow_offloading_settings = {
1691 		.offload = 0,
1692 		/* rest is irrelevant when offload is off. */
1693 	};
1694 
1695 	rw_assert(&sc->policy_lock, RA_LOCKED);
1696 
1697 	/*
1698 	 * If there's no Connection Offloading Policy attached to the device
1699 	 * then we need to return a default static policy.  If
1700 	 * "cop_managed_offloading" is true, then we need to disallow
1701 	 * offloading until a COP is attached to the device.  Otherwise we
1702 	 * allow offloading ...
1703 	 */
1704 	op = sc->policy;
1705 	if (op == NULL) {
1706 		if (sc->tt.cop_managed_offloading)
1707 			return (&disallow_offloading_settings);
1708 		else
1709 			return (&allow_offloading_settings);
1710 	}
1711 
1712 	switch (open_type) {
1713 	case OPEN_TYPE_ACTIVE:
1714 	case OPEN_TYPE_LISTEN:
1715 		pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen);
1716 		break;
1717 	case OPEN_TYPE_PASSIVE:
1718 		MPASS(m != NULL);
1719 		pkt = mtod(m, char *);
1720 		MPASS(*pkt == CPL_PASS_ACCEPT_REQ);
1721 		pkt += sizeof(struct cpl_pass_accept_req);
1722 		pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req);
1723 		buflen = m->m_len - sizeof(struct cpl_pass_accept_req);
1724 		break;
1725 	default:
1726 		MPASS(0);
1727 		return (&disallow_offloading_settings);
1728 	}
1729 
1730 	if (pkt == NULL || pktlen == 0 || buflen == 0)
1731 		return (&disallow_offloading_settings);
1732 
1733 	matched = 0;
1734 	r = &op->rule[0];
1735 	for (i = 0; i < op->nrules; i++, r++) {
1736 		if (r->open_type != open_type &&
1737 		    r->open_type != OPEN_TYPE_DONTCARE) {
1738 			continue;
1739 		}
1740 		matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen);
1741 		if (matched)
1742 			break;
1743 	}
1744 
1745 	if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN)
1746 		free(pkt, M_CXGBE);
1747 
1748 	return (matched ? &r->settings : &disallow_offloading_settings);
1749 }
1750 
1751 static void
1752 reclaim_wr_resources(void *arg, int count)
1753 {
1754 	struct tom_data *td = arg;
1755 	STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
1756 	struct cpl_act_open_req *cpl;
1757 	u_int opcode, atid, tid;
1758 	struct wrqe *wr;
1759 	struct adapter *sc = td_adapter(td);
1760 
1761 	mtx_lock(&td->unsent_wr_lock);
1762 	STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
1763 	mtx_unlock(&td->unsent_wr_lock);
1764 
1765 	while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
1766 		STAILQ_REMOVE_HEAD(&twr_list, link);
1767 
1768 		cpl = wrtod(wr);
1769 		opcode = GET_OPCODE(cpl);
1770 
1771 		switch (opcode) {
1772 		case CPL_ACT_OPEN_REQ:
1773 		case CPL_ACT_OPEN_REQ6:
1774 			atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1775 			CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1776 			act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1777 			free(wr, M_CXGBE);
1778 			break;
1779 		case CPL_PASS_ACCEPT_RPL:
1780 			tid = GET_TID(cpl);
1781 			CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid);
1782 			synack_failure_cleanup(sc, tid);
1783 			free(wr, M_CXGBE);
1784 			break;
1785 		default:
1786 			log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1787 			    "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1788 			/* WR not freed here; go look at it with a debugger.  */
1789 		}
1790 	}
1791 }
1792 
1793 /*
1794  * Ground control to Major TOM
1795  * Commencing countdown, engines on
1796  */
1797 static int
1798 t4_tom_activate(struct adapter *sc)
1799 {
1800 	struct tom_data *td;
1801 	struct toedev *tod;
1802 	struct vi_info *vi;
1803 	int i, rc, v;
1804 
1805 	ASSERT_SYNCHRONIZED_OP(sc);
1806 
1807 	/* per-adapter softc for TOM */
1808 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1809 	if (td == NULL)
1810 		return (ENOMEM);
1811 
1812 	/* List of TOE PCBs and associated lock */
1813 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1814 	TAILQ_INIT(&td->toep_list);
1815 
1816 	/* Listen context */
1817 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1818 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1819 	    &td->listen_mask, HASH_NOWAIT);
