xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision 9a3f7fb46c93eb0142a7c32bff3dbda2cd29196e)
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_activate = t4_tom_activate,
95 	.uld_deactivate = t4_tom_deactivate,
96 };
97 
98 static void release_offload_resources(struct toepcb *);
99 static void done_with_toepcb(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 
315 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
316 	    __func__, toep, tid, toep->l2te, toep->ce);
317 
318 	if (toep->l2te) {
319 		t4_l2t_release(toep->l2te);
320 		toep->l2te = NULL;
321 	}
322 	if (tid >= 0) {
323 		remove_tid(sc, tid, toep->ce ? 2 : 1);
324 		release_tid(sc, tid, toep->ctrlq);
325 		toep->tid = -1;
326 	}
327 	if (toep->ce) {
328 		t4_release_clip_entry(sc, toep->ce);
329 		toep->ce = NULL;
330 	}
331 	if (toep->params.tc_idx != -1)
332 		t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx);
333 }
334 
335 /*
336  * Both the driver and kernel are done with the toepcb.
337  */
338 static void
339 done_with_toepcb(struct toepcb *toep)
340 {
341 	struct tom_data *td = toep->td;
342 
343 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
344 	    ("%s: %p has CPL pending.", __func__, toep));
345 	KASSERT(!(toep->flags & TPF_ATTACHED),
346 	    ("%s: %p is still attached.", __func__, toep));
347 
348 	CTR(KTR_CXGBE, "%s: toep %p (0x%x)", __func__, toep, toep->flags);
349 
350 	/*
351 	 * These queues should have been emptied at approximately the same time
352 	 * that a normal connection's socket's so_snd would have been purged or
353 	 * drained.  Do _not_ clean up here.
354 	 */
355 	MPASS(mbufq_empty(&toep->ulp_pduq));
356 	MPASS(mbufq_empty(&toep->ulp_pdu_reclaimq));
357 #ifdef INVARIANTS
358 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
359 		ddp_assert_empty(toep);
360 #endif
361 	MPASS(TAILQ_EMPTY(&toep->aiotx_jobq));
362 	MPASS(toep->tid == -1);
363 	MPASS(toep->l2te == NULL);
364 	MPASS(toep->ce == NULL);
365 
366 	mtx_lock(&td->toep_list_lock);
367 	TAILQ_REMOVE(&td->toep_list, toep, link);
368 	mtx_unlock(&td->toep_list_lock);
369 
370 	free_toepcb(toep);
371 }
372 
373 /*
374  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
375  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
376  * pending CPL) then it is time to release all resources tied to the toepcb.
377  *
378  * Also gets called when an offloaded active open fails and the TOM wants the
379  * kernel to take the TCP PCB back.
380  */
381 void
382 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
383 {
384 #if defined(KTR) || defined(INVARIANTS)
385 	struct inpcb *inp = tptoinpcb(tp);
386 #endif
387 	struct toepcb *toep = tp->t_toe;
388 
389 	INP_WLOCK_ASSERT(inp);
390 
391 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
392 	KASSERT(toep->flags & TPF_ATTACHED,
393 	    ("%s: not attached", __func__));
394 
395 #ifdef KTR
396 	if (tp->t_state == TCPS_SYN_SENT) {
397 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
398 		    __func__, toep->tid, toep, toep->flags, inp,
399 		    inp->inp_flags);
400 	} else {
401 		CTR6(KTR_CXGBE,
402 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
403 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
404 		    inp->inp_flags);
405 	}
406 #endif
407 
408 	tp->tod = NULL;
409 	tp->t_toe = NULL;
410 	tp->t_flags &= ~TF_TOE;
411 	toep->flags &= ~TPF_ATTACHED;
412 
413 	if (!(toep->flags & TPF_CPL_PENDING))
414 		done_with_toepcb(toep);
415 }
416 
417 /*
418  * setsockopt handler.
419  */
420 static void
421 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
422 {
423 	struct adapter *sc = tod->tod_softc;
424 	struct toepcb *toep = tp->t_toe;
425 
426 	if (dir == SOPT_GET)
427 		return;
428 
429 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
430 
431 	switch (name) {
432 	case TCP_NODELAY:
433 		if (tp->t_state != TCPS_ESTABLISHED)
434 			break;
435 		toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
436 		t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
437 		    V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0);
438 		break;
439 	default:
440 		break;
441 	}
442 }
443 
444 static inline uint64_t
445 get_tcb_tflags(const uint64_t *tcb)
446 {
447 
448 	return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32));
449 }
450 
451 static inline uint32_t
452 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift)
453 {
454 #define LAST_WORD ((TCB_SIZE / 4) - 1)
455 	uint64_t t1, t2;
456 	int flit_idx;
457 
458 	MPASS(mask != 0);
459 	MPASS(word <= LAST_WORD);
460 	MPASS(shift < 32);
461 
462 	flit_idx = (LAST_WORD - word) / 2;
463 	if (word & 0x1)
464 		shift += 32;
465 	t1 = be64toh(tcb[flit_idx]) >> shift;
466 	t2 = 0;
467 	if (fls(mask) > 64 - shift) {
468 		/*
469 		 * Will spill over into the next logical flit, which is the flit
470 		 * before this one.  The flit_idx before this one must be valid.
471 		 */
472 		MPASS(flit_idx > 0);
473 		t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift);
474 	}
475 	return ((t2 | t1) & mask);
476 #undef LAST_WORD
477 }
478 #define GET_TCB_FIELD(tcb, F) \
479     get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F)
480 
481 /*
482  * Issues a CPL_GET_TCB to read the entire TCB for the tid.
483  */
484 static int
485 send_get_tcb(struct adapter *sc, u_int tid)
486 {
487 	struct cpl_get_tcb *cpl;
488 	struct wrq_cookie cookie;
489 
490 	MPASS(tid >= sc->tids.tid_base);
491 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
492 
493 	cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16),
494 	    &cookie);
495 	if (__predict_false(cpl == NULL))
496 		return (ENOMEM);
497 	bzero(cpl, sizeof(*cpl));
498 	INIT_TP_WR(cpl, tid);
499 	OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid));
500 	cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) |
501 	    V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id));
502 	cpl->cookie = 0xff;
503 	commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie);
504 
505 	return (0);
506 }
507 
508 static struct tcb_histent *
509 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags)
510 {
511 	struct tcb_histent *te;
512 
513 	MPASS(flags == M_NOWAIT || flags == M_WAITOK);
514 
515 	te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags);
516 	if (te == NULL)
517 		return (NULL);
518 	mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF);
519 	callout_init_mtx(&te->te_callout, &te->te_lock, 0);
520 	te->te_adapter = sc;
521 	te->te_tid = tid;
522 
523 	return (te);
524 }
525 
526 static void
527 free_tcb_histent(struct tcb_histent *te)
528 {
529 
530 	mtx_destroy(&te->te_lock);
531 	free(te, M_CXGBE);
532 }
533 
534 /*
535  * Start tracking the tid in the TCB history.
