xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision 2e3f49888ec8851bafb22011533217487764fdb0)
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 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
183 		ddp_init_toep(toep);
184 
185 	toep->flags |= TPF_INITIALIZED;
186 
187 	return (0);
188 }
189 
190 struct toepcb *
191 hold_toepcb(struct toepcb *toep)
192 {
193 
194 	refcount_acquire(&toep->refcount);
195 	return (toep);
196 }
197 
198 void
199 free_toepcb(struct toepcb *toep)
200 {
201 
202 	if (refcount_release(&toep->refcount) == 0)
203 		return;
204 
205 	KASSERT(!(toep->flags & TPF_ATTACHED),
206 	    ("%s: attached to an inpcb", __func__));
207 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
208 	    ("%s: CPL pending", __func__));
209 
210 	if (toep->flags & TPF_INITIALIZED) {
211 		if (ulp_mode(toep) == ULP_MODE_TCPDDP)
212 			ddp_uninit_toep(toep);
213 		tls_uninit_toep(toep);
214 	}
215 	free(toep, M_CXGBE);
216 }
217 
218 /*
219  * Set up the socket for TCP offload.
220  */
221 void
222 offload_socket(struct socket *so, struct toepcb *toep)
223 {
224 	struct tom_data *td = toep->td;
225 	struct inpcb *inp = sotoinpcb(so);
226 	struct tcpcb *tp = intotcpcb(inp);
227 	struct sockbuf *sb;
228 
229 	INP_WLOCK_ASSERT(inp);
230 
231 	/* Update socket */
232 	sb = &so->so_snd;
233 	SOCKBUF_LOCK(sb);
234 	sb->sb_flags |= SB_NOCOALESCE;
235 	SOCKBUF_UNLOCK(sb);
236 	sb = &so->so_rcv;
237 	SOCKBUF_LOCK(sb);
238 	sb->sb_flags |= SB_NOCOALESCE;
239 	if (inp->inp_vflag & INP_IPV6)
240 		so->so_proto = &toe6_protosw;
241 	else
242 		so->so_proto = &toe_protosw;
243 	SOCKBUF_UNLOCK(sb);
244 
245 	/* Update TCP PCB */
246 	tp->tod = &td->tod;
247 	tp->t_toe = toep;
248 	tp->t_flags |= TF_TOE;
249 
250 	/* Install an extra hold on inp */
251 	toep->inp = inp;
252 	toep->flags |= TPF_ATTACHED;
253 	in_pcbref(inp);
254 
255 	/* Add the TOE PCB to the active list */
256 	mtx_lock(&td->toep_list_lock);
257 	TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
258 	mtx_unlock(&td->toep_list_lock);
259 }
260 
261 void
262 restore_so_proto(struct socket *so, bool v6)
263 {
264 	if (v6)
265 		so->so_proto = &tcp6_protosw;
266 	else
267 		so->so_proto = &tcp_protosw;
268 }
269 
270 /* This is _not_ the normal way to "unoffload" a socket. */
271 void
272 undo_offload_socket(struct socket *so)
273 {
274 	struct inpcb *inp = sotoinpcb(so);
275 	struct tcpcb *tp = intotcpcb(inp);
276 	struct toepcb *toep = tp->t_toe;
277 	struct tom_data *td = toep->td;
278 	struct sockbuf *sb;
279 
280 	INP_WLOCK_ASSERT(inp);
281 
282 	sb = &so->so_snd;
283 	SOCKBUF_LOCK(sb);
284 	sb->sb_flags &= ~SB_NOCOALESCE;
285 	SOCKBUF_UNLOCK(sb);
286 	sb = &so->so_rcv;
287 	SOCKBUF_LOCK(sb);
288 	sb->sb_flags &= ~SB_NOCOALESCE;
289 	restore_so_proto(so, inp->inp_vflag & INP_IPV6);
290 	SOCKBUF_UNLOCK(sb);
291 
292 	tp->tod = NULL;
293 	tp->t_toe = NULL;
294 	tp->t_flags &= ~TF_TOE;
295 
296 	toep->inp = NULL;
297 	toep->flags &= ~TPF_ATTACHED;
298 	if (in_pcbrele_wlocked(inp))
299 		panic("%s: inp freed.", __func__);
300 
301 	mtx_lock(&td->toep_list_lock);
302 	TAILQ_REMOVE(&td->toep_list, toep, link);
303 	mtx_unlock(&td->toep_list_lock);
304 }
305 
306 static void
307 release_offload_resources(struct toepcb *toep)
308 {
309 	struct tom_data *td = toep->td;
310 	struct adapter *sc = td_adapter(td);
311 	int tid = toep->tid;
312 
313 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
314 	    ("%s: %p has CPL pending.", __func__, toep));
315 	KASSERT(!(toep->flags & TPF_ATTACHED),
316 	    ("%s: %p is still attached.", __func__, toep));
317 
318 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
319 	    __func__, toep, tid, toep->l2te, toep->ce);
320 
321 	/*
322 	 * These queues should have been emptied at approximately the same time
323 	 * that a normal connection's socket's so_snd would have been purged or
324 	 * drained.  Do _not_ clean up here.
325 	 */
326 	MPASS(mbufq_empty(&toep->ulp_pduq));
327 	MPASS(mbufq_empty(&toep->ulp_pdu_reclaimq));
328 #ifdef INVARIANTS
329 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
330 		ddp_assert_empty(toep);
331 #endif
332 	MPASS(TAILQ_EMPTY(&toep->aiotx_jobq));
333 
334 	if (toep->l2te)
335 		t4_l2t_release(toep->l2te);
336 
337 	if (tid >= 0) {
338 		remove_tid(sc, tid, toep->ce ? 2 : 1);
339 		release_tid(sc, tid, toep->ctrlq);
340 	}
341 
342 	if (toep->ce)
343 		t4_release_clip_entry(sc, toep->ce);
344 
345 	if (toep->params.tc_idx != -1)
346 		t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx);
347 
348 	mtx_lock(&td->toep_list_lock);
349 	TAILQ_REMOVE(&td->toep_list, toep, link);
350 	mtx_unlock(&td->toep_list_lock);
351 
352 	free_toepcb(toep);
353 }
354 
355 /*
356  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
357  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
358  * pending CPL) then it is time to release all resources tied to the toepcb.
359  *
360  * Also gets called when an offloaded active open fails and the TOM wants the
361  * kernel to take the TCP PCB back.
362  */
363 static void
364 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
365 {
366 #if defined(KTR) || defined(INVARIANTS)
367 	struct inpcb *inp = tptoinpcb(tp);
368 #endif
369 	struct toepcb *toep = tp->t_toe;
370 
371 	INP_WLOCK_ASSERT(inp);
372 
373 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
374 	KASSERT(toep->flags & TPF_ATTACHED,
375 	    ("%s: not attached", __func__));
376 
377 #ifdef KTR
378 	if (tp->t_state == TCPS_SYN_SENT) {
379 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
380 		    __func__, toep->tid, toep, toep->flags, inp,
381 		    inp->inp_flags);
382 	} else {
383 		CTR6(KTR_CXGBE,
384 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
385 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
386 		    inp->inp_flags);
387 	}
388 #endif
389 
390 	tp->tod = NULL;
391 	tp->t_toe = NULL;
392 	tp->t_flags &= ~TF_TOE;
393 	toep->flags &= ~TPF_ATTACHED;
394 
395 	if (!(toep->flags & TPF_CPL_PENDING))
396 		release_offload_resources(toep);
397 }
398 
399 /*
400  * setsockopt handler.
401  */
402 static void
403 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
404 {
405 	struct adapter *sc = tod->tod_softc;
406 	struct toepcb *toep = tp->t_toe;
407 
408 	if (dir == SOPT_GET)
409 		return;
410 
411 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
412 
413 	switch (name) {
414 	case TCP_NODELAY:
415 		if (tp->t_state != TCPS_ESTABLISHED)
416 			break;
417 		toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
418 		t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
419 		    V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0);
420 		break;
421 	default:
422 		break;
423 	}
424 }
425 
426 static inline uint64_t
427 get_tcb_tflags(const uint64_t *tcb)
428 {
429 
430 	return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32));
431 }
432 
433 static inline uint32_t
434 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift)
435 {
436 #define LAST_WORD ((TCB_SIZE / 4) - 1)
437 	uint64_t t1, t2;
438 	int flit_idx;
439 
440 	MPASS(mask != 0);
441 	MPASS(word <= LAST_WORD);
442 	MPASS(shift < 32);
443 
444 	flit_idx = (LAST_WORD - word) / 2;
445 	if (word & 0x1)
446 		shift += 32;
447 	t1 = be64toh(tcb[flit_idx]) >> shift;
448 	t2 = 0;
449 	if (fls(mask) > 64 - shift) {
450 		/*
451 		 * Will spill over into the next logical flit, which is the flit
452 		 * before this one.  The flit_idx before this one must be valid.
