xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision a2464ee12761660f50d0b6f59f233949ebcacc87)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2012 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_kern_tls.h"
36 #include "opt_ratelimit.h"
37 
38 #include <sys/param.h>
39 #include <sys/types.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/ktr.h>
43 #include <sys/lock.h>
44 #include <sys/limits.h>
45 #include <sys/module.h>
46 #include <sys/protosw.h>
47 #include <sys/domain.h>
48 #include <sys/refcount.h>
49 #include <sys/rmlock.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/taskqueue.h>
54 #include <net/if.h>
55 #include <net/if_var.h>
56 #include <net/if_types.h>
57 #include <net/if_vlan_var.h>
58 #include <netinet/in.h>
59 #include <netinet/in_pcb.h>
60 #include <netinet/in_var.h>
61 #include <netinet/ip.h>
62 #include <netinet/ip6.h>
63 #include <netinet6/scope6_var.h>
64 #define TCPSTATES
65 #include <netinet/tcp_fsm.h>
66 #include <netinet/tcp_seq.h>
67 #include <netinet/tcp_timer.h>
68 #include <netinet/tcp_var.h>
69 #include <netinet/toecore.h>
70 #include <netinet/cc/cc.h>
71 
72 #ifdef TCP_OFFLOAD
73 #include "common/common.h"
74 #include "common/t4_msg.h"
75 #include "common/t4_regs.h"
76 #include "common/t4_regs_values.h"
77 #include "common/t4_tcb.h"
78 #include "t4_clip.h"
79 #include "tom/t4_tom_l2t.h"
80 #include "tom/t4_tom.h"
81 #include "tom/t4_tls.h"
82 
83 static struct protosw toe_protosw;
84 static struct protosw toe6_protosw;
85 
86 /* Module ops */
87 static int t4_tom_mod_load(void);
88 static int t4_tom_mod_unload(void);
89 static int t4_tom_modevent(module_t, int, void *);
90 
91 /* ULD ops and helpers */
92 static int t4_tom_activate(struct adapter *);
93 static int t4_tom_deactivate(struct adapter *);
94 
95 static struct uld_info tom_uld_info = {
96 	.uld_id = ULD_TOM,
97 	.activate = t4_tom_activate,
98 	.deactivate = t4_tom_deactivate,
99 };
100 
101 static void release_offload_resources(struct toepcb *);
102 static int alloc_tid_tabs(struct tid_info *);
103 static void free_tid_tabs(struct tid_info *);
104 static void free_tom_data(struct adapter *, struct tom_data *);
105 static void reclaim_wr_resources(void *, int);
106 
107 struct toepcb *
108 alloc_toepcb(struct vi_info *vi, int flags)
109 {
110 	struct port_info *pi = vi->pi;
111 	struct adapter *sc = pi->adapter;
112 	struct toepcb *toep;
113 	int tx_credits, txsd_total, len;
114 
115 	/*
116 	 * The firmware counts tx work request credits in units of 16 bytes
117 	 * each.  Reserve room for an ABORT_REQ so the driver never has to worry
118 	 * about tx credits if it wants to abort a connection.
119 	 */
120 	tx_credits = sc->params.ofldq_wr_cred;
121 	tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
122 
123 	/*
124 	 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
125 	 * immediate payload, and firmware counts tx work request credits in
126 	 * units of 16 byte.  Calculate the maximum work requests possible.
127 	 */
128 	txsd_total = tx_credits /
129 	    howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
130 
131 	len = offsetof(struct toepcb, txsd) +
132 	    txsd_total * sizeof(struct ofld_tx_sdesc);
133 
134 	toep = malloc(len, M_CXGBE, M_ZERO | flags);
135 	if (toep == NULL)
136 		return (NULL);
137 
138 	refcount_init(&toep->refcount, 1);
139 	toep->td = sc->tom_softc;
140 	toep->vi = vi;
141 	toep->tid = -1;
142 	toep->tx_total = tx_credits;
143 	toep->tx_credits = tx_credits;
144 	mbufq_init(&toep->ulp_pduq, INT_MAX);
145 	mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
146 	toep->txsd_total = txsd_total;
147 	toep->txsd_avail = txsd_total;
148 	toep->txsd_pidx = 0;
149 	toep->txsd_cidx = 0;
150 	aiotx_init_toep(toep);
151 
152 	return (toep);
153 }
154 
155 /*
156  * Initialize a toepcb after its params have been filled out.