536  */
537 int
538 add_tid_to_history(struct adapter *sc, u_int tid)
539 {
540 	struct tcb_histent *te = NULL;
541 	struct tom_data *td = sc->tom_softc;
542 	int rc;
543 
544 	MPASS(tid >= sc->tids.tid_base);
545 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
546 
547 	if (td->tcb_history == NULL)
548 		return (ENXIO);
549 
550 	rw_wlock(&td->tcb_history_lock);
551 	if (td->tcb_history[tid] != NULL) {
552 		rc = EEXIST;
553 		goto done;
554 	}
555 	te = alloc_tcb_histent(sc, tid, M_NOWAIT);
556 	if (te == NULL) {
557 		rc = ENOMEM;
558 		goto done;
559 	}
560 	mtx_lock(&te->te_lock);
561 	rc = send_get_tcb(sc, tid);
562 	if (rc == 0) {
563 		te->te_flags |= TE_RPL_PENDING;
564 		td->tcb_history[tid] = te;
565 	} else {
566 		free(te, M_CXGBE);
567 	}
568 	mtx_unlock(&te->te_lock);
569 done:
570 	rw_wunlock(&td->tcb_history_lock);
571 	return (rc);
572 }
573 
574 static void
575 remove_tcb_histent(struct tcb_histent *te)
576 {
577 	struct adapter *sc = te->te_adapter;
578 	struct tom_data *td = sc->tom_softc;
579 
580 	rw_assert(&td->tcb_history_lock, RA_WLOCKED);
581 	mtx_assert(&te->te_lock, MA_OWNED);
582 	MPASS(td->tcb_history[te->te_tid] == te);
583 
584 	td->tcb_history[te->te_tid] = NULL;
585 	free_tcb_histent(te);
586 	rw_wunlock(&td->tcb_history_lock);
587 }
588 
589 static inline struct tcb_histent *
590 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem)
591 {
592 	struct tcb_histent *te;
593 	struct tom_data *td = sc->tom_softc;
594 
595 	MPASS(tid >= sc->tids.tid_base);
596 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
597 
598 	if (td->tcb_history == NULL)
599 		return (NULL);
600 
601 	if (addrem)
602 		rw_wlock(&td->tcb_history_lock);
603 	else
604 		rw_rlock(&td->tcb_history_lock);
605 	te = td->tcb_history[tid];
606 	if (te != NULL) {
607 		mtx_lock(&te->te_lock);
608 		return (te);	/* with both locks held */
609 	}
610 	if (addrem)
611 		rw_wunlock(&td->tcb_history_lock);
612 	else
613 		rw_runlock(&td->tcb_history_lock);
614 
615 	return (te);
616 }
617 
618 static inline void
619 release_tcb_histent(struct tcb_histent *te)
620 {
621 	struct adapter *sc = te->te_adapter;
622 	struct tom_data *td = sc->tom_softc;
623 
624 	mtx_assert(&te->te_lock, MA_OWNED);
625 	mtx_unlock(&te->te_lock);
626 	rw_assert(&td->tcb_history_lock, RA_RLOCKED);
627 	rw_runlock(&td->tcb_history_lock);
628 }
629 
630 static void
631 request_tcb(void *arg)
632 {
633 	struct tcb_histent *te = arg;
634 
635 	mtx_assert(&te->te_lock, MA_OWNED);
636 
637 	/* Noone else is supposed to update the histent. */
638 	MPASS(!(te->te_flags & TE_RPL_PENDING));
639 	if (send_get_tcb(te->te_adapter, te->te_tid) == 0)
640 		te->te_flags |= TE_RPL_PENDING;
641 	else
642 		callout_schedule(&te->te_callout, hz / 100);
643 }
644 
645 static void
646 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb)
647 {
648 	struct tom_data *td = te->te_adapter->tom_softc;
649 	uint64_t tflags = get_tcb_tflags(tcb);
650 	uint8_t sample = 0;
651 
652 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) {
653 		if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0)
654 			sample |= TS_RTO;
655 		if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0)
656 			sample |= TS_DUPACKS;
657 		if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold)
658 			sample |= TS_FASTREXMT;
659 	}
660 
661 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) {
662 		uint32_t snd_wnd;
663 
664 		sample |= TS_SND_BACKLOGGED;	/* for whatever reason. */
665 
666 		snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
667 		if (tflags & V_TF_RECV_SCALE(1))
668 			snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE);
669 		if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd)
670 			sample |= TS_CWND_LIMITED;	/* maybe due to CWND */
671 	}
672 
673 	if (tflags & V_TF_CCTRL_ECN(1)) {
674 
675 		/*
676 		 * CE marker on incoming IP hdr, echoing ECE back in the TCP
677 		 * hdr.  Indicates congestion somewhere on the way from the peer
678 		 * to this node.
679 		 */
680 		if (tflags & V_TF_CCTRL_ECE(1))
681 			sample |= TS_ECN_ECE;
682 
683 		/*
684 		 * ECE seen and CWR sent (or about to be sent).  Might indicate
685 		 * congestion on the way to the peer.  This node is reducing its
686 		 * congestion window in response.
687 		 */
688 		if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1)))
689 			sample |= TS_ECN_CWR;
690 	}
691 
692 	te->te_sample[te->te_pidx] = sample;
693 	if (++te->te_pidx == nitems(te->te_sample))
694 		te->te_pidx = 0;
695 	memcpy(te->te_tcb, tcb, TCB_SIZE);
696 	te->te_flags |= TE_ACTIVE;
697 }
698 
699 static int
700 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
701 {
702 	struct adapter *sc = iq->adapter;
703 	const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *);
704 	const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1);
705 	struct tcb_histent *te;
706 	const u_int tid = GET_TID(cpl);
707 	bool remove;
708 
709 	remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED;
710 	te = lookup_tcb_histent(sc, tid, remove);
711 	if (te == NULL) {
712 		/* Not in the history.  Who issued the GET_TCB for this? */
713 		device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, "
714 		    "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid,
715 		    (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE),
716 		    GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE),
717 		    GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie);
718 		goto done;
719 	}
720 
721 	MPASS(te->te_flags & TE_RPL_PENDING);
722 	te->te_flags &= ~TE_RPL_PENDING;
723 	if (remove) {
724 		remove_tcb_histent(te);
725 	} else {
726 		update_tcb_histent(te, tcb);
727 		callout_reset(&te->te_callout, hz / 10, request_tcb, te);
728 		release_tcb_histent(te);
729 	}
730 done:
731 	m_freem(m);
732 	return (0);
733 }
734 
735 static void
736 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti)
737 {
738 	uint32_t v;
739 
740 	ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE);
741 
742 	v = GET_TCB_FIELD(tcb, T_SRTT);
743 	ti->tcpi_rtt = tcp_ticks_to_us(sc, v);
744 
745 	v = GET_TCB_FIELD(tcb, T_RTTVAR);
746 	ti->tcpi_rttvar = tcp_ticks_to_us(sc, v);
747 
748 	ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH);
749 	ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND);
750 	ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT);
751 	ti->tcpi_rcv_adv = GET_TCB_FIELD(tcb, RCV_ADV);
752 	ti->tcpi_dupacks = GET_TCB_FIELD(tcb, T_DUPACKS);
753 
754 	v = GET_TCB_FIELD(tcb, TX_MAX);
755 	ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW);
756 	ti->tcpi_snd_una = v - GET_TCB_FIELD(tcb, SND_UNA_RAW);
757 	ti->tcpi_snd_max = v - GET_TCB_FIELD(tcb, SND_MAX_RAW);
758 
759 	/* Receive window being advertised by us. */
760 	ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE);	/* Yes, SND. */
761 	ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND);
762 
763 	/* Send window */
764 	ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE);	/* Yes, RCV. */
765 	ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
766 	if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1))
767 		ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale;
768 	else
769 		ti->tcpi_snd_wscale = 0;
770 
771 }
772 
773 static void
774 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te,
775     struct tcp_info *ti)
776 {
777 
778 	fill_tcp_info_from_tcb(sc, te->te_tcb, ti);
779 }
780 
781 /*
782  * Reads the TCB for the given tid using a memory window and copies it to 'buf'
783  * in the same format as CPL_GET_TCB_RPL.