453 		 */
454 		MPASS(flit_idx > 0);
455 		t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift);
456 	}
457 	return ((t2 | t1) & mask);
458 #undef LAST_WORD
459 }
460 #define GET_TCB_FIELD(tcb, F) \
461     get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F)
462 
463 /*
464  * Issues a CPL_GET_TCB to read the entire TCB for the tid.
465  */
466 static int
467 send_get_tcb(struct adapter *sc, u_int tid)
468 {
469 	struct cpl_get_tcb *cpl;
470 	struct wrq_cookie cookie;
471 
472 	MPASS(tid >= sc->tids.tid_base);
473 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
474 
475 	cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16),
476 	    &cookie);
477 	if (__predict_false(cpl == NULL))
478 		return (ENOMEM);
479 	bzero(cpl, sizeof(*cpl));
480 	INIT_TP_WR(cpl, tid);
481 	OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid));
482 	cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) |
483 	    V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id));
484 	cpl->cookie = 0xff;
485 	commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie);
486 
487 	return (0);
488 }
489 
490 static struct tcb_histent *
491 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags)
492 {
493 	struct tcb_histent *te;
494 
495 	MPASS(flags == M_NOWAIT || flags == M_WAITOK);
496 
497 	te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags);
498 	if (te == NULL)
499 		return (NULL);
500 	mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF);
501 	callout_init_mtx(&te->te_callout, &te->te_lock, 0);
502 	te->te_adapter = sc;
503 	te->te_tid = tid;
504 
505 	return (te);
506 }
507 
508 static void
509 free_tcb_histent(struct tcb_histent *te)
510 {
511 
512 	mtx_destroy(&te->te_lock);
513 	free(te, M_CXGBE);
514 }
515 
516 /*
517  * Start tracking the tid in the TCB history.
518  */
519 int
520 add_tid_to_history(struct adapter *sc, u_int tid)
521 {
522 	struct tcb_histent *te = NULL;
523 	struct tom_data *td = sc->tom_softc;
524 	int rc;
525 
526 	MPASS(tid >= sc->tids.tid_base);
527 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
528 
529 	if (td->tcb_history == NULL)
530 		return (ENXIO);
531 
532 	rw_wlock(&td->tcb_history_lock);
533 	if (td->tcb_history[tid] != NULL) {
534 		rc = EEXIST;
535 		goto done;
536 	}
537 	te = alloc_tcb_histent(sc, tid, M_NOWAIT);
538 	if (te == NULL) {
539 		rc = ENOMEM;
540 		goto done;
541 	}
542 	mtx_lock(&te->te_lock);
543 	rc = send_get_tcb(sc, tid);
544 	if (rc == 0) {
545 		te->te_flags |= TE_RPL_PENDING;
546 		td->tcb_history[tid] = te;
547 	} else {
548 		free(te, M_CXGBE);
549 	}
550 	mtx_unlock(&te->te_lock);
551 done:
552 	rw_wunlock(&td->tcb_history_lock);
553 	return (rc);
554 }
555 
556 static void
557 remove_tcb_histent(struct tcb_histent *te)
558 {
559 	struct adapter *sc = te->te_adapter;
560 	struct tom_data *td = sc->tom_softc;
561 
562 	rw_assert(&td->tcb_history_lock, RA_WLOCKED);
563 	mtx_assert(&te->te_lock, MA_OWNED);
564 	MPASS(td->tcb_history[te->te_tid] == te);
565 
566 	td->tcb_history[te->te_tid] = NULL;
567 	free_tcb_histent(te);
568 	rw_wunlock(&td->tcb_history_lock);
569 }
570 
571 static inline struct tcb_histent *
572 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem)
573 {
574 	struct tcb_histent *te;
575 	struct tom_data *td = sc->tom_softc;
576 
577 	MPASS(tid >= sc->tids.tid_base);
578 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
579 
580 	if (td->tcb_history == NULL)
581 		return (NULL);
582 
583 	if (addrem)
584 		rw_wlock(&td->tcb_history_lock);
585 	else
586 		rw_rlock(&td->tcb_history_lock);
587 	te = td->tcb_history[tid];
588 	if (te != NULL) {
589 		mtx_lock(&te->te_lock);
590 		return (te);	/* with both locks held */
591 	}
592 	if (addrem)
593 		rw_wunlock(&td->tcb_history_lock);
594 	else
595 		rw_runlock(&td->tcb_history_lock);
596 
597 	return (te);
598 }
599 
600 static inline void
601 release_tcb_histent(struct tcb_histent *te)
602 {
603 	struct adapter *sc = te->te_adapter;
604 	struct tom_data *td = sc->tom_softc;
605 
606 	mtx_assert(&te->te_lock, MA_OWNED);
607 	mtx_unlock(&te->te_lock);
608 	rw_assert(&td->tcb_history_lock, RA_RLOCKED);
609 	rw_runlock(&td->tcb_history_lock);
610 }
611 
612 static void
613 request_tcb(void *arg)
614 {
615 	struct tcb_histent *te = arg;
616 
617 	mtx_assert(&te->te_lock, MA_OWNED);
618 
619 	/* Noone else is supposed to update the histent. */
620 	MPASS(!(te->te_flags & TE_RPL_PENDING));
621 	if (send_get_tcb(te->te_adapter, te->te_tid) == 0)
622 		te->te_flags |= TE_RPL_PENDING;
623 	else
624 		callout_schedule(&te->te_callout, hz / 100);
625 }
626 
627 static void
628 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb)
629 {
630 	struct tom_data *td = te->te_adapter->tom_softc;
631 	uint64_t tflags = get_tcb_tflags(tcb);
632 	uint8_t sample = 0;
633 
634 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) {
635 		if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0)
636 			sample |= TS_RTO;
637 		if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0)
638 			sample |= TS_DUPACKS;
639 		if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold)
640 			sample |= TS_FASTREXMT;
641 	}
642 
643 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) {
644 		uint32_t snd_wnd;
645 
646 		sample |= TS_SND_BACKLOGGED;	/* for whatever reason. */
647 
648 		snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
649 		if (tflags & V_TF_RECV_SCALE(1))
650 			snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE);
651 		if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd)
652 			sample |= TS_CWND_LIMITED;	/* maybe due to CWND */
653 	}
654 
655 	if (tflags & V_TF_CCTRL_ECN(1)) {
656 
657 		/*
658 		 * CE marker on incoming IP hdr, echoing ECE back in the TCP
659 		 * hdr.  Indicates congestion somewhere on the way from the peer
660 		 * to this node.
661 		 */
662 		if (tflags & V_TF_CCTRL_ECE(1))
663 			sample |= TS_ECN_ECE;
664 
665 		/*
666 		 * ECE seen and CWR sent (or about to be sent).  Might indicate
667 		 * congestion on the way to the peer.  This node is reducing its
668 		 * congestion window in response.