157  */
158 int
159 init_toepcb(struct vi_info *vi, struct toepcb *toep)
160 {
161 	struct conn_params *cp = &toep->params;
162 	struct port_info *pi = vi->pi;
163 	struct adapter *sc = pi->adapter;
164 	struct tx_cl_rl_params *tc;
165 
166 	if (cp->tc_idx >= 0 && cp->tc_idx < sc->params.nsched_cls) {
167 		tc = &pi->sched_params->cl_rl[cp->tc_idx];
168 		mtx_lock(&sc->tc_lock);
169 		if (tc->state != CS_HW_CONFIGURED) {
170 			CH_ERR(vi, "tid %d cannot be bound to traffic class %d "
171 			    "because it is not configured (its state is %d)\n",
172 			    toep->tid, cp->tc_idx, tc->state);
173 			cp->tc_idx = -1;
174 		} else {
175 			tc->refcount++;
176 		}
177 		mtx_unlock(&sc->tc_lock);
178 	}
179 	toep->ofld_txq = &sc->sge.ofld_txq[cp->txq_idx];
180 	toep->ofld_rxq = &sc->sge.ofld_rxq[cp->rxq_idx];
181 	toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
182 
183 	tls_init_toep(toep);
184 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
185 		ddp_init_toep(toep);
186 
187 	toep->flags |= TPF_INITIALIZED;
188 
189 	return (0);
190 }
191 
192 struct toepcb *
193 hold_toepcb(struct toepcb *toep)
194 {
195 
196 	refcount_acquire(&toep->refcount);
197 	return (toep);
198 }
199 
200 void
201 free_toepcb(struct toepcb *toep)
202 {
203 
204 	if (refcount_release(&toep->refcount) == 0)
205 		return;
206 
207 	KASSERT(!(toep->flags & TPF_ATTACHED),
208 	    ("%s: attached to an inpcb", __func__));
209 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
210 	    ("%s: CPL pending", __func__));
211 
212 	if (toep->flags & TPF_INITIALIZED) {
213 		if (ulp_mode(toep) == ULP_MODE_TCPDDP)
214 			ddp_uninit_toep(toep);
215 		tls_uninit_toep(toep);
216 	}
217 	free(toep, M_CXGBE);
218 }
219 
220 /*
221  * Set up the socket for TCP offload.
222  */
223 void
224 offload_socket(struct socket *so, struct toepcb *toep)
225 {
226 	struct tom_data *td = toep->td;
227 	struct inpcb *inp = sotoinpcb(so);
228 	struct tcpcb *tp = intotcpcb(inp);
229 	struct sockbuf *sb;
230 
231 	INP_WLOCK_ASSERT(inp);
232 
233 	/* Update socket */
234 	sb = &so->so_snd;
235 	SOCKBUF_LOCK(sb);
236 	sb->sb_flags |= SB_NOCOALESCE;
237 	SOCKBUF_UNLOCK(sb);
238 	sb = &so->so_rcv;
239 	SOCKBUF_LOCK(sb);
240 	sb->sb_flags |= SB_NOCOALESCE;
241 	if (inp->inp_vflag & INP_IPV6)
242 		so->so_proto = &toe6_protosw;
243 	else
244 		so->so_proto = &toe_protosw;
245 	SOCKBUF_UNLOCK(sb);
246 
247 	/* Update TCP PCB */
248 	tp->tod = &td->tod;
249 	tp->t_toe = toep;
250 	tp->t_flags |= TF_TOE;
251 
252 	/* Install an extra hold on inp */
253 	toep->inp = inp;
254 	toep->flags |= TPF_ATTACHED;
255 	in_pcbref(inp);
256 
257 	/* Add the TOE PCB to the active list */
258 	mtx_lock(&td->toep_list_lock);
259 	TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
260 	mtx_unlock(&td->toep_list_lock);
261 }
262 
263 void
264 restore_so_proto(struct socket *so, bool v6)
265 {
266 	if (v6)
267 		so->so_proto = &tcp6_protosw;
268 	else
269 		so->so_proto = &tcp_protosw;
270 }
271 
272 /* This is _not_ the normal way to "unoffload" a socket. */
273 void
274 undo_offload_socket(struct socket *so)
275 {
276 	struct inpcb *inp = sotoinpcb(so);
277 	struct tcpcb *tp = intotcpcb(inp);
278 	struct toepcb *toep = tp->t_toe;
279 	struct tom_data *td = toep->td;
280 	struct sockbuf *sb;
281 
282 	INP_WLOCK_ASSERT(inp);
283 
284 	sb = &so->so_snd;
285 	SOCKBUF_LOCK(sb);
286 	sb->sb_flags &= ~SB_NOCOALESCE;
287 	SOCKBUF_UNLOCK(sb);
288 	sb = &so->so_rcv;
289 	SOCKBUF_LOCK(sb);
290 	sb->sb_flags &= ~SB_NOCOALESCE;
291 	restore_so_proto(so, inp->inp_vflag & INP_IPV6);
292 	SOCKBUF_UNLOCK(sb);
293 
294 	tp->tod = NULL;
295 	tp->t_toe = NULL;
296 	tp->t_flags &= ~TF_TOE;
297 
298 	toep->inp = NULL;
299 	toep->flags &= ~TPF_ATTACHED;
300 	if (in_pcbrele_wlocked(inp))
301 		panic("%s: inp freed.", __func__);
302 
303 	mtx_lock(&td->toep_list_lock);
304 	TAILQ_REMOVE(&td->toep_list, toep, link);
305 	mtx_unlock(&td->toep_list_lock);
306 }
307 
308 static void
309 release_offload_resources(struct toepcb *toep)
310 {
311 	struct tom_data *td = toep->td;
312 	struct adapter *sc = td_adapter(td);
313 	int tid = toep->tid;
314 
315 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
316 	    ("%s: %p has CPL pending.", __func__, toep));
317 	KASSERT(!(toep->flags & TPF_ATTACHED),
318 	    ("%s: %p is still attached.", __func__, toep));
319 
320 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
321 	    __func__, toep, tid, toep->l2te, toep->ce);
322 
323 	/*
324 	 * These queues should have been emptied at approximately the same time
325 	 * that a normal connection's socket's so_snd would have been purged or
326 	 * drained.  Do _not_ clean up here.