784  */
785 static void
786 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf)
787 {
788 	int i, j, k, rc;
789 	uint32_t addr;
790 	u_char *tcb, tmp;
791 
792 	MPASS(tid >= sc->tids.tid_base);
793 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
794 
795 	addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE;
796 	rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE);
797 	if (rc != 0)
798 		return;
799 
800 	tcb = (u_char *)buf;
801 	for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) {
802 		for (k = 0; k < 16; k++) {
803 			tmp = tcb[i + k];
804 			tcb[i + k] = tcb[j + k];
805 			tcb[j + k] = tmp;
806 		}
807 	}
808 }
809 
810 static void
811 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti)
812 {
813 	uint64_t tcb[TCB_SIZE / sizeof(uint64_t)];
814 	struct tcb_histent *te;
815 
816 	ti->tcpi_toe_tid = tid;
817 	te = lookup_tcb_histent(sc, tid, false);
818 	if (te != NULL) {
819 		fill_tcp_info_from_history(sc, te, ti);
820 		release_tcb_histent(te);
821 	} else {
822 		if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) {
823 			/* XXX: tell firmware to flush TCB cache. */
824 		}
825 		read_tcb_using_memwin(sc, tid, tcb);
826 		fill_tcp_info_from_tcb(sc, tcb, ti);
827 	}
828 }
829 
830 /*
831  * Called by the kernel to allow the TOE driver to "refine" values filled up in
832  * the tcp_info for an offloaded connection.
833  */
834 static void
835 t4_tcp_info(struct toedev *tod, const struct tcpcb *tp, struct tcp_info *ti)
836 {
837 	struct adapter *sc = tod->tod_softc;
838 	struct toepcb *toep = tp->t_toe;
839 
840 	INP_LOCK_ASSERT(tptoinpcb(tp));
841 	MPASS(ti != NULL);
842 
843 	fill_tcp_info(sc, toep->tid, ti);
844 }
845 
846 #ifdef KERN_TLS
847 static int
848 t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp,
849     struct ktls_session *tls, int direction)
850 {
851 	struct toepcb *toep = tp->t_toe;
852 
853 	INP_WLOCK_ASSERT(tptoinpcb(tp));
854 	MPASS(tls != NULL);
855 
856 	return (tls_alloc_ktls(toep, tls, direction));
857 }
858 #endif
859 
860 static void
861 send_mss_flowc_wr(struct adapter *sc, struct toepcb *toep)
862 {
863 	struct wrq_cookie cookie;
864 	struct fw_flowc_wr *flowc;
865 	struct ofld_tx_sdesc *txsd;
866 	const int flowclen = sizeof(*flowc) + sizeof(struct fw_flowc_mnemval);
867 	const int flowclen16 = howmany(flowclen, 16);
868 
869 	if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0) {
870 		CH_ERR(sc, "%s: tid %u out of tx credits (%d, %d).\n", __func__,
871 		    toep->tid, toep->tx_credits, toep->txsd_avail);
872 		return;
873 	}
874 
875 	flowc = start_wrq_wr(&toep->ofld_txq->wrq, flowclen16, &cookie);
876 	if (__predict_false(flowc == NULL)) {
877 		CH_ERR(sc, "ENOMEM in %s for tid %u.\n", __func__, toep->tid);
878 		return;
879 	}
880 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
881 	    V_FW_FLOWC_WR_NPARAMS(1));
882 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
883 	    V_FW_WR_FLOWID(toep->tid));
884 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_MSS;
885 	flowc->mnemval[0].val = htobe32(toep->params.emss);
886 
887 	txsd = &toep->txsd[toep->txsd_pidx];
888 	txsd->tx_credits = flowclen16;
889 	txsd->plen = 0;
890 	toep->tx_credits -= txsd->tx_credits;
891 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
892 		toep->txsd_pidx = 0;
893 	toep->txsd_avail--;
894 	commit_wrq_wr(&toep->ofld_txq->wrq, flowc, &cookie);
895 }
896 
897 static void
898 t4_pmtu_update(struct toedev *tod, struct tcpcb *tp, tcp_seq seq, int mtu)
899 {
900 	struct work_request_hdr *wrh;
901 	struct ulp_txpkt *ulpmc;
902 	int idx, len;
903 	struct wrq_cookie cookie;
904 	struct inpcb *inp = tptoinpcb(tp);
905 	struct toepcb *toep = tp->t_toe;
906 	struct adapter *sc = td_adapter(toep->td);
907 	unsigned short *mtus = &sc->params.mtus[0];
908 
909 	INP_WLOCK_ASSERT(inp);
910 	MPASS(mtu > 0);	/* kernel is supposed to provide something usable. */
911 
912 	/* tp->snd_una and snd_max are in host byte order too. */
913 	seq = be32toh(seq);
914 
915 	CTR6(KTR_CXGBE, "%s: tid %d, seq 0x%08x, mtu %u, mtu_idx %u (%d)",
916 	    __func__, toep->tid, seq, mtu, toep->params.mtu_idx,
917 	    mtus[toep->params.mtu_idx]);
918 
919 	if (ulp_mode(toep) == ULP_MODE_NONE &&	/* XXX: Read TCB otherwise? */
920 	    (SEQ_LT(seq, tp->snd_una) || SEQ_GEQ(seq, tp->snd_max))) {
921 		CTR5(KTR_CXGBE,
922 		    "%s: tid %d, seq 0x%08x not in range [0x%08x, 0x%08x).",
923 		    __func__, toep->tid, seq, tp->snd_una, tp->snd_max);
924 		return;
925 	}
926 
927 	/* Find the best mtu_idx for the suggested MTU. */
928 	for (idx = 0; idx < NMTUS - 1 && mtus[idx + 1] <= mtu; idx++)
929 		continue;
930 	if (idx >= toep->params.mtu_idx)
931 		return;	/* Never increase the PMTU (just like the kernel). */
932 
933 	/*
934 	 * We'll send a compound work request with 2 SET_TCB_FIELDs -- the first
935 	 * one updates the mtu_idx and the second one triggers a retransmit.
936 	 */
937 	len = sizeof(*wrh) + 2 * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
938 	wrh = start_wrq_wr(toep->ctrlq, howmany(len, 16), &cookie);
939 	if (wrh == NULL) {
940 		CH_ERR(sc, "failed to change mtu_idx of tid %d (%u -> %u).\n",
941 		    toep->tid, toep->params.mtu_idx, idx);
942 		return;
943 	}
944 	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
945 	ulpmc = (struct ulp_txpkt *)(wrh + 1);
946 	ulpmc = mk_set_tcb_field_ulp(sc, ulpmc, toep->tid, W_TCB_T_MAXSEG,
947 	    V_TCB_T_MAXSEG(M_TCB_T_MAXSEG), V_TCB_T_MAXSEG(idx));
948 	ulpmc = mk_set_tcb_field_ulp(sc, ulpmc, toep->tid, W_TCB_TIMESTAMP,
949 	    V_TCB_TIMESTAMP(0x7FFFFULL << 11), 0);
950 	commit_wrq_wr(toep->ctrlq, wrh, &cookie);
951 
952 	/* Update the software toepcb and tcpcb. */
953 	toep->params.mtu_idx = idx;
954 	tp->t_maxseg = mtus[toep->params.mtu_idx];
955 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
956 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
957 	else
958 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
959 	toep->params.emss = tp->t_maxseg;
960 	if (tp->t_flags & TF_RCVD_TSTMP)
961 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
962 
963 	/* Update the firmware flowc. */
964 	send_mss_flowc_wr(sc, toep);
965 
966 	/* Update the MTU in the kernel's hostcache. */
967 	if (sc->tt.update_hc_on_pmtu_change != 0) {
968 		struct in_conninfo inc = {0};
969 
970 		inc.inc_fibnum = inp->inp_inc.inc_fibnum;
971 		if (inp->inp_inc.inc_flags & INC_ISIPV6) {
972 			inc.inc_flags |= INC_ISIPV6;
973 			inc.inc6_faddr = inp->inp_inc.inc6_faddr;
974 		} else {
975 			inc.inc_faddr = inp->inp_inc.inc_faddr;
976 		}
977 		tcp_hc_updatemtu(&inc, mtu);
978 	}
979 
980 	CTR6(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u), t_maxseg %u, emss %u",
981 	    __func__, toep->tid, toep->params.mtu_idx,
982 	    mtus[toep->params.mtu_idx], tp->t_maxseg, toep->params.emss);
983 }
984 
985 /*
986  * The TOE driver will not receive any more CPLs for the tid associated with the
987  * toepcb; release the hold on the inpcb.