669 		 */
670 		if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1)))
671 			sample |= TS_ECN_CWR;
672 	}
673 
674 	te->te_sample[te->te_pidx] = sample;
675 	if (++te->te_pidx == nitems(te->te_sample))
676 		te->te_pidx = 0;
677 	memcpy(te->te_tcb, tcb, TCB_SIZE);
678 	te->te_flags |= TE_ACTIVE;
679 }
680 
681 static int
682 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
683 {
684 	struct adapter *sc = iq->adapter;
685 	const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *);
686 	const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1);
687 	struct tcb_histent *te;
688 	const u_int tid = GET_TID(cpl);
689 	bool remove;
690 
691 	remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED;
692 	te = lookup_tcb_histent(sc, tid, remove);
693 	if (te == NULL) {
694 		/* Not in the history.  Who issued the GET_TCB for this? */
695 		device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, "
696 		    "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid,
697 		    (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE),
698 		    GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE),
699 		    GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie);
700 		goto done;
701 	}
702 
703 	MPASS(te->te_flags & TE_RPL_PENDING);
704 	te->te_flags &= ~TE_RPL_PENDING;
705 	if (remove) {
706 		remove_tcb_histent(te);
707 	} else {
708 		update_tcb_histent(te, tcb);
709 		callout_reset(&te->te_callout, hz / 10, request_tcb, te);
710 		release_tcb_histent(te);
711 	}
712 done:
713 	m_freem(m);
714 	return (0);
715 }
716 
717 static void
718 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti)
719 {
720 	uint32_t v;
721 
722 	ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE);
723 
724 	v = GET_TCB_FIELD(tcb, T_SRTT);
725 	ti->tcpi_rtt = tcp_ticks_to_us(sc, v);
726 
727 	v = GET_TCB_FIELD(tcb, T_RTTVAR);
728 	ti->tcpi_rttvar = tcp_ticks_to_us(sc, v);
729 
730 	ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH);
731 	ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND);
732 	ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT);
733 	ti->tcpi_rcv_adv = GET_TCB_FIELD(tcb, RCV_ADV);
734 	ti->tcpi_dupacks = GET_TCB_FIELD(tcb, T_DUPACKS);
735 
736 	v = GET_TCB_FIELD(tcb, TX_MAX);
737 	ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW);
738 	ti->tcpi_snd_una = v - GET_TCB_FIELD(tcb, SND_UNA_RAW);
739 	ti->tcpi_snd_max = v - GET_TCB_FIELD(tcb, SND_MAX_RAW);
740 
741 	/* Receive window being advertised by us. */
742 	ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE);	/* Yes, SND. */
743 	ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND);
744 
745 	/* Send window */
746 	ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE);	/* Yes, RCV. */
747 	ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
748 	if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1))
749 		ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale;
750 	else
751 		ti->tcpi_snd_wscale = 0;
752 
753 }
754 
755 static void
756 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te,
757     struct tcp_info *ti)
758 {
759 
760 	fill_tcp_info_from_tcb(sc, te->te_tcb, ti);
761 }
762 
763 /*
764  * Reads the TCB for the given tid using a memory window and copies it to 'buf'
765  * in the same format as CPL_GET_TCB_RPL.
766  */
767 static void
768 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf)
769 {
770 	int i, j, k, rc;
771 	uint32_t addr;
772 	u_char *tcb, tmp;
773 
774 	MPASS(tid >= sc->tids.tid_base);
775 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
776 
777 	addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE;
778 	rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE);
779 	if (rc != 0)
780 		return;
781 
782 	tcb = (u_char *)buf;
783 	for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) {
784 		for (k = 0; k < 16; k++) {
785 			tmp = tcb[i + k];
786 			tcb[i + k] = tcb[j + k];
787 			tcb[j + k] = tmp;
788 		}
789 	}
790 }
791 
792 static void
793 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti)
794 {
795 	uint64_t tcb[TCB_SIZE / sizeof(uint64_t)];
796 	struct tcb_histent *te;
797 
798 	ti->tcpi_toe_tid = tid;
799 	te = lookup_tcb_histent(sc, tid, false);
800 	if (te != NULL) {
801 		fill_tcp_info_from_history(sc, te, ti);
802 		release_tcb_histent(te);
803 	} else {
804 		if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) {
805 			/* XXX: tell firmware to flush TCB cache. */
806 		}
807 		read_tcb_using_memwin(sc, tid, tcb);
808 		fill_tcp_info_from_tcb(sc, tcb, ti);
809 	}
810 }
811 
812 /*
813  * Called by the kernel to allow the TOE driver to "refine" values filled up in
814  * the tcp_info for an offloaded connection.
815  */
816 static void
817 t4_tcp_info(struct toedev *tod, const struct tcpcb *tp, struct tcp_info *ti)
818 {
819 	struct adapter *sc = tod->tod_softc;
820 	struct toepcb *toep = tp->t_toe;
821 
822 	INP_LOCK_ASSERT(tptoinpcb(tp));
823 	MPASS(ti != NULL);
824 
825 	fill_tcp_info(sc, toep->tid, ti);
826 }
827 
828 #ifdef KERN_TLS
829 static int
830 t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp,
831     struct ktls_session *tls, int direction)
832 {
833 	struct toepcb *toep = tp->t_toe;
834 
835 	INP_WLOCK_ASSERT(tptoinpcb(tp));
836 	MPASS(tls != NULL);
837 
838 	return (tls_alloc_ktls(toep, tls, direction));
839 }
840 #endif
841 
842 /* SET_TCB_FIELD sent as a ULP command looks like this */
843 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
844     sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
845 
846 static void *
847 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, uint64_t word, uint64_t mask,
848 		uint64_t val, uint32_t tid)
849 {
850 	struct ulptx_idata *ulpsc;
851 	struct cpl_set_tcb_field_core *req;
852 
853 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
854 	ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
855 
856 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
857 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
858 	ulpsc->len = htobe32(sizeof(*req));
859 
860 	req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
861 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, tid));
862 	req->reply_ctrl = htobe16(V_NO_REPLY(1));
863 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
864 	req->mask = htobe64(mask);
865 	req->val = htobe64(val);
866 
867 	ulpsc = (struct ulptx_idata *)(req + 1);
868 	if (LEN__SET_TCB_FIELD_ULP % 16) {
869 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
870 		ulpsc->len = htobe32(0);
871 		return (ulpsc + 1);
872 	}
873 	return (ulpsc);
874 }
875 
876 static void
877 send_mss_flowc_wr(struct adapter *sc, struct toepcb *toep)
878 {
879 	struct wrq_cookie cookie;
880 	struct fw_flowc_wr *flowc;
881 	struct ofld_tx_sdesc *txsd;
882 	const int flowclen = sizeof(*flowc) + sizeof(struct fw_flowc_mnemval);
883 	const int flowclen16 = howmany(flowclen, 16);
884 
885 	if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0) {
886 		CH_ERR(sc, "%s: tid %u out of tx credits (%d, %d).\n", __func__,
887 		    toep->tid, toep->tx_credits, toep->txsd_avail);
888 		return;
889 	}
890 
891 	flowc = start_wrq_wr(&toep->ofld_txq->wrq, flowclen16, &cookie);
892 	if (__predict_false(flowc == NULL)) {
893 		CH_ERR(sc, "ENOMEM in %s for tid %u.\n", __func__, toep->tid);
894 		return;
895 	}
896 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
897 	    V_FW_FLOWC_WR_NPARAMS(1));
898 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
899 	    V_FW_WR_FLOWID(toep->tid));
900 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_MSS;
901 	flowc->mnemval[0].val = htobe32(toep->params.emss);
902 
903 	txsd = &toep->txsd[toep->txsd_pidx];
904 	txsd->tx_credits = flowclen16;
905 	txsd->plen = 0;
906 	toep->tx_credits -= txsd->tx_credits;
907 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
908 		toep->txsd_pidx = 0;
909 	toep->txsd_avail--;
910 	commit_wrq_wr(&toep->ofld_txq->wrq, flowc, &cookie);
911 }
912 
913 static void
914 t4_pmtu_update(struct toedev *tod, struct tcpcb *tp, tcp_seq seq, int mtu)
915 {
916 	struct work_request_hdr *wrh;
917 	struct ulp_txpkt *ulpmc;
918 	int idx, len;
919 	struct wrq_cookie cookie;
920 	struct inpcb *inp = tptoinpcb(tp);
921 	struct toepcb *toep = tp->t_toe;
922 	struct adapter *sc = td_adapter(toep->td);
923 	unsigned short *mtus = &sc->params.mtus[0];
924 
925 	INP_WLOCK_ASSERT(inp);
926 	MPASS(mtu > 0);	/* kernel is supposed to provide something usable. */
927 
928 	/* tp->snd_una and snd_max are in host byte order too. */
929 	seq = be32toh(seq);
930 
931 	CTR6(KTR_CXGBE, "%s: tid %d, seq 0x%08x, mtu %u, mtu_idx %u (%d)",
932 	    __func__, toep->tid, seq, mtu, toep->params.mtu_idx,
933 	    mtus[toep->params.mtu_idx]);
934 
935 	if (ulp_mode(toep) == ULP_MODE_NONE &&	/* XXX: Read TCB otherwise? */
936 	    (SEQ_LT(seq, tp->snd_una) || SEQ_GEQ(seq, tp->snd_max))) {
937 		CTR5(KTR_CXGBE,
938 		    "%s: tid %d, seq 0x%08x not in range [0x%08x, 0x%08x).",
939 		    __func__, toep->tid, seq, tp->snd_una, tp->snd_max);
940 		return;
941 	}
942 
943 	/* Find the best mtu_idx for the suggested MTU. */
944 	for (idx = 0; idx < NMTUS - 1 && mtus[idx + 1] <= mtu; idx++)
945 		continue;
946 	if (idx >= toep->params.mtu_idx)
947 		return;	/* Never increase the PMTU (just like the kernel). */
948 
949 	/*
950 	 * We'll send a compound work request with 2 SET_TCB_FIELDs -- the first
951 	 * one updates the mtu_idx and the second one triggers a retransmit.