327 	 */
328 	MPASS(mbufq_len(&toep->ulp_pduq) == 0);
329 	MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0);
330 #ifdef INVARIANTS
331 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
332 		ddp_assert_empty(toep);
333 #endif
334 	MPASS(TAILQ_EMPTY(&toep->aiotx_jobq));
335 
336 	if (toep->l2te)
337 		t4_l2t_release(toep->l2te);
338 
339 	if (tid >= 0) {
340 		remove_tid(sc, tid, toep->ce ? 2 : 1);
341 		release_tid(sc, tid, toep->ctrlq);
342 	}
343 
344 	if (toep->ce)
345 		t4_release_clip_entry(sc, toep->ce);
346 
347 	if (toep->params.tc_idx != -1)
348 		t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx);
349 
350 	mtx_lock(&td->toep_list_lock);
351 	TAILQ_REMOVE(&td->toep_list, toep, link);
352 	mtx_unlock(&td->toep_list_lock);
353 
354 	free_toepcb(toep);
355 }
356 
357 /*
358  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
359  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
360  * pending CPL) then it is time to release all resources tied to the toepcb.
361  *
362  * Also gets called when an offloaded active open fails and the TOM wants the
363  * kernel to take the TCP PCB back.
364  */
365 static void
366 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
367 {
368 #if defined(KTR) || defined(INVARIANTS)
369 	struct inpcb *inp = tp->t_inpcb;
370 #endif
371 	struct toepcb *toep = tp->t_toe;
372 
373 	INP_WLOCK_ASSERT(inp);
374 
375 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
376 	KASSERT(toep->flags & TPF_ATTACHED,
377 	    ("%s: not attached", __func__));
378 
379 #ifdef KTR
380 	if (tp->t_state == TCPS_SYN_SENT) {
381 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
382 		    __func__, toep->tid, toep, toep->flags, inp,
383 		    inp->inp_flags);
384 	} else {
385 		CTR6(KTR_CXGBE,
386 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
387 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
388 		    inp->inp_flags);
389 	}
390 #endif
391 
392 	if (ulp_mode(toep) == ULP_MODE_TLS)
393 		tls_detach(toep);
394 
395 	tp->tod = NULL;
396 	tp->t_toe = NULL;
397 	tp->t_flags &= ~TF_TOE;
398 	toep->flags &= ~TPF_ATTACHED;
399 
400 	if (!(toep->flags & TPF_CPL_PENDING))
401 		release_offload_resources(toep);
402 }
403 
404 /*
405  * setsockopt handler.
406  */
407 static void
408 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
409 {
410 	struct adapter *sc = tod->tod_softc;
411 	struct toepcb *toep = tp->t_toe;
412 
413 	if (dir == SOPT_GET)
414 		return;
415 
416 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
417 
418 	switch (name) {
419 	case TCP_NODELAY:
420 		if (tp->t_state != TCPS_ESTABLISHED)
421 			break;
422 		toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
423 		t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
424 		    V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0);
425 		break;
426 	default:
427 		break;
428 	}
429 }
430 
431 static inline uint64_t
432 get_tcb_tflags(const uint64_t *tcb)
433 {
434 
435 	return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32));
436 }
437 
438 static inline uint32_t
439 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift)
440 {
441 #define LAST_WORD ((TCB_SIZE / 4) - 1)
442 	uint64_t t1, t2;
443 	int flit_idx;
444 
445 	MPASS(mask != 0);
446 	MPASS(word <= LAST_WORD);
447 	MPASS(shift < 32);
448 
449 	flit_idx = (LAST_WORD - word) / 2;
450 	if (word & 0x1)
451 		shift += 32;
452 	t1 = be64toh(tcb[flit_idx]) >> shift;
453 	t2 = 0;
454 	if (fls(mask) > 64 - shift) {
455 		/*
456 		 * Will spill over into the next logical flit, which is the flit
457 		 * before this one.  The flit_idx before this one must be valid.
458 		 */
459 		MPASS(flit_idx > 0);
460 		t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift);
461 	}
462 	return ((t2 | t1) & mask);
463 #undef LAST_WORD
464 }
465 #define GET_TCB_FIELD(tcb, F) \
466     get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F)
467 
468 /*
469  * Issues a CPL_GET_TCB to read the entire TCB for the tid.