988  */
989 void
990 final_cpl_received(struct toepcb *toep)
991 {
992 	struct inpcb *inp = toep->inp;
993 	bool need_wakeup;
994 
995 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
996 	INP_WLOCK_ASSERT(inp);
997 	KASSERT(toep->flags & TPF_CPL_PENDING,
998 	    ("%s: CPL not pending already?", __func__));
999 
1000 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
1001 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
1002 
1003 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1004 		release_ddp_resources(toep);
1005 	toep->inp = NULL;
1006 	need_wakeup = (toep->flags & TPF_WAITING_FOR_FINAL) != 0;
1007 	toep->flags &= ~(TPF_CPL_PENDING | TPF_WAITING_FOR_FINAL);
1008 	mbufq_drain(&toep->ulp_pduq);
1009 	mbufq_drain(&toep->ulp_pdu_reclaimq);
1010 	release_offload_resources(toep);
1011 	if (!(toep->flags & TPF_ATTACHED))
1012 		done_with_toepcb(toep);
1013 
1014 	if (!in_pcbrele_wlocked(inp))
1015 		INP_WUNLOCK(inp);
1016 
1017 	if (need_wakeup) {
1018 		struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep);
1019 
1020 		mtx_lock(lock);
1021 		wakeup(toep);
1022 		mtx_unlock(lock);
1023 	}
1024 }
1025 
1026 void
1027 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
1028 {
1029 	struct tid_info *t = &sc->tids;
1030 
1031 	MPASS(tid >= t->tid_base);
1032 	MPASS(tid - t->tid_base < t->ntids);
1033 
1034 	t->tid_tab[tid - t->tid_base] = ctx;
1035 	atomic_add_int(&t->tids_in_use, ntids);
1036 }
1037 
1038 void *
1039 lookup_tid(struct adapter *sc, int tid)
1040 {
1041 	struct tid_info *t = &sc->tids;
1042 
1043 	return (t->tid_tab[tid - t->tid_base]);
1044 }
1045 
1046 void
1047 update_tid(struct adapter *sc, int tid, void *ctx)
1048 {
1049 	struct tid_info *t = &sc->tids;
1050 
1051 	t->tid_tab[tid - t->tid_base] = ctx;
1052 }
1053 
1054 void
1055 remove_tid(struct adapter *sc, int tid, int ntids)
1056 {
1057 	struct tid_info *t = &sc->tids;
1058 
1059 	t->tid_tab[tid - t->tid_base] = NULL;
1060 	atomic_subtract_int(&t->tids_in_use, ntids);
1061 }
1062 
1063 /*
1064  * What mtu_idx to use, given a 4-tuple.  Note that both s->mss and tcp_mssopt
1065  * have the MSS that we should advertise in our SYN.  Advertised MSS doesn't
1066  * account for any TCP options so the effective MSS (only payload, no headers or
1067  * options) could be different.
1068  */
1069 static int
1070 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc,
1071     struct offload_settings *s)
1072 {
1073 	unsigned short *mtus = &sc->params.mtus[0];
1074 	int i, mss, mtu;
1075 
1076 	MPASS(inc != NULL);
1077 
1078 	mss = s->mss > 0 ? s->mss : tcp_mssopt(inc);
1079 	if (inc->inc_flags & INC_ISIPV6)
1080 		mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1081 	else
1082 		mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr);
1083 
1084 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++)
1085 		continue;
1086 
1087 	return (i);
1088 }
1089 
1090 /*
1091  * Determine the receive window size for a socket.
1092  */
1093 u_long
1094 select_rcv_wnd(struct socket *so)
1095 {
1096 	unsigned long wnd;
1097 
1098 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1099 
1100 	wnd = sbspace(&so->so_rcv);
1101 	if (wnd < MIN_RCV_WND)
1102 		wnd = MIN_RCV_WND;
1103 
1104 	return min(wnd, MAX_RCV_WND);
1105 }
1106 
1107 int
1108 select_rcv_wscale(void)
1109 {
1110 	int wscale = 0;
1111 	unsigned long space = sb_max;
1112 
1113 	if (space > MAX_RCV_WND)
1114 		space = MAX_RCV_WND;
1115 
1116 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
1117 		wscale++;
1118 
1119 	return (wscale);
1120 }
1121 
1122 __be64
1123 calc_options0(struct vi_info *vi, struct conn_params *cp)
1124 {
1125 	uint64_t opt0 = 0;
1126 
1127 	opt0 |= F_TCAM_BYPASS;
1128 
1129 	MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE);
1130 	opt0 |= V_WND_SCALE(cp->wscale);
1131 
1132 	MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS);
1133 	opt0 |= V_MSS_IDX(cp->mtu_idx);
1134 
1135 	MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE);
1136 	opt0 |= V_ULP_MODE(cp->ulp_mode);
1137 
1138 	MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ);
1139 	opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize);
1140 
1141 	MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->adapter->vres.l2t.size);
1142 	opt0 |= V_L2T_IDX(cp->l2t_idx);
1143 
1144 	opt0 |= V_SMAC_SEL(vi->smt_idx);
1145 	opt0 |= V_TX_CHAN(vi->pi->tx_chan);
1146 
1147 	MPASS(cp->keepalive == 0 || cp->keepalive == 1);
1148 	opt0 |= V_KEEP_ALIVE(cp->keepalive);
1149 
1150 	MPASS(cp->nagle == 0 || cp->nagle == 1);
1151 	opt0 |= V_NAGLE(cp->nagle);
1152 
1153 	return (htobe64(opt0));
1154 }
1155 
1156 __be32
1157 calc_options2(struct vi_info *vi, struct conn_params *cp)
1158 {
1159 	uint32_t opt2 = 0;
1160 	struct port_info *pi = vi->pi;
1161 	struct adapter *sc = pi->adapter;
1162 
1163 	/*
1164 	 * rx flow control, rx coalesce, congestion control, and tx pace are all
1165 	 * explicitly set by the driver.  On T5+ the ISS is also set by the
1166 	 * driver to the value picked by the kernel.
1167 	 */
1168 	if (is_t4(sc)) {
1169 		opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID;
1170 		opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID;
1171 	} else {
1172 		opt2 |= F_T5_OPT_2_VALID;	/* all 4 valid */
1173 		opt2 |= F_T5_ISS;		/* ISS provided in CPL */
1174 	}
1175 
1176 	MPASS(cp->sack == 0 || cp->sack == 1);
1177 	opt2 |= V_SACK_EN(cp->sack);
1178 
1179 	MPASS(cp->tstamp == 0 || cp->tstamp == 1);
1180 	opt2 |= V_TSTAMPS_EN(cp->tstamp);
1181 
1182 	if (cp->wscale > 0)
1183 		opt2 |= F_WND_SCALE_EN;
1184 
1185 	MPASS(cp->ecn == 0 || cp->ecn == 1);
1186 	opt2 |= V_CCTRL_ECN(cp->ecn);
1187 
1188 	opt2 |= V_TX_QUEUE(TX_MODQ(pi->tx_chan));
1189 	opt2 |= V_PACE(0);
1190 	opt2 |= F_RSS_QUEUE_VALID;
1191 	opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id);
1192 	if (chip_id(sc) <= CHELSIO_T6) {
1193 		MPASS(pi->rx_chan == 0 || pi->rx_chan == 1);
1194 		opt2 |= V_RX_CHANNEL(pi->rx_chan);
1195 	}
1196 
1197 	MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL);
1198 	opt2 |= V_CONG_CNTRL(cp->cong_algo);
1199 
1200 	MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1);
1201 	if (cp->rx_coalesce == 1)
1202 		opt2 |= V_RX_COALESCE(M_RX_COALESCE);
1203 
1204 	opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
1205 	MPASS(cp->ulp_mode != ULP_MODE_TCPDDP);
1206 
1207 	return (htobe32(opt2));
1208 }
1209 
1210 uint64_t
1211 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
1212 {
1213 	struct adapter *sc = vi->adapter;
1214 	struct tp_params *tp = &sc->params.tp;
1215 	uint64_t ntuple = 0;
1216 
1217 	/*
1218 	 * Initialize each of the fields which we care about which are present
1219 	 * in the Compressed Filter Tuple.