952 	 */
953 	len = sizeof(*wrh) + 2 * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
954 	wrh = start_wrq_wr(toep->ctrlq, howmany(len, 16), &cookie);
955 	if (wrh == NULL) {
956 		CH_ERR(sc, "failed to change mtu_idx of tid %d (%u -> %u).\n",
957 		    toep->tid, toep->params.mtu_idx, idx);
958 		return;
959 	}
960 	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
961 	ulpmc = (struct ulp_txpkt *)(wrh + 1);
962 	ulpmc = mk_set_tcb_field_ulp(ulpmc, W_TCB_T_MAXSEG,
963 	    V_TCB_T_MAXSEG(M_TCB_T_MAXSEG), V_TCB_T_MAXSEG(idx), toep->tid);
964 	ulpmc = mk_set_tcb_field_ulp(ulpmc, W_TCB_TIMESTAMP,
965 	    V_TCB_TIMESTAMP(0x7FFFFULL << 11), 0, toep->tid);
966 	commit_wrq_wr(toep->ctrlq, wrh, &cookie);
967 
968 	/* Update the software toepcb and tcpcb. */
969 	toep->params.mtu_idx = idx;
970 	tp->t_maxseg = mtus[toep->params.mtu_idx];
971 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
972 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
973 	else
974 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
975 	toep->params.emss = tp->t_maxseg;
976 	if (tp->t_flags & TF_RCVD_TSTMP)
977 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
978 
979 	/* Update the firmware flowc. */
980 	send_mss_flowc_wr(sc, toep);
981 
982 	/* Update the MTU in the kernel's hostcache. */
983 	if (sc->tt.update_hc_on_pmtu_change != 0) {
984 		struct in_conninfo inc = {0};
985 
986 		inc.inc_fibnum = inp->inp_inc.inc_fibnum;
987 		if (inp->inp_inc.inc_flags & INC_ISIPV6) {
988 			inc.inc_flags |= INC_ISIPV6;
989 			inc.inc6_faddr = inp->inp_inc.inc6_faddr;
990 		} else {
991 			inc.inc_faddr = inp->inp_inc.inc_faddr;
992 		}
993 		tcp_hc_updatemtu(&inc, mtu);
994 	}
995 
996 	CTR6(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u), t_maxseg %u, emss %u",
997 	    __func__, toep->tid, toep->params.mtu_idx,
998 	    mtus[toep->params.mtu_idx], tp->t_maxseg, toep->params.emss);
999 }
1000 
1001 /*
1002  * The TOE driver will not receive any more CPLs for the tid associated with the
1003  * toepcb; release the hold on the inpcb.
1004  */
1005 void
1006 final_cpl_received(struct toepcb *toep)
1007 {
1008 	struct inpcb *inp = toep->inp;
1009 	bool need_wakeup;
1010 
1011 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
1012 	INP_WLOCK_ASSERT(inp);
1013 	KASSERT(toep->flags & TPF_CPL_PENDING,
1014 	    ("%s: CPL not pending already?", __func__));
1015 
1016 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
1017 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
1018 
1019 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1020 		release_ddp_resources(toep);
1021 	toep->inp = NULL;
1022 	need_wakeup = (toep->flags & TPF_WAITING_FOR_FINAL) != 0;
1023 	toep->flags &= ~(TPF_CPL_PENDING | TPF_WAITING_FOR_FINAL);
1024 	mbufq_drain(&toep->ulp_pduq);
1025 	mbufq_drain(&toep->ulp_pdu_reclaimq);
1026 
1027 	if (!(toep->flags & TPF_ATTACHED))
1028 		release_offload_resources(toep);
1029 
1030 	if (!in_pcbrele_wlocked(inp))
1031 		INP_WUNLOCK(inp);
1032 
1033 	if (need_wakeup) {
1034 		struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep);
1035 
1036 		mtx_lock(lock);
1037 		wakeup(toep);
1038 		mtx_unlock(lock);
1039 	}
1040 }
1041 
1042 void
1043 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
1044 {
1045 	struct tid_info *t = &sc->tids;
1046 
1047 	MPASS(tid >= t->tid_base);
1048 	MPASS(tid - t->tid_base < t->ntids);
1049 
1050 	t->tid_tab[tid - t->tid_base] = ctx;
1051 	atomic_add_int(&t->tids_in_use, ntids);
1052 }
1053 
1054 void *
1055 lookup_tid(struct adapter *sc, int tid)
1056 {
1057 	struct tid_info *t = &sc->tids;
1058 
1059 	return (t->tid_tab[tid - t->tid_base]);
1060 }
1061 
1062 void
1063 update_tid(struct adapter *sc, int tid, void *ctx)
1064 {
1065 	struct tid_info *t = &sc->tids;
1066 
1067 	t->tid_tab[tid - t->tid_base] = ctx;
1068 }
1069 
1070 void
1071 remove_tid(struct adapter *sc, int tid, int ntids)
1072 {
1073 	struct tid_info *t = &sc->tids;
1074 
1075 	t->tid_tab[tid - t->tid_base] = NULL;
1076 	atomic_subtract_int(&t->tids_in_use, ntids);
1077 }
1078 
1079 /*
1080  * What mtu_idx to use, given a 4-tuple.  Note that both s->mss and tcp_mssopt
1081  * have the MSS that we should advertise in our SYN.  Advertised MSS doesn't
1082  * account for any TCP options so the effective MSS (only payload, no headers or
1083  * options) could be different.
1084  */
1085 static int
1086 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc,
1087     struct offload_settings *s)
1088 {
1089 	unsigned short *mtus = &sc->params.mtus[0];
1090 	int i, mss, mtu;
1091 
1092 	MPASS(inc != NULL);
1093 
1094 	mss = s->mss > 0 ? s->mss : tcp_mssopt(inc);
1095 	if (inc->inc_flags & INC_ISIPV6)
1096 		mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1097 	else
1098 		mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr);
1099 
1100 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++)
1101 		continue;
1102 
1103 	return (i);
1104 }
1105 
1106 /*
1107  * Determine the receive window size for a socket.
1108  */
1109 u_long
1110 select_rcv_wnd(struct socket *so)
1111 {
1112 	unsigned long wnd;
1113 
1114 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1115 
1116 	wnd = sbspace(&so->so_rcv);
1117 	if (wnd < MIN_RCV_WND)
1118 		wnd = MIN_RCV_WND;
1119 
1120 	return min(wnd, MAX_RCV_WND);
1121 }
1122 
1123 int
1124 select_rcv_wscale(void)
1125 {
1126 	int wscale = 0;
1127 	unsigned long space = sb_max;
1128 
1129 	if (space > MAX_RCV_WND)
1130 		space = MAX_RCV_WND;
1131 
1132 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
1133 		wscale++;
1134 
1135 	return (wscale);
1136 }
1137 
1138 __be64
1139 calc_options0(struct vi_info *vi, struct conn_params *cp)
1140 {
1141 	uint64_t opt0 = 0;
1142 
1143 	opt0 |= F_TCAM_BYPASS;
1144 
1145 	MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE);
1146 	opt0 |= V_WND_SCALE(cp->wscale);
1147 
1148 	MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS);
1149 	opt0 |= V_MSS_IDX(cp->mtu_idx);
1150 
1151 	MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE);
1152 	opt0 |= V_ULP_MODE(cp->ulp_mode);
1153 
1154 	MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ);
1155 	opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize);
1156 
1157 	MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->adapter->vres.l2t.size);
1158 	opt0 |= V_L2T_IDX(cp->l2t_idx);
1159 
1160 	opt0 |= V_SMAC_SEL(vi->smt_idx);
1161 	opt0 |= V_TX_CHAN(vi->pi->tx_chan);
1162 
1163 	MPASS(cp->keepalive == 0 || cp->keepalive == 1);
1164 	opt0 |= V_KEEP_ALIVE(cp->keepalive);
1165 
1166 	MPASS(cp->nagle == 0 || cp->nagle == 1);
1167 	opt0 |= V_NAGLE(cp->nagle);
1168 
1169 	return (htobe64(opt0));
1170 }
1171 
1172 __be32
1173 calc_options2(struct vi_info *vi, struct conn_params *cp)
1174 {
1175 	uint32_t opt2 = 0;
1176 	struct port_info *pi = vi->pi;
1177 	struct adapter *sc = pi->adapter;
1178 
1179 	/*
1180 	 * rx flow control, rx coalesce, congestion control, and tx pace are all
1181 	 * explicitly set by the driver.  On T5+ the ISS is also set by the
1182 	 * driver to the value picked by the kernel.