1820 
1821 	/* List of WRs for which L2 resolution failed */
1822 	mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1823 	STAILQ_INIT(&td->unsent_wr_list);
1824 	TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1825 
1826 	/* TID tables */
1827 	rc = alloc_tid_tabs(&sc->tids);
1828 	if (rc != 0)
1829 		goto done;
1830 
1831 	rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1832 	    t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1833 	if (rc != 0)
1834 		goto done;
1835 	t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1836 	    V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1837 
1838 	alloc_tcb_history(sc, td);
1839 
1840 	/* toedev ops */
1841 	tod = &td->tod;
1842 	init_toedev(tod);
1843 	tod->tod_softc = sc;
1844 	tod->tod_connect = t4_connect;
1845 	tod->tod_listen_start = t4_listen_start;
1846 	tod->tod_listen_stop = t4_listen_stop;
1847 	tod->tod_rcvd = t4_rcvd;
1848 	tod->tod_output = t4_tod_output;
1849 	tod->tod_send_rst = t4_send_rst;
1850 	tod->tod_send_fin = t4_send_fin;
1851 	tod->tod_pcb_detach = t4_pcb_detach;
1852 	tod->tod_l2_update = t4_l2_update;
1853 	tod->tod_syncache_added = t4_syncache_added;
1854 	tod->tod_syncache_removed = t4_syncache_removed;
1855 	tod->tod_syncache_respond = t4_syncache_respond;
1856 	tod->tod_offload_socket = t4_offload_socket;
1857 	tod->tod_ctloutput = t4_ctloutput;
1858 	tod->tod_tcp_info = t4_tcp_info;
1859 #ifdef KERN_TLS
1860 	tod->tod_alloc_tls_session = t4_alloc_tls_session;
1861 #endif
1862 	tod->tod_pmtu_update = t4_pmtu_update;
1863 
1864 	for_each_port(sc, i) {
1865 		for_each_vi(sc->port[i], v, vi) {
1866 			SETTOEDEV(vi->ifp, &td->tod);
1867 		}
1868 	}
1869 
1870 	sc->tom_softc = td;
1871 	register_toedev(sc->tom_softc);
1872 
1873 done:
1874 	if (rc != 0)
1875 		free_tom_data(sc, td);
1876 	return (rc);
1877 }
1878 
1879 static int
1880 t4_tom_deactivate(struct adapter *sc)
1881 {
1882 	int rc = 0;
1883 	struct tom_data *td = sc->tom_softc;
1884 
1885 	ASSERT_SYNCHRONIZED_OP(sc);
1886 
1887 	if (td == NULL)
1888 		return (0);	/* XXX. KASSERT? */
1889 
1890 	if (sc->offload_map != 0)
1891 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1892 
1893 	if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1894 		return (EBUSY);	/* both iWARP and iSCSI rely on the TOE. */
1895 
1896 	mtx_lock(&td->toep_list_lock);
1897 	if (!TAILQ_EMPTY(&td->toep_list))
1898 		rc = EBUSY;
1899 	mtx_unlock(&td->toep_list_lock);
1900 
1901 	mtx_lock(&td->lctx_hash_lock);
1902 	if (td->lctx_count > 0)
1903 		rc = EBUSY;
1904 	mtx_unlock(&td->lctx_hash_lock);
1905 
1906 	taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1907 	mtx_lock(&td->unsent_wr_lock);
1908 	if (!STAILQ_EMPTY(&td->unsent_wr_list))
1909 		rc = EBUSY;
1910 	mtx_unlock(&td->unsent_wr_lock);
1911 
1912 	if (rc == 0) {
1913 		unregister_toedev(sc->tom_softc);
1914 		free_tom_data(sc, td);
1915 		sc->tom_softc = NULL;
1916 	}
1917 
1918 	return (rc);
1919 }
1920 
1921 static int
1922 t4_ctloutput_tom(struct socket *so, struct sockopt *sopt)
1923 {
1924 	struct tcpcb *tp = sototcpcb(so);
1925 	struct toepcb *toep = tp->t_toe;
1926 	int error, optval;
1927 
1928 	if (sopt->sopt_level == IPPROTO_TCP && sopt->sopt_name == TCP_USE_DDP) {
1929 		if (sopt->sopt_dir != SOPT_SET)
1930 			return (EOPNOTSUPP);
1931 
1932 		if (sopt->sopt_td != NULL) {
1933 			/* Only settable by the kernel. */
1934 			return (EPERM);
1935 		}
1936 
1937 		error = sooptcopyin(sopt, &optval, sizeof(optval),
1938 		    sizeof(optval));
1939 		if (error != 0)
1940 			return (error);
1941 
1942 		if (optval != 0)
1943 			return (t4_enable_ddp_rcv(so, toep));
1944 		else
1945 			return (EOPNOTSUPP);
1946 	}
1947 	return (tcp_ctloutput(so, sopt));
1948 }
1949 
1950 static int
1951 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1952 {
1953 	struct tcpcb *tp = sototcpcb(so);
1954 	struct toepcb *toep = tp->t_toe;
1955 	int error;
1956 
1957 	/*
1958 	 * No lock is needed as TOE sockets never change between
1959 	 * active and passive.