470  */
471 static int
472 send_get_tcb(struct adapter *sc, u_int tid)
473 {
474 	struct cpl_get_tcb *cpl;
475 	struct wrq_cookie cookie;
476 
477 	MPASS(tid >= sc->tids.tid_base);
478 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
479 
480 	cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16),
481 	    &cookie);
482 	if (__predict_false(cpl == NULL))
483 		return (ENOMEM);
484 	bzero(cpl, sizeof(*cpl));
485 	INIT_TP_WR(cpl, tid);
486 	OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid));
487 	cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) |
488 	    V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id));
489 	cpl->cookie = 0xff;
490 	commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie);
491 
492 	return (0);
493 }
494 
495 static struct tcb_histent *
496 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags)
497 {
498 	struct tcb_histent *te;
499 
500 	MPASS(flags == M_NOWAIT || flags == M_WAITOK);
501 
502 	te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags);
503 	if (te == NULL)
504 		return (NULL);
505 	mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF);
506 	callout_init_mtx(&te->te_callout, &te->te_lock, 0);
507 	te->te_adapter = sc;
508 	te->te_tid = tid;
509 
510 	return (te);
511 }
512 
513 static void
514 free_tcb_histent(struct tcb_histent *te)
515 {
516 
517 	mtx_destroy(&te->te_lock);
518 	free(te, M_CXGBE);
519 }
520 
521 /*
522  * Start tracking the tid in the TCB history.
523  */
524 int
525 add_tid_to_history(struct adapter *sc, u_int tid)
526 {
527 	struct tcb_histent *te = NULL;
528 	struct tom_data *td = sc->tom_softc;
529 	int rc;
530 
531 	MPASS(tid >= sc->tids.tid_base);
532 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
533 
534 	if (td->tcb_history == NULL)
535 		return (ENXIO);
536 
537 	rw_wlock(&td->tcb_history_lock);
538 	if (td->tcb_history[tid] != NULL) {
539 		rc = EEXIST;
540 		goto done;
541 	}
542 	te = alloc_tcb_histent(sc, tid, M_NOWAIT);
543 	if (te == NULL) {
544 		rc = ENOMEM;
545 		goto done;
546 	}
547 	mtx_lock(&te->te_lock);
548 	rc = send_get_tcb(sc, tid);
549 	if (rc == 0) {
550 		te->te_flags |= TE_RPL_PENDING;
551 		td->tcb_history[tid] = te;
552 	} else {
553 		free(te, M_CXGBE);
554 	}
555 	mtx_unlock(&te->te_lock);
556 done:
557 	rw_wunlock(&td->tcb_history_lock);
558 	return (rc);
559 }
560 
561 static void
562 remove_tcb_histent(struct tcb_histent *te)
563 {
564 	struct adapter *sc = te->te_adapter;
565 	struct tom_data *td = sc->tom_softc;
566 
567 	rw_assert(&td->tcb_history_lock, RA_WLOCKED);
568 	mtx_assert(&te->te_lock, MA_OWNED);
569 	MPASS(td->tcb_history[te->te_tid] == te);
570 
571 	td->tcb_history[te->te_tid] = NULL;
572 	free_tcb_histent(te);
573 	rw_wunlock(&td->tcb_history_lock);
574 }
575 
576 static inline struct tcb_histent *
577 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem)
578 {
579 	struct tcb_histent *te;
580 	struct tom_data *td = sc->tom_softc;
581 
582 	MPASS(tid >= sc->tids.tid_base);
583 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
584 
585 	if (td->tcb_history == NULL)
586 		return (NULL);
587 
588 	if (addrem)
589 		rw_wlock(&td->tcb_history_lock);
590 	else
591 		rw_rlock(&td->tcb_history_lock);
592 	te = td->tcb_history[tid];
593 	if (te != NULL) {
594 		mtx_lock(&te->te_lock);
595 		return (te);	/* with both locks held */
596 	}
597 	if (addrem)
598 		rw_wunlock(&td->tcb_history_lock);
599 	else
600 		rw_runlock(&td->tcb_history_lock);
601 
602 	return (te);
603 }
604 
605 static inline void
606 release_tcb_histent(struct tcb_histent *te)
607 {
608 	struct adapter *sc = te->te_adapter;
609 	struct tom_data *td = sc->tom_softc;
610 
611 	mtx_assert(&te->te_lock, MA_OWNED);
612 	mtx_unlock(&te->te_lock);
613 	rw_assert(&td->tcb_history_lock, RA_RLOCKED);
614 	rw_runlock(&td->tcb_history_lock);
615 }
616 
617 static void
618 request_tcb(void *arg)
619 {
620 	struct tcb_histent *te = arg;
621 
622 	mtx_assert(&te->te_lock, MA_OWNED);
623 
624 	/* Noone else is supposed to update the histent. */
625 	MPASS(!(te->te_flags & TE_RPL_PENDING));
626 	if (send_get_tcb(te->te_adapter, te->te_tid) == 0)
627 		te->te_flags |= TE_RPL_PENDING;
628 	else
629 		callout_schedule(&te->te_callout, hz / 100);
630 }
631 
632 static void
633 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb)
634 {
635 	struct tom_data *td = te->te_adapter->tom_softc;
636 	uint64_t tflags = get_tcb_tflags(tcb);
637 	uint8_t sample = 0;
638 
639 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) {
640 		if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0)
641 			sample |= TS_RTO;
642 		if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0)
643 			sample |= TS_DUPACKS;
644 		if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold)
645 			sample |= TS_FASTREXMT;
646 	}
647 
648 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) {
649 		uint32_t snd_wnd;
650 
651 		sample |= TS_SND_BACKLOGGED;	/* for whatever reason. */
652 
653 		snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
654 		if (tflags & V_TF_RECV_SCALE(1))
655 			snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE);
656 		if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd)
657 			sample |= TS_CWND_LIMITED;	/* maybe due to CWND */
658 	}
659 
660 	if (tflags & V_TF_CCTRL_ECN(1)) {
661 
662 		/*
663 		 * CE marker on incoming IP hdr, echoing ECE back in the TCP
664 		 * hdr.  Indicates congestion somewhere on the way from the peer
665 		 * to this node.