1183 	 */
1184 	if (is_t4(sc)) {
1185 		opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID;
1186 		opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID;
1187 	} else {
1188 		opt2 |= F_T5_OPT_2_VALID;	/* all 4 valid */
1189 		opt2 |= F_T5_ISS;		/* ISS provided in CPL */
1190 	}
1191 
1192 	MPASS(cp->sack == 0 || cp->sack == 1);
1193 	opt2 |= V_SACK_EN(cp->sack);
1194 
1195 	MPASS(cp->tstamp == 0 || cp->tstamp == 1);
1196 	opt2 |= V_TSTAMPS_EN(cp->tstamp);
1197 
1198 	if (cp->wscale > 0)
1199 		opt2 |= F_WND_SCALE_EN;
1200 
1201 	MPASS(cp->ecn == 0 || cp->ecn == 1);
1202 	opt2 |= V_CCTRL_ECN(cp->ecn);
1203 
1204 	/* XXX: F_RX_CHANNEL for multiple rx c-chan support goes here. */
1205 
1206 	opt2 |= V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]);
1207 	opt2 |= V_PACE(0);
1208 	opt2 |= F_RSS_QUEUE_VALID;
1209 	opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id);
1210 
1211 	MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL);
1212 	opt2 |= V_CONG_CNTRL(cp->cong_algo);
1213 
1214 	MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1);
1215 	if (cp->rx_coalesce == 1)
1216 		opt2 |= V_RX_COALESCE(M_RX_COALESCE);
1217 
1218 	opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
1219 #ifdef USE_DDP_RX_FLOW_CONTROL
1220 	if (cp->ulp_mode == ULP_MODE_TCPDDP)
1221 		opt2 |= F_RX_FC_DDP;
1222 #endif
1223 
1224 	return (htobe32(opt2));
1225 }
1226 
1227 uint64_t
1228 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
1229 {
1230 	struct adapter *sc = vi->adapter;
1231 	struct tp_params *tp = &sc->params.tp;
1232 	uint64_t ntuple = 0;
1233 
1234 	/*
1235 	 * Initialize each of the fields which we care about which are present
1236 	 * in the Compressed Filter Tuple.
1237 	 */
1238 	if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE)
1239 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
1240 
1241 	if (tp->port_shift >= 0)
1242 		ntuple |= (uint64_t)e->lport << tp->port_shift;
1243 
1244 	if (tp->protocol_shift >= 0)
1245 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
1246 
1247 	if (tp->vnic_shift >= 0 && tp->vnic_mode == FW_VNIC_MODE_PF_VF) {
1248 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) |
1249 		    V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) <<
1250 		    tp->vnic_shift;
1251 	}
1252 
1253 	if (is_t4(sc))
1254 		return (htobe32((uint32_t)ntuple));
1255 	else
1256 		return (htobe64(V_FILTER_TUPLE(ntuple)));
1257 }
1258 
1259 /*
1260  * Initialize various connection parameters.
1261  */
1262 void
1263 init_conn_params(struct vi_info *vi , struct offload_settings *s,
1264     struct in_conninfo *inc, struct socket *so,
1265     const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp)
1266 {
1267 	struct port_info *pi = vi->pi;
1268 	struct adapter *sc = pi->adapter;
1269 	struct tom_tunables *tt = &sc->tt;
1270 	struct inpcb *inp = sotoinpcb(so);
1271 	struct tcpcb *tp = intotcpcb(inp);
1272 	u_long wnd;
1273 	u_int q_idx;
1274 
1275 	MPASS(s->offload != 0);
1276 
1277 	/* Congestion control algorithm */
1278 	if (s->cong_algo >= 0)
1279 		cp->cong_algo = s->cong_algo & M_CONG_CNTRL;
1280 	else if (sc->tt.cong_algorithm >= 0)
1281 		cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL;
1282 	else {
1283 		struct cc_algo *cc = CC_ALGO(tp);
1284 
1285 		if (strcasecmp(cc->name, "reno") == 0)
1286 			cp->cong_algo = CONG_ALG_RENO;
1287 		else if (strcasecmp(cc->name, "tahoe") == 0)
1288 			cp->cong_algo = CONG_ALG_TAHOE;
1289 		if (strcasecmp(cc->name, "newreno") == 0)
1290 			cp->cong_algo = CONG_ALG_NEWRENO;
1291 		if (strcasecmp(cc->name, "highspeed") == 0)
1292 			cp->cong_algo = CONG_ALG_HIGHSPEED;
1293 		else {
1294 			/*
1295 			 * Use newreno in case the algorithm selected by the
1296 			 * host stack is not supported by the hardware.
1297 			 */
1298 			cp->cong_algo = CONG_ALG_NEWRENO;
1299 		}
1300 	}
1301 
1302 	/* Tx traffic scheduling class. */
1303 	if (s->sched_class >= 0 && s->sched_class < sc->params.nsched_cls)
1304 		cp->tc_idx = s->sched_class;
1305 	else
1306 		cp->tc_idx = -1;
1307 
1308 	/* Nagle's algorithm. */
1309 	if (s->nagle >= 0)
1310 		cp->nagle = s->nagle > 0 ? 1 : 0;
1311 	else
1312 		cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
1313 
1314 	/* TCP Keepalive. */
1315 	if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE)
1316 		cp->keepalive = 1;
1317 	else
1318 		cp->keepalive = 0;
1319 
1320 	/* Optimization that's specific to T5 @ 40G. */
1321 	if (tt->tx_align >= 0)
1322 		cp->tx_align =  tt->tx_align > 0 ? 1 : 0;
1323 	else if (chip_id(sc) == CHELSIO_T5 &&
1324 	    (port_top_speed(pi) > 10 || sc->params.nports > 2))
1325 		cp->tx_align = 1;
1326 	else
1327 		cp->tx_align = 0;
1328 
1329 	/* ULP mode. */
1330 	if (s->ddp > 0 ||
1331 	    (s->ddp < 0 && sc->tt.ddp && (so_options_get(so) & SO_NO_DDP) == 0))
1332 		cp->ulp_mode = ULP_MODE_TCPDDP;
1333 	else
1334 		cp->ulp_mode = ULP_MODE_NONE;
1335 
1336 	/* Rx coalescing. */
1337 	if (s->rx_coalesce >= 0)
1338 		cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0;
1339 	else if (tt->rx_coalesce >= 0)
1340 		cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0;
1341 	else
1342 		cp->rx_coalesce = 1;	/* default */
1343 
1344 	/*
1345 	 * Index in the PMTU table.  This controls the MSS that we announce in
1346 	 * our SYN initially, but after ESTABLISHED it controls the MSS that we
1347 	 * use to send data.