1220 	 */
1221 	if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE)
1222 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
1223 
1224 	if (tp->port_shift >= 0)
1225 		ntuple |= (uint64_t)e->lport << tp->port_shift;
1226 
1227 	if (tp->protocol_shift >= 0)
1228 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
1229 
1230 	if (tp->vnic_shift >= 0 && tp->vnic_mode == FW_VNIC_MODE_PF_VF) {
1231 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) |
1232 		    V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) <<
1233 		    tp->vnic_shift;
1234 	}
1235 
1236 	if (is_t4(sc))
1237 		return (htobe32((uint32_t)ntuple));
1238 	else
1239 		return (htobe64(V_FILTER_TUPLE(ntuple)));
1240 }
1241 
1242 /*
1243  * Initialize various connection parameters.
1244  */
1245 void
1246 init_conn_params(struct vi_info *vi , struct offload_settings *s,
1247     struct in_conninfo *inc, struct socket *so,
1248     const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp)
1249 {
1250 	struct port_info *pi = vi->pi;
1251 	struct adapter *sc = pi->adapter;
1252 	struct tom_tunables *tt = &sc->tt;
1253 	struct inpcb *inp = sotoinpcb(so);
1254 	struct tcpcb *tp = intotcpcb(inp);
1255 	u_long wnd;
1256 	u_int q_idx;
1257 
1258 	MPASS(s->offload != 0);
1259 
1260 	/* Congestion control algorithm */
1261 	if (s->cong_algo >= 0)
1262 		cp->cong_algo = s->cong_algo & M_CONG_CNTRL;
1263 	else if (sc->tt.cong_algorithm >= 0)
1264 		cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL;
1265 	else {
1266 		struct cc_algo *cc = CC_ALGO(tp);
1267 
1268 		if (strcasecmp(cc->name, "reno") == 0)
1269 			cp->cong_algo = CONG_ALG_RENO;
1270 		else if (strcasecmp(cc->name, "tahoe") == 0)
1271 			cp->cong_algo = CONG_ALG_TAHOE;
1272 		if (strcasecmp(cc->name, "newreno") == 0)
1273 			cp->cong_algo = CONG_ALG_NEWRENO;
1274 		if (strcasecmp(cc->name, "highspeed") == 0)
1275 			cp->cong_algo = CONG_ALG_HIGHSPEED;
1276 		else {
1277 			/*
1278 			 * Use newreno in case the algorithm selected by the
1279 			 * host stack is not supported by the hardware.
1280 			 */
1281 			cp->cong_algo = CONG_ALG_NEWRENO;
1282 		}
1283 	}
1284 
1285 	/* Tx traffic scheduling class. */
1286 	if (s->sched_class >= 0 && s->sched_class < sc->params.nsched_cls)
1287 		cp->tc_idx = s->sched_class;
1288 	else
1289 		cp->tc_idx = -1;
1290 
1291 	/* Nagle's algorithm. */
1292 	if (s->nagle >= 0)
1293 		cp->nagle = s->nagle > 0 ? 1 : 0;
1294 	else
1295 		cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
1296 
1297 	/* TCP Keepalive. */
1298 	if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE)
1299 		cp->keepalive = 1;
1300 	else
1301 		cp->keepalive = 0;
1302 
1303 	/* Optimization that's specific to T5 @ 40G. */
1304 	if (tt->tx_align >= 0)
1305 		cp->tx_align =  tt->tx_align > 0 ? 1 : 0;
1306 	else if (chip_id(sc) == CHELSIO_T5 &&
1307 	    (port_top_speed(pi) > 10 || sc->params.nports > 2))
1308 		cp->tx_align = 1;
1309 	else
1310 		cp->tx_align = 0;
1311 
1312 	/* ULP mode. */
1313 	cp->ulp_mode = ULP_MODE_NONE;
1314 
1315 	/* Rx coalescing. */
1316 	if (s->rx_coalesce >= 0)
1317 		cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0;
1318 	else if (tt->rx_coalesce >= 0)
1319 		cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0;
1320 	else
1321 		cp->rx_coalesce = 1;	/* default */
1322 
1323 	/*
1324 	 * Index in the PMTU table.  This controls the MSS that we announce in
1325 	 * our SYN initially, but after ESTABLISHED it controls the MSS that we
1326 	 * use to send data.
1327 	 */
1328 	cp->mtu_idx = find_best_mtu_idx(sc, inc, s);
1329 
1330 	/* Tx queue for this connection. */
1331 	if (s->txq == QUEUE_RANDOM)
1332 		q_idx = arc4random();
1333 	else if (s->txq == QUEUE_ROUNDROBIN)
1334 		q_idx = atomic_fetchadd_int(&vi->txq_rr, 1);
1335 	else
1336 		q_idx = s->txq;
1337 	cp->txq_idx = vi->first_ofld_txq + q_idx % vi->nofldtxq;
1338 
1339 	/* Rx queue for this connection. */
1340 	if (s->rxq == QUEUE_RANDOM)
1341 		q_idx = arc4random();
1342 	else if (s->rxq == QUEUE_ROUNDROBIN)
1343 		q_idx = atomic_fetchadd_int(&vi->rxq_rr, 1);
1344 	else
1345 		q_idx = s->rxq;
1346 	cp->rxq_idx = vi->first_ofld_rxq + q_idx % vi->nofldrxq;
1347 
1348 	if (SOLISTENING(so)) {
1349 		/* Passive open */
1350 		MPASS(tcpopt != NULL);
1351 
1352 		/* TCP timestamp option */
1353 		if (tcpopt->tstamp &&
1354 		    (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323)))
1355 			cp->tstamp = 1;
1356 		else
1357 			cp->tstamp = 0;
1358 
1359 		/* SACK */
1360 		if (tcpopt->sack &&
1361 		    (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack)))
1362 			cp->sack = 1;
1363 		else
1364 			cp->sack = 0;
1365 
1366 		/* Receive window scaling. */
1367 		if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323)
1368 			cp->wscale = select_rcv_wscale();
1369 		else
1370 			cp->wscale = 0;
1371 
1372 		/* ECN */
1373 		if (tcpopt->ecn &&	/* XXX: review. */
1374 		    (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn)))
1375 			cp->ecn = 1;
1376 		else
1377 			cp->ecn = 0;
1378 
1379 		wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND);
1380 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1381 
1382 		if (tt->sndbuf > 0)
1383 			cp->sndbuf = tt->sndbuf;
1384 		else if (so->sol_sbsnd_flags & SB_AUTOSIZE &&
1385 		    V_tcp_do_autosndbuf)
1386 			cp->sndbuf = 256 * 1024;
1387 		else
1388 			cp->sndbuf = so->sol_sbsnd_hiwat;
1389 	} else {
1390 		/* Active open */
1391 
1392 		/* TCP timestamp option */
1393 		if (s->tstamp > 0 ||
1394 		    (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP)))
1395 			cp->tstamp = 1;
1396 		else
1397 			cp->tstamp = 0;
1398 
1399 		/* SACK */
1400 		if (s->sack > 0 ||
1401 		    (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT)))
1402 			cp->sack = 1;
1403 		else
1404 			cp->sack = 0;
1405 
1406 		/* Receive window scaling */
1407 		if (tp->t_flags & TF_REQ_SCALE)
1408 			cp->wscale = select_rcv_wscale();
1409 		else
1410 			cp->wscale = 0;
1411 
1412 		/* ECN */
1413 		if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1))
1414 			cp->ecn = 1;
1415 		else
1416 			cp->ecn = 0;
1417 
1418 		SOCKBUF_LOCK(&so->so_rcv);
1419 		wnd = max(select_rcv_wnd(so), MIN_RCV_WND);
1420 		SOCKBUF_UNLOCK(&so->so_rcv);