666 		 */
667 		if (tflags & V_TF_CCTRL_ECE(1))
668 			sample |= TS_ECN_ECE;
669 
670 		/*
671 		 * ECE seen and CWR sent (or about to be sent).  Might indicate
672 		 * congestion on the way to the peer.  This node is reducing its
673 		 * congestion window in response.
674 		 */
675 		if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1)))
676 			sample |= TS_ECN_CWR;
677 	}
678 
679 	te->te_sample[te->te_pidx] = sample;
680 	if (++te->te_pidx == nitems(te->te_sample))
681 		te->te_pidx = 0;
682 	memcpy(te->te_tcb, tcb, TCB_SIZE);
683 	te->te_flags |= TE_ACTIVE;
684 }
685 
686 static int
687 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
688 {
689 	struct adapter *sc = iq->adapter;
690 	const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *);
691 	const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1);
692 	struct tcb_histent *te;
693 	const u_int tid = GET_TID(cpl);
694 	bool remove;
695 
696 	remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED;
697 	te = lookup_tcb_histent(sc, tid, remove);
698 	if (te == NULL) {
699 		/* Not in the history.  Who issued the GET_TCB for this? */
700 		device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, "
701 		    "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid,
702 		    (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE),
703 		    GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE),
704 		    GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie);
705 		goto done;
706 	}
707 
708 	MPASS(te->te_flags & TE_RPL_PENDING);
709 	te->te_flags &= ~TE_RPL_PENDING;
710 	if (remove) {
711 		remove_tcb_histent(te);
712 	} else {
713 		update_tcb_histent(te, tcb);
714 		callout_reset(&te->te_callout, hz / 10, request_tcb, te);
715 		release_tcb_histent(te);
716 	}
717 done:
718 	m_freem(m);
719 	return (0);
720 }
721 
722 static void
723 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti)
724 {
725 	uint32_t v;
726 
727 	ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE);
728 
729 	v = GET_TCB_FIELD(tcb, T_SRTT);
730 	ti->tcpi_rtt = tcp_ticks_to_us(sc, v);
731 
732 	v = GET_TCB_FIELD(tcb, T_RTTVAR);
733 	ti->tcpi_rttvar = tcp_ticks_to_us(sc, v);
734 
735 	ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH);
736 	ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND);
737 	ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT);
738 
739 	v = GET_TCB_FIELD(tcb, TX_MAX);
740 	ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW);
741 
742 	/* Receive window being advertised by us. */
743 	ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE);	/* Yes, SND. */
744 	ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND);
745 
746 	/* Send window */
747 	ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE);	/* Yes, RCV. */
748 	ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
749 	if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1))
750 		ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale;
751 	else
752 		ti->tcpi_snd_wscale = 0;
753 
754 }
755 
756 static void
757 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te,
758     struct tcp_info *ti)
759 {
760 
761 	fill_tcp_info_from_tcb(sc, te->te_tcb, ti);
762 }
763 
764 /*
765  * Reads the TCB for the given tid using a memory window and copies it to 'buf'
766  * in the same format as CPL_GET_TCB_RPL.
767  */
768 static void
769 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf)
770 {
771 	int i, j, k, rc;
772 	uint32_t addr;
773 	u_char *tcb, tmp;
774 
775 	MPASS(tid >= sc->tids.tid_base);
776 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
777 
778 	addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE;
779 	rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE);
780 	if (rc != 0)
781 		return;
782 
783 	tcb = (u_char *)buf;
784 	for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) {
785 		for (k = 0; k < 16; k++) {
786 			tmp = tcb[i + k];
787 			tcb[i + k] = tcb[j + k];
788 			tcb[j + k] = tmp;
789 		}
790 	}
791 }
792 
793 static void
794 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti)
795 {
796 	uint64_t tcb[TCB_SIZE / sizeof(uint64_t)];
797 	struct tcb_histent *te;
798 
799 	ti->tcpi_toe_tid = tid;
800 	te = lookup_tcb_histent(sc, tid, false);
801 	if (te != NULL) {
802 		fill_tcp_info_from_history(sc, te, ti);
803 		release_tcb_histent(te);
804 	} else {
805 		if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) {
806 			/* XXX: tell firmware to flush TCB cache. */
807 		}
808 		read_tcb_using_memwin(sc, tid, tcb);
809 		fill_tcp_info_from_tcb(sc, tcb, ti);
810 	}
811 }
812 
813 /*
814  * Called by the kernel to allow the TOE driver to "refine" values filled up in
815  * the tcp_info for an offloaded connection.