1348 	 */
1349 	cp->mtu_idx = find_best_mtu_idx(sc, inc, s);
1350 
1351 	/* Tx queue for this connection. */
1352 	if (s->txq == QUEUE_RANDOM)
1353 		q_idx = arc4random();
1354 	else if (s->txq == QUEUE_ROUNDROBIN)
1355 		q_idx = atomic_fetchadd_int(&vi->txq_rr, 1);
1356 	else
1357 		q_idx = s->txq;
1358 	cp->txq_idx = vi->first_ofld_txq + q_idx % vi->nofldtxq;
1359 
1360 	/* Rx queue for this connection. */
1361 	if (s->rxq == QUEUE_RANDOM)
1362 		q_idx = arc4random();
1363 	else if (s->rxq == QUEUE_ROUNDROBIN)
1364 		q_idx = atomic_fetchadd_int(&vi->rxq_rr, 1);
1365 	else
1366 		q_idx = s->rxq;
1367 	cp->rxq_idx = vi->first_ofld_rxq + q_idx % vi->nofldrxq;
1368 
1369 	if (SOLISTENING(so)) {
1370 		/* Passive open */
1371 		MPASS(tcpopt != NULL);
1372 
1373 		/* TCP timestamp option */
1374 		if (tcpopt->tstamp &&
1375 		    (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323)))
1376 			cp->tstamp = 1;
1377 		else
1378 			cp->tstamp = 0;
1379 
1380 		/* SACK */
1381 		if (tcpopt->sack &&
1382 		    (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack)))
1383 			cp->sack = 1;
1384 		else
1385 			cp->sack = 0;
1386 
1387 		/* Receive window scaling. */
1388 		if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323)
1389 			cp->wscale = select_rcv_wscale();
1390 		else
1391 			cp->wscale = 0;
1392 
1393 		/* ECN */
1394 		if (tcpopt->ecn &&	/* XXX: review. */
1395 		    (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn)))
1396 			cp->ecn = 1;
1397 		else
1398 			cp->ecn = 0;
1399 
1400 		wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND);
1401 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1402 
1403 		if (tt->sndbuf > 0)
1404 			cp->sndbuf = tt->sndbuf;
1405 		else if (so->sol_sbsnd_flags & SB_AUTOSIZE &&
1406 		    V_tcp_do_autosndbuf)
1407 			cp->sndbuf = 256 * 1024;
1408 		else
1409 			cp->sndbuf = so->sol_sbsnd_hiwat;
1410 	} else {
1411 		/* Active open */
1412 
1413 		/* TCP timestamp option */
1414 		if (s->tstamp > 0 ||
1415 		    (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP)))
1416 			cp->tstamp = 1;
1417 		else
1418 			cp->tstamp = 0;
1419 
1420 		/* SACK */
1421 		if (s->sack > 0 ||
1422 		    (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT)))
1423 			cp->sack = 1;
1424 		else
1425 			cp->sack = 0;
1426 
1427 		/* Receive window scaling */
1428 		if (tp->t_flags & TF_REQ_SCALE)
1429 			cp->wscale = select_rcv_wscale();
1430 		else
1431 			cp->wscale = 0;
1432 
1433 		/* ECN */
1434 		if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1))
1435 			cp->ecn = 1;
1436 		else
1437 			cp->ecn = 0;
1438 
1439 		SOCKBUF_LOCK(&so->so_rcv);
1440 		wnd = max(select_rcv_wnd(so), MIN_RCV_WND);
1441 		SOCKBUF_UNLOCK(&so->so_rcv);
1442 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1443 
1444 		if (tt->sndbuf > 0)
1445 			cp->sndbuf = tt->sndbuf;
1446 		else {
1447 			SOCKBUF_LOCK(&so->so_snd);
1448 			if (so->so_snd.sb_flags & SB_AUTOSIZE &&
1449 			    V_tcp_do_autosndbuf)
1450 				cp->sndbuf = 256 * 1024;
1451 			else
1452 				cp->sndbuf = so->so_snd.sb_hiwat;
1453 			SOCKBUF_UNLOCK(&so->so_snd);
1454 		}
1455 	}
1456 
1457 	cp->l2t_idx = l2t_idx;
1458 
1459 	/* This will be initialized on ESTABLISHED. */
1460 	cp->emss = 0;
1461 }
1462 
1463 int
1464 negative_advice(int status)
1465 {
1466 
1467 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
1468 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
1469 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
1470 }
1471 
1472 static int
1473 alloc_tid_tab(struct tid_info *t, int flags)
1474 {
1475 
1476 	MPASS(t->ntids > 0);
1477 	MPASS(t->tid_tab == NULL);
1478 
1479 	t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE,
1480 	    M_ZERO | flags);
1481 	if (t->tid_tab == NULL)
1482 		return (ENOMEM);
1483 	atomic_store_rel_int(&t->tids_in_use, 0);
1484 
1485 	return (0);
1486 }
1487 
1488 static void
1489 free_tid_tab(struct tid_info *t)
1490 {
1491 
1492 	KASSERT(t->tids_in_use == 0,
1493 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
1494 
1495 	free(t->tid_tab, M_CXGBE);
1496 	t->tid_tab = NULL;
1497 }
1498 
1499 static int
1500 alloc_stid_tab(struct tid_info *t, int flags)
1501 {
1502 
1503 	MPASS(t->nstids > 0);
1504 	MPASS(t->stid_tab == NULL);
1505 
1506 	t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE,
1507 	    M_ZERO | flags);
1508 	if (t->stid_tab == NULL)
1509 		return (ENOMEM);
1510 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
1511 	t->stids_in_use = 0;
1512 	TAILQ_INIT(&t->stids);
1513 	t->nstids_free_head = t->nstids;
1514 
1515 	return (0);
1516 }
1517 
1518 static void
1519 free_stid_tab(struct tid_info *t)
1520 {
1521 
1522 	KASSERT(t->stids_in_use == 0,
1523 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
1524 
1525 	if (mtx_initialized(&t->stid_lock))
1526 		mtx_destroy(&t->stid_lock);
1527 	free(t->stid_tab, M_CXGBE);
1528 	t->stid_tab = NULL;
1529 }
1530 
1531 static void
1532 free_tid_tabs(struct tid_info *t)
1533 {
1534 
1535 	free_tid_tab(t);
1536 	free_stid_tab(t);
1537 }
1538 
1539 static int
1540 alloc_tid_tabs(struct tid_info *t)
1541 {
1542 	int rc;
1543 
1544 	rc = alloc_tid_tab(t, M_NOWAIT);
1545 	if (rc != 0)
1546 		goto failed;
1547 
1548 	rc = alloc_stid_tab(t, M_NOWAIT);
1549 	if (rc != 0)
1550 		goto failed;
1551 
1552 	return (0);
1553 failed:
1554 	free_tid_tabs(t);
1555 	return (rc);
1556 }
1557 
1558 static inline void
1559 alloc_tcb_history(struct adapter *sc, struct tom_data *td)
1560 {
1561 
1562 	if (sc->tids.ntids == 0 || sc->tids.ntids > 1024)
1563 		return;
1564 	rw_init(&td->tcb_history_lock, "TCB history");
1565 	td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history),
1566 	    M_CXGBE, M_ZERO | M_NOWAIT);
1567 	td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0));
1568 }
1569 
1570 static inline void
1571 free_tcb_history(struct adapter *sc, struct tom_data *td)
1572 {
1573 #ifdef INVARIANTS
1574 	int i;
1575 
1576 	if (td->tcb_history != NULL) {
1577 		for (i = 0; i < sc->tids.ntids; i++) {
1578 			MPASS(td->tcb_history[i] == NULL);
1579 		}
1580 	}
1581 #endif
1582 	free(td->tcb_history, M_CXGBE);
1583 	if (rw_initialized(&td->tcb_history_lock))
1584 		rw_destroy(&td->tcb_history_lock);
1585 }
1586 
1587 static void
1588 free_tom_data(struct adapter *sc, struct tom_data *td)
1589 {
1590 
1591 	ASSERT_SYNCHRONIZED_OP(sc);
1592 
1593 	KASSERT(TAILQ_EMPTY(&td->toep_list),
1594 	    ("%s: TOE PCB list is not empty.", __func__));
1595 	KASSERT(td->lctx_count == 0,
1596 	    ("%s: lctx hash table is not empty.", __func__));
1597 
1598 	t4_free_ppod_region(&td->pr);
1599 
1600 	if (td->listen_mask != 0)
1601 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
1602 
1603 	if (mtx_initialized(&td->unsent_wr_lock))
1604 		mtx_destroy(&td->unsent_wr_lock);
1605 	if (mtx_initialized(&td->lctx_hash_lock))