1421 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1422 
1423 		if (tt->sndbuf > 0)
1424 			cp->sndbuf = tt->sndbuf;
1425 		else {
1426 			SOCKBUF_LOCK(&so->so_snd);
1427 			if (so->so_snd.sb_flags & SB_AUTOSIZE &&
1428 			    V_tcp_do_autosndbuf)
1429 				cp->sndbuf = 256 * 1024;
1430 			else
1431 				cp->sndbuf = so->so_snd.sb_hiwat;
1432 			SOCKBUF_UNLOCK(&so->so_snd);
1433 		}
1434 	}
1435 
1436 	cp->l2t_idx = l2t_idx;
1437 
1438 	/* This will be initialized on ESTABLISHED. */
1439 	cp->emss = 0;
1440 }
1441 
1442 int
1443 negative_advice(int status)
1444 {
1445 
1446 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
1447 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
1448 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
1449 }
1450 
1451 static int
1452 alloc_tid_tab(struct tid_info *t, int flags)
1453 {
1454 
1455 	MPASS(t->ntids > 0);
1456 	MPASS(t->tid_tab == NULL);
1457 
1458 	t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE,
1459 	    M_ZERO | flags);
1460 	if (t->tid_tab == NULL)
1461 		return (ENOMEM);
1462 	atomic_store_rel_int(&t->tids_in_use, 0);
1463 
1464 	return (0);
1465 }
1466 
1467 static void
1468 free_tid_tab(struct tid_info *t)
1469 {
1470 
1471 	KASSERT(t->tids_in_use == 0,
1472 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
1473 
1474 	free(t->tid_tab, M_CXGBE);
1475 	t->tid_tab = NULL;
1476 }
1477 
1478 static int
1479 alloc_stid_tab(struct tid_info *t, int flags)
1480 {
1481 
1482 	MPASS(t->nstids > 0);
1483 	MPASS(t->stid_tab == NULL);
1484 
1485 	t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE,
1486 	    M_ZERO | flags);
1487 	if (t->stid_tab == NULL)
1488 		return (ENOMEM);
1489 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
1490 	t->stids_in_use = 0;
1491 	TAILQ_INIT(&t->stids);
1492 	t->nstids_free_head = t->nstids;
1493 
1494 	return (0);
1495 }
1496 
1497 static void
1498 free_stid_tab(struct tid_info *t)
1499 {
1500 
1501 	KASSERT(t->stids_in_use == 0,
1502 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
1503 
1504 	if (mtx_initialized(&t->stid_lock))
1505 		mtx_destroy(&t->stid_lock);
1506 	free(t->stid_tab, M_CXGBE);
1507 	t->stid_tab = NULL;
1508 }
1509 
1510 static void
1511 free_tid_tabs(struct tid_info *t)
1512 {
1513 
1514 	free_tid_tab(t);
1515 	free_stid_tab(t);
1516 }
1517 
1518 static int
1519 alloc_tid_tabs(struct tid_info *t)
1520 {
1521 	int rc;
1522 
1523 	rc = alloc_tid_tab(t, M_NOWAIT);
1524 	if (rc != 0)
1525 		goto failed;
1526 
1527 	rc = alloc_stid_tab(t, M_NOWAIT);
1528 	if (rc != 0)
1529 		goto failed;
1530 
1531 	return (0);
1532 failed:
1533 	free_tid_tabs(t);
1534 	return (rc);
1535 }
1536 
1537 static inline void
1538 alloc_tcb_history(struct adapter *sc, struct tom_data *td)
1539 {
1540 
1541 	if (sc->tids.ntids == 0 || sc->tids.ntids > 1024)
1542 		return;
1543 	rw_init(&td->tcb_history_lock, "TCB history");
1544 	td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history),
1545 	    M_CXGBE, M_ZERO | M_NOWAIT);
1546 	td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0));
1547 }
1548 
1549 static inline void
1550 free_tcb_history(struct adapter *sc, struct tom_data *td)
1551 {
1552 #ifdef INVARIANTS
1553 	int i;
1554 
1555 	if (td->tcb_history != NULL) {
1556 		for (i = 0; i < sc->tids.ntids; i++) {
1557 			MPASS(td->tcb_history[i] == NULL);
1558 		}
1559 	}
1560 #endif
1561 	free(td->tcb_history, M_CXGBE);
1562 	if (rw_initialized(&td->tcb_history_lock))
1563 		rw_destroy(&td->tcb_history_lock);
1564 }
1565 
1566 static void
1567 free_tom_data(struct adapter *sc, struct tom_data *td)
1568 {
1569 
1570 	ASSERT_SYNCHRONIZED_OP(sc);
1571 
1572 	KASSERT(TAILQ_EMPTY(&td->toep_list),
1573 	    ("%s: TOE PCB list is not empty.", __func__));
1574 	KASSERT(td->lctx_count == 0,
1575 	    ("%s: lctx hash table is not empty.", __func__));
1576 
1577 	t4_free_ppod_region(&td->pr);
1578 
1579 	if (td->listen_mask != 0)
1580 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
1581 
1582 	if (mtx_initialized(&td->unsent_wr_lock))
1583 		mtx_destroy(&td->unsent_wr_lock);
1584 	if (mtx_initialized(&td->lctx_hash_lock))
1585 		mtx_destroy(&td->lctx_hash_lock);
1586 	if (mtx_initialized(&td->toep_list_lock))
1587 		mtx_destroy(&td->toep_list_lock);
1588 
1589 	free_tcb_history(sc, td);
1590 	free_tid_tabs(&sc->tids);
1591 	free(td, M_CXGBE);
1592 }
1593 
1594 static char *
1595 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen,
1596     int *buflen)
1597 {
1598 	char *pkt;
1599 	struct tcphdr *th;
1600 	int ipv6, len;
1601 	const int maxlen =
1602 	    max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) +
1603 	    max(sizeof(struct ip), sizeof(struct ip6_hdr)) +
1604 	    sizeof(struct tcphdr);
1605 
1606 	MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN);
1607 
1608 	pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT);
1609 	if (pkt == NULL)
1610 		return (NULL);
1611 
1612 	ipv6 = inp->inp_vflag & INP_IPV6;
1613 	len = 0;
1614 
1615 	if (EVL_VLANOFTAG(vtag) == 0xfff) {
1616 		struct ether_header *eh = (void *)pkt;
1617 
1618 		if (ipv6)
1619 			eh->ether_type = htons(ETHERTYPE_IPV6);
1620 		else
1621 			eh->ether_type = htons(ETHERTYPE_IP);
1622 
1623 		len += sizeof(*eh);
1624 	} else {
1625 		struct ether_vlan_header *evh = (void *)pkt;
1626 
1627 		evh->evl_encap_proto = htons(ETHERTYPE_VLAN);
1628 		evh->evl_tag = htons(vtag);
1629 		if (ipv6)
1630 			evh->evl_proto = htons(ETHERTYPE_IPV6);
1631 		else
1632 			evh->evl_proto = htons(ETHERTYPE_IP);
1633 
1634 		len += sizeof(*evh);
1635 	}
1636 
1637 	if (ipv6) {
1638 		struct ip6_hdr *ip6 = (void *)&pkt[len];
1639 
1640 		ip6->ip6_vfc = IPV6_VERSION;
1641 		ip6->ip6_plen = htons(sizeof(struct tcphdr));
1642 		ip6->ip6_nxt = IPPROTO_TCP;
1643 		if (open_type == OPEN_TYPE_ACTIVE) {
1644 			ip6->ip6_src = inp->in6p_laddr;
1645 			ip6->ip6_dst = inp->in6p_faddr;
1646 		} else if (open_type == OPEN_TYPE_LISTEN) {
1647 			ip6->ip6_src = inp->in6p_laddr;
1648 			ip6->ip6_dst = ip6->ip6_src;
1649 		}
1650 
1651 		len += sizeof(*ip6);
1652 	} else {
1653 		struct ip *ip = (void *)&pkt[len];
1654 
1655 		ip->ip_v = IPVERSION;
1656 		ip->ip_hl = sizeof(*ip) >> 2;
1657 		ip->ip_tos = inp->inp_ip_tos;