816  */
817 static void
818 t4_tcp_info(struct toedev *tod, struct tcpcb *tp, struct tcp_info *ti)
819 {
820 	struct adapter *sc = tod->tod_softc;
821 	struct toepcb *toep = tp->t_toe;
822 
823 	INP_WLOCK_ASSERT(tp->t_inpcb);
824 	MPASS(ti != NULL);
825 
826 	fill_tcp_info(sc, toep->tid, ti);
827 }
828 
829 #ifdef KERN_TLS
830 static int
831 t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp,
832     struct ktls_session *tls, int direction)
833 {
834 	struct toepcb *toep = tp->t_toe;
835 
836 	INP_WLOCK_ASSERT(tp->t_inpcb);
837 	MPASS(tls != NULL);
838 
839 	return (tls_alloc_ktls(toep, tls, direction));
840 }
841 #endif
842 
843 /* SET_TCB_FIELD sent as a ULP command looks like this */
844 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
845     sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
846 
847 static void *
848 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, uint64_t word, uint64_t mask,
849 		uint64_t val, uint32_t tid)
850 {
851 	struct ulptx_idata *ulpsc;
852 	struct cpl_set_tcb_field_core *req;
853 
854 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
855 	ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
856 
857 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
858 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
859 	ulpsc->len = htobe32(sizeof(*req));
860 
861 	req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
862 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, tid));
863 	req->reply_ctrl = htobe16(V_NO_REPLY(1));
864 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
865 	req->mask = htobe64(mask);
866 	req->val = htobe64(val);
867 
868 	ulpsc = (struct ulptx_idata *)(req + 1);
869 	if (LEN__SET_TCB_FIELD_ULP % 16) {
870 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
871 		ulpsc->len = htobe32(0);
872 		return (ulpsc + 1);
873 	}
874 	return (ulpsc);
875 }
876 
877 static void
878 send_mss_flowc_wr(struct adapter *sc, struct toepcb *toep)
879 {
880 	struct wrq_cookie cookie;
881 	struct fw_flowc_wr *flowc;
882 	struct ofld_tx_sdesc *txsd;
883 	const int flowclen = sizeof(*flowc) + sizeof(struct fw_flowc_mnemval);
884 	const int flowclen16 = howmany(flowclen, 16);
885 
886 	if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0) {
887 		CH_ERR(sc, "%s: tid %u out of tx credits (%d, %d).\n", __func__,
888 		    toep->tid, toep->tx_credits, toep->txsd_avail);
889 		return;
890 	}
891 
892 	flowc = start_wrq_wr(&toep->ofld_txq->wrq, flowclen16, &cookie);
893 	if (__predict_false(flowc == NULL)) {
894 		CH_ERR(sc, "ENOMEM in %s for tid %u.\n", __func__, toep->tid);
895 		return;
896 	}
897 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
898 	    V_FW_FLOWC_WR_NPARAMS(1));
899 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
900 	    V_FW_WR_FLOWID(toep->tid));
901 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_MSS;
902 	flowc->mnemval[0].val = htobe32(toep->params.emss);
903 
904 	txsd = &toep->txsd[toep->txsd_pidx];
905 	txsd->tx_credits = flowclen16;
906 	txsd->plen = 0;
907 	toep->tx_credits -= txsd->tx_credits;
908 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
909 		toep->txsd_pidx = 0;
910 	toep->txsd_avail--;
911 	commit_wrq_wr(&toep->ofld_txq->wrq, flowc, &cookie);
912 }
913 
914 static void
915 t4_pmtu_update(struct toedev *tod, struct tcpcb *tp, tcp_seq seq, int mtu)
916 {
917 	struct work_request_hdr *wrh;
918 	struct ulp_txpkt *ulpmc;
919 	int idx, len;
920 	struct wrq_cookie cookie;
921 	struct inpcb *inp = tp->t_inpcb;
922 	struct toepcb *toep = tp->t_toe;
923 	struct adapter *sc = td_adapter(toep->td);
924 	unsigned short *mtus = &sc->params.mtus[0];
925 
926 	INP_WLOCK_ASSERT(inp);
927 	MPASS(mtu > 0);	/* kernel is supposed to provide something usable. */
928 
929 	/* tp->snd_una and snd_max are in host byte order too. */
930 	seq = be32toh(seq);
931 
932 	CTR6(KTR_CXGBE, "%s: tid %d, seq 0x%08x, mtu %u, mtu_idx %u (%d)",
933 	    __func__, toep->tid, seq, mtu, toep->params.mtu_idx,
934 	    mtus[toep->params.mtu_idx]);
935 
936 	if (ulp_mode(toep) == ULP_MODE_NONE &&	/* XXX: Read TCB otherwise? */
937 	    (SEQ_LT(seq, tp->snd_una) || SEQ_GEQ(seq, tp->snd_max))) {
938 		CTR5(KTR_CXGBE,
939 		    "%s: tid %d, seq 0x%08x not in range [0x%08x, 0x%08x).",
940 		    __func__, toep->tid, seq, tp->snd_una, tp->snd_max);
941 		return;
942 	}
943 
944 	/* Find the best mtu_idx for the suggested MTU. */
945 	for (idx = 0; idx < NMTUS - 1 && mtus[idx + 1] <= mtu; idx++)
946 		continue;
947 	if (idx >= toep->params.mtu_idx)
948 		return;	/* Never increase the PMTU (just like the kernel). */
949 
950 	/*
951 	 * We'll send a compound work request with 2 SET_TCB_FIELDs -- the first
952 	 * one updates the mtu_idx and the second one triggers a retransmit.