1606 		mtx_destroy(&td->lctx_hash_lock);
1607 	if (mtx_initialized(&td->toep_list_lock))
1608 		mtx_destroy(&td->toep_list_lock);
1609 
1610 	free_tcb_history(sc, td);
1611 	free_tid_tabs(&sc->tids);
1612 	free(td, M_CXGBE);
1613 }
1614 
1615 static char *
1616 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen,
1617     int *buflen)
1618 {
1619 	char *pkt;
1620 	struct tcphdr *th;
1621 	int ipv6, len;
1622 	const int maxlen =
1623 	    max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) +
1624 	    max(sizeof(struct ip), sizeof(struct ip6_hdr)) +
1625 	    sizeof(struct tcphdr);
1626 
1627 	MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN);
1628 
1629 	pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT);
1630 	if (pkt == NULL)
1631 		return (NULL);
1632 
1633 	ipv6 = inp->inp_vflag & INP_IPV6;
1634 	len = 0;
1635 
1636 	if (EVL_VLANOFTAG(vtag) == 0xfff) {
1637 		struct ether_header *eh = (void *)pkt;
1638 
1639 		if (ipv6)
1640 			eh->ether_type = htons(ETHERTYPE_IPV6);
1641 		else
1642 			eh->ether_type = htons(ETHERTYPE_IP);
1643 
1644 		len += sizeof(*eh);
1645 	} else {
1646 		struct ether_vlan_header *evh = (void *)pkt;
1647 
1648 		evh->evl_encap_proto = htons(ETHERTYPE_VLAN);
1649 		evh->evl_tag = htons(vtag);
1650 		if (ipv6)
1651 			evh->evl_proto = htons(ETHERTYPE_IPV6);
1652 		else
1653 			evh->evl_proto = htons(ETHERTYPE_IP);
1654 
1655 		len += sizeof(*evh);
1656 	}
1657 
1658 	if (ipv6) {
1659 		struct ip6_hdr *ip6 = (void *)&pkt[len];
1660 
1661 		ip6->ip6_vfc = IPV6_VERSION;
1662 		ip6->ip6_plen = htons(sizeof(struct tcphdr));
1663 		ip6->ip6_nxt = IPPROTO_TCP;
1664 		if (open_type == OPEN_TYPE_ACTIVE) {
1665 			ip6->ip6_src = inp->in6p_laddr;
1666 			ip6->ip6_dst = inp->in6p_faddr;
1667 		} else if (open_type == OPEN_TYPE_LISTEN) {
1668 			ip6->ip6_src = inp->in6p_laddr;
1669 			ip6->ip6_dst = ip6->ip6_src;
1670 		}
1671 
1672 		len += sizeof(*ip6);
1673 	} else {
1674 		struct ip *ip = (void *)&pkt[len];
1675 
1676 		ip->ip_v = IPVERSION;
1677 		ip->ip_hl = sizeof(*ip) >> 2;
1678 		ip->ip_tos = inp->inp_ip_tos;
1679 		ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr));
1680 		ip->ip_ttl = inp->inp_ip_ttl;
1681 		ip->ip_p = IPPROTO_TCP;
1682 		if (open_type == OPEN_TYPE_ACTIVE) {
1683 			ip->ip_src = inp->inp_laddr;
1684 			ip->ip_dst = inp->inp_faddr;
1685 		} else if (open_type == OPEN_TYPE_LISTEN) {
1686 			ip->ip_src = inp->inp_laddr;
1687 			ip->ip_dst = ip->ip_src;
1688 		}
1689 
1690 		len += sizeof(*ip);
1691 	}
1692 
1693 	th = (void *)&pkt[len];
1694 	if (open_type == OPEN_TYPE_ACTIVE) {
1695 		th->th_sport = inp->inp_lport;	/* network byte order already */
1696 		th->th_dport = inp->inp_fport;	/* ditto */
1697 	} else if (open_type == OPEN_TYPE_LISTEN) {
1698 		th->th_sport = inp->inp_lport;	/* network byte order already */
1699 		th->th_dport = th->th_sport;
1700 	}
1701 	len += sizeof(th);
1702 
1703 	*pktlen = *buflen = len;
1704 	return (pkt);
1705 }
1706 
1707 const struct offload_settings *
1708 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m,
1709     uint16_t vtag, struct inpcb *inp)
1710 {
1711 	const struct t4_offload_policy *op;
1712 	char *pkt;
1713 	struct offload_rule *r;
1714 	int i, matched, pktlen, buflen;
1715 	static const struct offload_settings allow_offloading_settings = {
1716 		.offload = 1,
1717 		.rx_coalesce = -1,
1718 		.cong_algo = -1,
1719 		.sched_class = -1,
1720 		.tstamp = -1,
1721 		.sack = -1,
1722 		.nagle = -1,
1723 		.ecn = -1,
1724 		.ddp = -1,
1725 		.tls = -1,
1726 		.txq = QUEUE_RANDOM,
1727 		.rxq = QUEUE_RANDOM,
1728 		.mss = -1,
1729 	};
1730 	static const struct offload_settings disallow_offloading_settings = {
1731 		.offload = 0,
1732 		/* rest is irrelevant when offload is off. */
1733 	};
1734 
1735 	rw_assert(&sc->policy_lock, RA_LOCKED);
1736 
1737 	/*
1738 	 * If there's no Connection Offloading Policy attached to the device
1739 	 * then we need to return a default static policy.  If
1740 	 * "cop_managed_offloading" is true, then we need to disallow
1741 	 * offloading until a COP is attached to the device.  Otherwise we
1742 	 * allow offloading ...
1743 	 */
1744 	op = sc->policy;
1745 	if (op == NULL) {
1746 		if (sc->tt.cop_managed_offloading)
1747 			return (&disallow_offloading_settings);
1748 		else
1749 			return (&allow_offloading_settings);
1750 	}
1751 
1752 	switch (open_type) {
1753 	case OPEN_TYPE_ACTIVE:
1754 	case OPEN_TYPE_LISTEN:
1755 		pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen);
1756 		break;
1757 	case OPEN_TYPE_PASSIVE:
1758 		MPASS(m != NULL);
1759 		pkt = mtod(m, char *);
1760 		MPASS(*pkt == CPL_PASS_ACCEPT_REQ);
1761 		pkt += sizeof(struct cpl_pass_accept_req);
1762 		pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req);
1763 		buflen = m->m_len - sizeof(struct cpl_pass_accept_req);
1764 		break;
1765 	default:
1766 		MPASS(0);
1767 		return (&disallow_offloading_settings);
1768 	}
1769 
1770 	if (pkt == NULL || pktlen == 0 || buflen == 0)
1771 		return (&disallow_offloading_settings);
1772 
1773 	matched = 0;
1774 	r = &op->rule[0];
1775 	for (i = 0; i < op->nrules; i++, r++) {
1776 		if (r->open_type != open_type &&
1777 		    r->open_type != OPEN_TYPE_DONTCARE) {
1778 			continue;
1779 		}
1780 		matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen);
1781 		if (matched)
1782 			break;
1783 	}
1784 
1785 	if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN)
1786 		free(pkt, M_CXGBE);
1787 
1788 	return (matched ? &r->settings : &disallow_offloading_settings);
1789 }
1790 
1791 static void
1792 reclaim_wr_resources(void *arg, int count)
1793 {
1794 	struct tom_data *td = arg;
1795 	STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
1796 	struct cpl_act_open_req *cpl;
1797 	u_int opcode, atid, tid;
1798 	struct wrqe *wr;
1799 	struct adapter *sc = td_adapter(td);
1800 
1801 	mtx_lock(&td->unsent_wr_lock);
1802 	STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
1803 	mtx_unlock(&td->unsent_wr_lock);
1804 
1805 	while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
1806 		STAILQ_REMOVE_HEAD(&twr_list, link);
1807 
1808 		cpl = wrtod(wr);
1809 		opcode = GET_OPCODE(cpl);
1810 
1811 		switch (opcode) {
1812 		case CPL_ACT_OPEN_REQ:
1813 		case CPL_ACT_OPEN_REQ6:
1814 			atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1815 			CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1816 			act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1817 			free(wr, M_CXGBE);
1818 			break;
1819 		case CPL_PASS_ACCEPT_RPL:
1820 			tid = GET_TID(cpl);
1821 			CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid);
1822 			synack_failure_cleanup(sc, tid);
1823 			free(wr, M_CXGBE);
1824 			break;
1825 		default:
1826 			log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1827 			    "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1828 			/* WR not freed here; go look at it with a debugger.  */