1658 		ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr));
1659 		ip->ip_ttl = inp->inp_ip_ttl;
1660 		ip->ip_p = IPPROTO_TCP;
1661 		if (open_type == OPEN_TYPE_ACTIVE) {
1662 			ip->ip_src = inp->inp_laddr;
1663 			ip->ip_dst = inp->inp_faddr;
1664 		} else if (open_type == OPEN_TYPE_LISTEN) {
1665 			ip->ip_src = inp->inp_laddr;
1666 			ip->ip_dst = ip->ip_src;
1667 		}
1668 
1669 		len += sizeof(*ip);
1670 	}
1671 
1672 	th = (void *)&pkt[len];
1673 	if (open_type == OPEN_TYPE_ACTIVE) {
1674 		th->th_sport = inp->inp_lport;	/* network byte order already */
1675 		th->th_dport = inp->inp_fport;	/* ditto */
1676 	} else if (open_type == OPEN_TYPE_LISTEN) {
1677 		th->th_sport = inp->inp_lport;	/* network byte order already */
1678 		th->th_dport = th->th_sport;
1679 	}
1680 	len += sizeof(th);
1681 
1682 	*pktlen = *buflen = len;
1683 	return (pkt);
1684 }
1685 
1686 const struct offload_settings *
1687 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m,
1688     uint16_t vtag, struct inpcb *inp)
1689 {
1690 	const struct t4_offload_policy *op;
1691 	char *pkt;
1692 	struct offload_rule *r;
1693 	int i, matched, pktlen, buflen;
1694 	static const struct offload_settings allow_offloading_settings = {
1695 		.offload = 1,
1696 		.rx_coalesce = -1,
1697 		.cong_algo = -1,
1698 		.sched_class = -1,
1699 		.tstamp = -1,
1700 		.sack = -1,
1701 		.nagle = -1,
1702 		.ecn = -1,
1703 		.ddp = -1,
1704 		.tls = -1,
1705 		.txq = QUEUE_RANDOM,
1706 		.rxq = QUEUE_RANDOM,
1707 		.mss = -1,
1708 	};
1709 	static const struct offload_settings disallow_offloading_settings = {
1710 		.offload = 0,
1711 		/* rest is irrelevant when offload is off. */
1712 	};
1713 
1714 	rw_assert(&sc->policy_lock, RA_LOCKED);
1715 
1716 	/*
1717 	 * If there's no Connection Offloading Policy attached to the device
1718 	 * then we need to return a default static policy.  If
1719 	 * "cop_managed_offloading" is true, then we need to disallow
1720 	 * offloading until a COP is attached to the device.  Otherwise we
1721 	 * allow offloading ...
1722 	 */
1723 	op = sc->policy;
1724 	if (op == NULL) {
1725 		if (sc->tt.cop_managed_offloading)
1726 			return (&disallow_offloading_settings);
1727 		else
1728 			return (&allow_offloading_settings);
1729 	}
1730 
1731 	switch (open_type) {
1732 	case OPEN_TYPE_ACTIVE:
1733 	case OPEN_TYPE_LISTEN:
1734 		pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen);
1735 		break;
1736 	case OPEN_TYPE_PASSIVE:
1737 		MPASS(m != NULL);
1738 		pkt = mtod(m, char *);
1739 		MPASS(*pkt == CPL_PASS_ACCEPT_REQ);
1740 		pkt += sizeof(struct cpl_pass_accept_req);
1741 		pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req);
1742 		buflen = m->m_len - sizeof(struct cpl_pass_accept_req);
1743 		break;
1744 	default:
1745 		MPASS(0);
1746 		return (&disallow_offloading_settings);
1747 	}
1748 
1749 	if (pkt == NULL || pktlen == 0 || buflen == 0)
1750 		return (&disallow_offloading_settings);
1751 
1752 	matched = 0;
1753 	r = &op->rule[0];
1754 	for (i = 0; i < op->nrules; i++, r++) {
1755 		if (r->open_type != open_type &&
1756 		    r->open_type != OPEN_TYPE_DONTCARE) {
1757 			continue;
1758 		}
1759 		matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen);
1760 		if (matched)
1761 			break;
1762 	}
1763 
1764 	if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN)
1765 		free(pkt, M_CXGBE);
1766 
1767 	return (matched ? &r->settings : &disallow_offloading_settings);
1768 }
1769 
1770 static void
1771 reclaim_wr_resources(void *arg, int count)
1772 {
1773 	struct tom_data *td = arg;
1774 	STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
1775 	struct cpl_act_open_req *cpl;
1776 	u_int opcode, atid, tid;
1777 	struct wrqe *wr;
1778 	struct adapter *sc = td_adapter(td);
1779 
1780 	mtx_lock(&td->unsent_wr_lock);
1781 	STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
1782 	mtx_unlock(&td->unsent_wr_lock);
1783 
1784 	while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
1785 		STAILQ_REMOVE_HEAD(&twr_list, link);
1786 
1787 		cpl = wrtod(wr);
1788 		opcode = GET_OPCODE(cpl);
1789 
1790 		switch (opcode) {
1791 		case CPL_ACT_OPEN_REQ:
1792 		case CPL_ACT_OPEN_REQ6:
1793 			atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1794 			CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1795 			act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1796 			free(wr, M_CXGBE);
1797 			break;
1798 		case CPL_PASS_ACCEPT_RPL:
1799 			tid = GET_TID(cpl);
1800 			CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid);
1801 			synack_failure_cleanup(sc, tid);
1802 			free(wr, M_CXGBE);
1803 			break;
1804 		default:
1805 			log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1806 			    "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1807 			/* WR not freed here; go look at it with a debugger.  */
1808 		}
1809 	}
1810 }
1811 
1812 /*
1813  * Ground control to Major TOM
1814  * Commencing countdown, engines on
1815  */
1816 static int
1817 t4_tom_activate(struct adapter *sc)
1818 {
1819 	struct tom_data *td;
1820 	struct toedev *tod;
1821 	struct vi_info *vi;
1822 	int i, rc, v;
1823 
1824 	ASSERT_SYNCHRONIZED_OP(sc);
1825 
1826 	/* per-adapter softc for TOM */
1827 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1828 	if (td == NULL)
1829 		return (ENOMEM);
1830 
1831 	/* List of TOE PCBs and associated lock */
1832 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1833 	TAILQ_INIT(&td->toep_list);
1834 
1835 	/* Listen context */
1836 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1837 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1838 	    &td->listen_mask, HASH_NOWAIT);
1839 
1840 	/* List of WRs for which L2 resolution failed */
1841 	mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1842 	STAILQ_INIT(&td->unsent_wr_list);
1843 	TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1844 
1845 	/* TID tables */
1846 	rc = alloc_tid_tabs(&sc->tids);
1847 	if (rc != 0)
1848 		goto done;
1849 
1850 	rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1851 	    t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1852 	if (rc != 0)
1853 		goto done;
1854 	t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1855 	    V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1856 