953 	 */
954 	len = sizeof(*wrh) + 2 * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
955 	wrh = start_wrq_wr(toep->ctrlq, howmany(len, 16), &cookie);
956 	if (wrh == NULL) {
957 		CH_ERR(sc, "failed to change mtu_idx of tid %d (%u -> %u).\n",
958 		    toep->tid, toep->params.mtu_idx, idx);
959 		return;
960 	}
961 	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
962 	ulpmc = (struct ulp_txpkt *)(wrh + 1);
963 	ulpmc = mk_set_tcb_field_ulp(ulpmc, W_TCB_T_MAXSEG,
964 	    V_TCB_T_MAXSEG(M_TCB_T_MAXSEG), V_TCB_T_MAXSEG(idx), toep->tid);
965 	ulpmc = mk_set_tcb_field_ulp(ulpmc, W_TCB_TIMESTAMP,
966 	    V_TCB_TIMESTAMP(0x7FFFFULL << 11), 0, toep->tid);
967 	commit_wrq_wr(toep->ctrlq, wrh, &cookie);
968 
969 	/* Update the software toepcb and tcpcb. */
970 	toep->params.mtu_idx = idx;
971 	tp->t_maxseg = mtus[toep->params.mtu_idx];
972 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
973 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
974 	else
975 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
976 	toep->params.emss = tp->t_maxseg;
977 	if (tp->t_flags & TF_RCVD_TSTMP)
978 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
979 
980 	/* Update the firmware flowc. */
981 	send_mss_flowc_wr(sc, toep);
982 
983 	/* Update the MTU in the kernel's hostcache. */
984 	if (sc->tt.update_hc_on_pmtu_change != 0) {
985 		struct in_conninfo inc = {0};
986 
987 		inc.inc_fibnum = inp->inp_inc.inc_fibnum;
988 		if (inp->inp_inc.inc_flags & INC_ISIPV6) {
989 			inc.inc_flags |= INC_ISIPV6;
990 			inc.inc6_faddr = inp->inp_inc.inc6_faddr;
991 		} else {
992 			inc.inc_faddr = inp->inp_inc.inc_faddr;
993 		}
994 		tcp_hc_updatemtu(&inc, mtu);
995 	}
996 
997 	CTR6(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u), t_maxseg %u, emss %u",
998 	    __func__, toep->tid, toep->params.mtu_idx,
999 	    mtus[toep->params.mtu_idx], tp->t_maxseg, toep->params.emss);
1000 }
1001 
1002 /*
1003  * The TOE driver will not receive any more CPLs for the tid associated with the
1004  * toepcb; release the hold on the inpcb.