1829 		}
1830 	}
1831 }
1832 
1833 /*
1834  * Ground control to Major TOM
1835  * Commencing countdown, engines on
1836  */
1837 static int
1838 t4_tom_activate(struct adapter *sc)
1839 {
1840 	struct tom_data *td;
1841 	struct toedev *tod;
1842 	struct vi_info *vi;
1843 	int i, rc, v;
1844 
1845 	ASSERT_SYNCHRONIZED_OP(sc);
1846 
1847 	/* per-adapter softc for TOM */
1848 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1849 	if (td == NULL)
1850 		return (ENOMEM);
1851 
1852 	/* List of TOE PCBs and associated lock */
1853 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1854 	TAILQ_INIT(&td->toep_list);
1855 
1856 	/* Listen context */
1857 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1858 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1859 	    &td->listen_mask, HASH_NOWAIT);
1860 
1861 	/* List of WRs for which L2 resolution failed */
1862 	mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1863 	STAILQ_INIT(&td->unsent_wr_list);
1864 	TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1865 
1866 	/* TID tables */
1867 	rc = alloc_tid_tabs(&sc->tids);
1868 	if (rc != 0)
1869 		goto done;
1870 
1871 	rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1872 	    t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1873 	if (rc != 0)
1874 		goto done;
1875 	t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1876 	    V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1877 
1878 	alloc_tcb_history(sc, td);
1879 
1880 	/* toedev ops */
1881 	tod = &td->tod;
1882 	init_toedev(tod);
1883 	tod->tod_softc = sc;
1884 	tod->tod_connect = t4_connect;
1885 	tod->tod_listen_start = t4_listen_start;
1886 	tod->tod_listen_stop = t4_listen_stop;
1887 	tod->tod_rcvd = t4_rcvd;
1888 	tod->tod_output = t4_tod_output;
1889 	tod->tod_send_rst = t4_send_rst;
1890 	tod->tod_send_fin = t4_send_fin;
1891 	tod->tod_pcb_detach = t4_pcb_detach;
1892 	tod->tod_l2_update = t4_l2_update;
1893 	tod->tod_syncache_added = t4_syncache_added;
1894 	tod->tod_syncache_removed = t4_syncache_removed;
1895 	tod->tod_syncache_respond = t4_syncache_respond;
1896 	tod->tod_offload_socket = t4_offload_socket;
1897 	tod->tod_ctloutput = t4_ctloutput;
1898 	tod->tod_tcp_info = t4_tcp_info;
1899 #ifdef KERN_TLS
1900 	tod->tod_alloc_tls_session = t4_alloc_tls_session;
1901 #endif
1902 	tod->tod_pmtu_update = t4_pmtu_update;
1903 
1904 	for_each_port(sc, i) {
1905 		for_each_vi(sc->port[i], v, vi) {
1906 			SETTOEDEV(vi->ifp, &td->tod);
1907 		}
1908 	}
1909 
1910 	sc->tom_softc = td;
1911 	register_toedev(sc->tom_softc);
1912 
1913 done:
1914 	if (rc != 0)
1915 		free_tom_data(sc, td);
1916 	return (rc);
1917 }
1918 
1919 static int
1920 t4_tom_deactivate(struct adapter *sc)
1921 {
1922 	int rc = 0;
1923 	struct tom_data *td = sc->tom_softc;
1924 
1925 	ASSERT_SYNCHRONIZED_OP(sc);
1926 
1927 	if (td == NULL)
1928 		return (0);	/* XXX. KASSERT? */
1929 
1930 	if (sc->offload_map != 0)
1931 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1932 
1933 	if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1934 		return (EBUSY);	/* both iWARP and iSCSI rely on the TOE. */
1935 
1936 	mtx_lock(&td->toep_list_lock);
1937 	if (!TAILQ_EMPTY(&td->toep_list))
1938 		rc = EBUSY;
1939 	mtx_unlock(&td->toep_list_lock);
1940 
1941 	mtx_lock(&td->lctx_hash_lock);
1942 	if (td->lctx_count > 0)
1943 		rc = EBUSY;
1944 	mtx_unlock(&td->lctx_hash_lock);
1945 
1946 	taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1947 	mtx_lock(&td->unsent_wr_lock);
1948 	if (!STAILQ_EMPTY(&td->unsent_wr_list))
1949 		rc = EBUSY;
1950 	mtx_unlock(&td->unsent_wr_lock);
1951 
1952 	if (rc == 0) {
1953 		unregister_toedev(sc->tom_softc);
1954 		free_tom_data(sc, td);
1955 		sc->tom_softc = NULL;
1956 	}
1957 
1958 	return (rc);
1959 }
1960 
1961 static int
1962 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1963 {
1964 	struct tcpcb *tp = sototcpcb(so);
1965 	struct toepcb *toep = tp->t_toe;
1966 	int error;
1967 
1968 	/*
1969 	 * No lock is needed as TOE sockets never change between
1970 	 * active and passive.
1971 	 */
1972 	if (SOLISTENING(so))
1973 		return (EINVAL);
1974 
1975 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1976 		error = t4_aio_queue_ddp(so, job);
1977 		if (error != EOPNOTSUPP)
1978 			return (error);
1979 	}
1980 
1981 	return (t4_aio_queue_aiotx(so, job));
1982 }
1983 
1984 static int
1985 t4_tom_mod_load(void)
1986 {
1987 	/* CPL handlers */
1988 	t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl);
1989 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2,
1990 	    CPL_COOKIE_TOM);
1991 	t4_init_connect_cpl_handlers();
1992 	t4_init_listen_cpl_handlers();
1993 	t4_init_cpl_io_handlers();
1994 
1995 	t4_ddp_mod_load();
1996 	t4_tls_mod_load();
1997 
1998 	bcopy(&tcp_protosw, &toe_protosw, sizeof(toe_protosw));
1999 	toe_protosw.pr_aio_queue = t4_aio_queue_tom;
2000 
2001 	bcopy(&tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
2002 	toe6_protosw.pr_aio_queue = t4_aio_queue_tom;
2003 
2004 	return (t4_register_uld(&tom_uld_info));
2005 }
2006 
2007 static void
2008 tom_uninit(struct adapter *sc, void *arg __unused)
2009 {
2010 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
2011 		return;
2012 
2013 	/* Try to free resources (works only if no port has IFCAP_TOE) */
2014 	if (uld_active(sc, ULD_TOM))
2015 		t4_deactivate_uld(sc, ULD_TOM);
2016 
2017 	end_synchronized_op(sc, 0);
2018 }
2019 
2020 static int
2021 t4_tom_mod_unload(void)
2022 {
2023 	t4_iterate(tom_uninit, NULL);
2024 
2025 	if (t4_unregister_uld(&tom_uld_info) == EBUSY)
2026 		return (EBUSY);
2027 
2028 	t4_tls_mod_unload();
2029 	t4_ddp_mod_unload();
2030 
2031 	t4_uninit_connect_cpl_handlers();
2032 	t4_uninit_listen_cpl_handlers();
2033 	t4_uninit_cpl_io_handlers();
2034 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM);
2035 	t4_register_cpl_handler(CPL_GET_TCB_RPL, NULL);
2036 
2037 	return (0);
2038 }
2039 #endif	/* TCP_OFFLOAD */
2040 
2041 static int
2042 t4_tom_modevent(module_t mod, int cmd, void *arg)
2043 {
2044 	int rc = 0;
2045 
2046 #ifdef TCP_OFFLOAD
2047 	switch (cmd) {
2048 	case MOD_LOAD:
2049 		rc = t4_tom_mod_load();
2050 		break;
2051 
2052 	case MOD_UNLOAD:
2053 		rc = t4_tom_mod_unload();
2054 		break;
2055 
2056 	default:
2057 		rc = EINVAL;
2058 	}
2059 #else
2060 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
2061 	rc = EOPNOTSUPP;
2062 #endif
2063 	return (rc);
2064 }
2065 
2066 static moduledata_t t4_tom_moddata= {
2067 	"t4_tom",
2068 	t4_tom_modevent,
2069 	0
2070 };
2071 
2072 MODULE_VERSION(t4_tom, 1);
2073 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
2074 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
2075 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
2076