1857 	alloc_tcb_history(sc, td);
1858 
1859 	/* toedev ops */
1860 	tod = &td->tod;
1861 	init_toedev(tod);
1862 	tod->tod_softc = sc;
1863 	tod->tod_connect = t4_connect;
1864 	tod->tod_listen_start = t4_listen_start;
1865 	tod->tod_listen_stop = t4_listen_stop;
1866 	tod->tod_rcvd = t4_rcvd;
1867 	tod->tod_output = t4_tod_output;
1868 	tod->tod_send_rst = t4_send_rst;
1869 	tod->tod_send_fin = t4_send_fin;
1870 	tod->tod_pcb_detach = t4_pcb_detach;
1871 	tod->tod_l2_update = t4_l2_update;
1872 	tod->tod_syncache_added = t4_syncache_added;
1873 	tod->tod_syncache_removed = t4_syncache_removed;
1874 	tod->tod_syncache_respond = t4_syncache_respond;
1875 	tod->tod_offload_socket = t4_offload_socket;
1876 	tod->tod_ctloutput = t4_ctloutput;
1877 	tod->tod_tcp_info = t4_tcp_info;
1878 #ifdef KERN_TLS
1879 	tod->tod_alloc_tls_session = t4_alloc_tls_session;
1880 #endif
1881 	tod->tod_pmtu_update = t4_pmtu_update;
1882 
1883 	for_each_port(sc, i) {
1884 		for_each_vi(sc->port[i], v, vi) {
1885 			SETTOEDEV(vi->ifp, &td->tod);
1886 		}
1887 	}
1888 
1889 	sc->tom_softc = td;
1890 	register_toedev(sc->tom_softc);
1891 
1892 done:
1893 	if (rc != 0)
1894 		free_tom_data(sc, td);
1895 	return (rc);
1896 }
1897 
1898 static int
1899 t4_tom_deactivate(struct adapter *sc)
1900 {
1901 	int rc = 0;
1902 	struct tom_data *td = sc->tom_softc;
1903 
1904 	ASSERT_SYNCHRONIZED_OP(sc);
1905 
1906 	if (td == NULL)
1907 		return (0);	/* XXX. KASSERT? */
1908 
1909 	if (sc->offload_map != 0)
1910 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1911 
1912 	if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1913 		return (EBUSY);	/* both iWARP and iSCSI rely on the TOE. */
1914 
1915 	mtx_lock(&td->toep_list_lock);
1916 	if (!TAILQ_EMPTY(&td->toep_list))
1917 		rc = EBUSY;
1918 	mtx_unlock(&td->toep_list_lock);
1919 
1920 	mtx_lock(&td->lctx_hash_lock);
1921 	if (td->lctx_count > 0)
1922 		rc = EBUSY;
1923 	mtx_unlock(&td->lctx_hash_lock);
1924 
1925 	taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1926 	mtx_lock(&td->unsent_wr_lock);
1927 	if (!STAILQ_EMPTY(&td->unsent_wr_list))
1928 		rc = EBUSY;
1929 	mtx_unlock(&td->unsent_wr_lock);
1930 
1931 	if (rc == 0) {
1932 		unregister_toedev(sc->tom_softc);
1933 		free_tom_data(sc, td);
1934 		sc->tom_softc = NULL;
1935 	}
1936 
1937 	return (rc);
1938 }
1939 
1940 static int
1941 t4_ctloutput_tom(struct socket *so, struct sockopt *sopt)
1942 {
1943 	struct tcpcb *tp = sototcpcb(so);
1944 	struct toepcb *toep = tp->t_toe;
1945 	int error, optval;
1946 
1947 	if (sopt->sopt_level == IPPROTO_TCP && sopt->sopt_name == TCP_USE_DDP) {
1948 		if (sopt->sopt_dir != SOPT_SET)
1949 			return (EOPNOTSUPP);
1950 
1951 		if (sopt->sopt_td != NULL) {
1952 			/* Only settable by the kernel. */
1953 			return (EPERM);
1954 		}
1955 
1956 		error = sooptcopyin(sopt, &optval, sizeof(optval),
1957 		    sizeof(optval));
1958 		if (error != 0)
1959 			return (error);
1960 
1961 		if (optval != 0)
1962 			return (t4_enable_ddp_rcv(so, toep));
1963 		else
1964 			return (EOPNOTSUPP);
1965 	}
1966 	return (tcp_ctloutput(so, sopt));
1967 }
1968 
1969 static int
1970 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1971 {
1972 	struct tcpcb *tp = sototcpcb(so);
1973 	struct toepcb *toep = tp->t_toe;
1974 	int error;
1975 
1976 	/*
1977 	 * No lock is needed as TOE sockets never change between
1978 	 * active and passive.
1979 	 */
1980 	if (SOLISTENING(so))
1981 		return (EINVAL);
1982 
1983 	if (ulp_mode(toep) == ULP_MODE_TCPDDP ||
1984 	    ulp_mode(toep) == ULP_MODE_NONE) {
1985 		error = t4_aio_queue_ddp(so, job);
1986 		if (error != EOPNOTSUPP)
1987 			return (error);
1988 	}
1989 
1990 	return (t4_aio_queue_aiotx(so, job));
1991 }
1992 
1993 static int
1994 t4_tom_mod_load(void)
1995 {
1996 	/* CPL handlers */
1997 	t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl);
1998 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2,
1999 	    CPL_COOKIE_TOM);
2000 	t4_init_connect_cpl_handlers();
2001 	t4_init_listen_cpl_handlers();
2002 	t4_init_cpl_io_handlers();
2003 
2004 	t4_ddp_mod_load();
2005 	t4_tls_mod_load();
2006 
2007 	bcopy(&tcp_protosw, &toe_protosw, sizeof(toe_protosw));
2008 	toe_protosw.pr_ctloutput = t4_ctloutput_tom;
2009 	toe_protosw.pr_aio_queue = t4_aio_queue_tom;
2010 
2011 	bcopy(&tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
2012 	toe6_protosw.pr_ctloutput = t4_ctloutput_tom;
2013 	toe6_protosw.pr_aio_queue = t4_aio_queue_tom;
2014 
2015 	return (t4_register_uld(&tom_uld_info, ULD_TOM));
2016 }
2017 
2018 static void
2019 tom_uninit(struct adapter *sc, void *arg __unused)
2020 {
2021 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
2022 		return;
2023 
2024 	/* Try to free resources (works only if no port has IFCAP_TOE) */
2025 	if (uld_active(sc, ULD_TOM))
2026 		t4_deactivate_uld(sc, ULD_TOM);
2027 
2028 	end_synchronized_op(sc, 0);
2029 }
2030 
2031 static int
2032 t4_tom_mod_unload(void)
2033 {
2034 	t4_iterate(tom_uninit, NULL);
2035 
2036 	if (t4_unregister_uld(&tom_uld_info, ULD_TOM) == EBUSY)
2037 		return (EBUSY);
2038 
2039 	t4_tls_mod_unload();
2040 	t4_ddp_mod_unload();
2041 
2042 	t4_uninit_connect_cpl_handlers();
2043 	t4_uninit_listen_cpl_handlers();
2044 	t4_uninit_cpl_io_handlers();
2045 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM);
2046 	t4_register_cpl_handler(CPL_GET_TCB_RPL, NULL);
2047 
2048 	return (0);
2049 }
2050 #endif	/* TCP_OFFLOAD */
2051 
2052 static int
2053 t4_tom_modevent(module_t mod, int cmd, void *arg)
2054 {
2055 	int rc = 0;
2056 
2057 #ifdef TCP_OFFLOAD
2058 	switch (cmd) {
2059 	case MOD_LOAD:
2060 		rc = t4_tom_mod_load();
2061 		break;
2062 
2063 	case MOD_UNLOAD:
2064 		rc = t4_tom_mod_unload();
2065 		break;
2066 
2067 	default:
2068 		rc = EINVAL;
2069 	}
2070 #else
2071 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
2072 	rc = EOPNOTSUPP;
2073 #endif
2074 	return (rc);
2075 }
2076 
2077 static moduledata_t t4_tom_moddata= {
2078 	"t4_tom",
2079 	t4_tom_modevent,
2080 	0
2081 };
2082 
2083 MODULE_VERSION(t4_tom, 1);
2084 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
2085 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
2086 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
2087