1960 	 */
1961 	if (SOLISTENING(so))
1962 		return (EINVAL);
1963 
1964 	if (ulp_mode(toep) == ULP_MODE_TCPDDP ||
1965 	    ulp_mode(toep) == ULP_MODE_NONE) {
1966 		error = t4_aio_queue_ddp(so, job);
1967 		if (error != EOPNOTSUPP)
1968 			return (error);
1969 	}
1970 
1971 	return (t4_aio_queue_aiotx(so, job));
1972 }
1973 
1974 static int
1975 t4_tom_mod_load(void)
1976 {
1977 	/* CPL handlers */
1978 	t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl);
1979 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2,
1980 	    CPL_COOKIE_TOM);
1981 	t4_init_connect_cpl_handlers();
1982 	t4_init_listen_cpl_handlers();
1983 	t4_init_cpl_io_handlers();
1984 
1985 	t4_ddp_mod_load();
1986 	t4_tls_mod_load();
1987 
1988 	bcopy(&tcp_protosw, &toe_protosw, sizeof(toe_protosw));
1989 	toe_protosw.pr_ctloutput = t4_ctloutput_tom;
1990 	toe_protosw.pr_aio_queue = t4_aio_queue_tom;
1991 
1992 	bcopy(&tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
1993 	toe6_protosw.pr_ctloutput = t4_ctloutput_tom;
1994 	toe6_protosw.pr_aio_queue = t4_aio_queue_tom;
1995 
1996 	return (t4_register_uld(&tom_uld_info));
1997 }
1998 
1999 static void
2000 tom_uninit(struct adapter *sc, void *arg __unused)
2001 {
2002 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
2003 		return;
2004 
2005 	/* Try to free resources (works only if no port has IFCAP_TOE) */
2006 	if (uld_active(sc, ULD_TOM))
2007 		t4_deactivate_uld(sc, ULD_TOM);
2008 
2009 	end_synchronized_op(sc, 0);
2010 }
2011 
2012 static int
2013 t4_tom_mod_unload(void)
2014 {
2015 	t4_iterate(tom_uninit, NULL);
2016 
2017 	if (t4_unregister_uld(&tom_uld_info) == EBUSY)
2018 		return (EBUSY);
2019 
2020 	t4_tls_mod_unload();
2021 	t4_ddp_mod_unload();
2022 
2023 	t4_uninit_connect_cpl_handlers();
2024 	t4_uninit_listen_cpl_handlers();
2025 	t4_uninit_cpl_io_handlers();
2026 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM);
2027 	t4_register_cpl_handler(CPL_GET_TCB_RPL, NULL);
2028 
2029 	return (0);
2030 }
2031 #endif	/* TCP_OFFLOAD */
2032 
2033 static int
2034 t4_tom_modevent(module_t mod, int cmd, void *arg)
2035 {
2036 	int rc = 0;
2037 
2038 #ifdef TCP_OFFLOAD
2039 	switch (cmd) {
2040 	case MOD_LOAD:
2041 		rc = t4_tom_mod_load();
2042 		break;
2043 
2044 	case MOD_UNLOAD:
2045 		rc = t4_tom_mod_unload();
2046 		break;
2047 
2048 	default:
2049 		rc = EINVAL;
2050 	}
2051 #else
2052 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
2053 	rc = EOPNOTSUPP;
2054 #endif
2055 	return (rc);
2056 }
2057 
2058 static moduledata_t t4_tom_moddata= {
2059 	"t4_tom",
2060 	t4_tom_modevent,
2061 	0
2062 };
2063 
2064 MODULE_VERSION(t4_tom, 1);
2065 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
2066 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
2067 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
2068