1005  */
1006 void
1007 final_cpl_received(struct toepcb *toep)
1008 {
1009 	struct inpcb *inp = toep->inp;
1010 	bool need_wakeup;
1011 
1012 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
1013 	INP_WLOCK_ASSERT(inp);
1014 	KASSERT(toep->flags & TPF_CPL_PENDING,
1015 	    ("%s: CPL not pending already?", __func__));
1016 
1017 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
1018 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
1019 
1020 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1021 		release_ddp_resources(toep);
1022 	else if (ulp_mode(toep) == ULP_MODE_TLS)
1023 		tls_detach(toep);
1024 	toep->inp = NULL;
1025 	need_wakeup = (toep->flags & TPF_WAITING_FOR_FINAL) != 0;
1026 	toep->flags &= ~(TPF_CPL_PENDING | TPF_WAITING_FOR_FINAL);
1027 	mbufq_drain(&toep->ulp_pduq);
1028 	mbufq_drain(&toep->ulp_pdu_reclaimq);
1029 
1030 	if (!(toep->flags & TPF_ATTACHED))
1031 		release_offload_resources(toep);
1032 
1033 	if (!in_pcbrele_wlocked(inp))
1034 		INP_WUNLOCK(inp);
1035 
1036 	if (need_wakeup) {
1037 		struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep);
1038 
1039 		mtx_lock(lock);
1040 		wakeup(toep);
1041 		mtx_unlock(lock);
1042 	}
1043 }
1044 
1045 void
1046 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
1047 {
1048 	struct tid_info *t = &sc->tids;
1049 
1050 	MPASS(tid >= t->tid_base);
1051 	MPASS(tid - t->tid_base < t->ntids);
1052 
1053 	t->tid_tab[tid - t->tid_base] = ctx;
1054 	atomic_add_int(&t->tids_in_use, ntids);
1055 }
1056 
1057 void *
1058 lookup_tid(struct adapter *sc, int tid)
1059 {
1060 	struct tid_info *t = &sc->tids;
1061 
1062 	return (t->tid_tab[tid - t->tid_base]);
1063 }
1064 
1065 void
1066 update_tid(struct adapter *sc, int tid, void *ctx)
1067 {
1068 	struct tid_info *t = &sc->tids;
1069 
1070 	t->tid_tab[tid - t->tid_base] = ctx;
1071 }
1072 
1073 void
1074 remove_tid(struct adapter *sc, int tid, int ntids)
1075 {
1076 	struct tid_info *t = &sc->tids;
1077 
1078 	t->tid_tab[tid - t->tid_base] = NULL;
1079 	atomic_subtract_int(&t->tids_in_use, ntids);
1080 }
1081 
1082 /*
1083  * What mtu_idx to use, given a 4-tuple.  Note that both s->mss and tcp_mssopt
1084  * have the MSS that we should advertise in our SYN.  Advertised MSS doesn't
1085  * account for any TCP options so the effective MSS (only payload, no headers or
1086  * options) could be different.
1087  */
1088 static int
1089 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc,
1090     struct offload_settings *s)
1091 {
1092 	unsigned short *mtus = &sc->params.mtus[0];
1093 	int i, mss, mtu;
1094 
1095 	MPASS(inc != NULL);
1096 
1097 	mss = s->mss > 0 ? s->mss : tcp_mssopt(inc);
1098 	if (inc->inc_flags & INC_ISIPV6)
1099 		mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1100 	else
1101 		mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr);
1102 
1103 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++)
1104 		continue;
1105 
1106 	return (i);
1107 }
1108 
1109 /*
1110  * Determine the receive window size for a socket.
1111  */
1112 u_long
1113 select_rcv_wnd(struct socket *so)
1114 {
1115 	unsigned long wnd;
1116 
1117 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1118 
1119 	wnd = sbspace(&so->so_rcv);
1120 	if (wnd < MIN_RCV_WND)
1121 		wnd = MIN_RCV_WND;
1122 
1123 	return min(wnd, MAX_RCV_WND);
1124 }
1125 
1126 int
1127 select_rcv_wscale(void)
1128 {
1129 	int wscale = 0;
1130 	unsigned long space = sb_max;
1131 
1132 	if (space > MAX_RCV_WND)
1133 		space = MAX_RCV_WND;
1134 
1135 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
1136 		wscale++;
1137 
1138 	return (wscale);
1139 }
1140 
1141 __be64
1142 calc_options0(struct vi_info *vi, struct conn_params *cp)
1143 {
1144 	uint64_t opt0 = 0;
1145 
1146 	opt0 |= F_TCAM_BYPASS;
1147 
1148 	MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE);
1149 	opt0 |= V_WND_SCALE(cp->wscale);
1150 
1151 	MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS);
1152 	opt0 |= V_MSS_IDX(cp->mtu_idx);
1153 
1154 	MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE);
1155 	opt0 |= V_ULP_MODE(cp->ulp_mode);
1156 
1157 	MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ);
1158 	opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize);
1159 
1160 	MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->adapter->vres.l2t.size);
1161 	opt0 |= V_L2T_IDX(cp->l2t_idx);
1162 
1163 	opt0 |= V_SMAC_SEL(vi->smt_idx);
1164 	opt0 |= V_TX_CHAN(vi->pi->tx_chan);
1165 
1166 	MPASS(cp->keepalive == 0 || cp->keepalive == 1);
1167 	opt0 |= V_KEEP_ALIVE(cp->keepalive);
1168 
1169 	MPASS(cp->nagle == 0 || cp->nagle == 1);
1170 	opt0 |= V_NAGLE(cp->nagle);
1171 
1172 	return (htobe64(opt0));
1173 }
1174 
1175 __be32
1176 calc_options2(struct vi_info *vi, struct conn_params *cp)
1177 {
1178 	uint32_t opt2 = 0;
1179 	struct port_info *pi = vi->pi;
1180 	struct adapter *sc = pi->adapter;
1181 
1182 	/*
1183 	 * rx flow control, rx coalesce, congestion control, and tx pace are all
1184 	 * explicitly set by the driver.  On T5+ the ISS is also set by the
1185 	 * driver to the value picked by the kernel.
1186 	 */
1187 	if (is_t4(sc)) {
1188 		opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID;
1189 		opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID;
1190 	} else {
1191 		opt2 |= F_T5_OPT_2_VALID;	/* all 4 valid */
1192 		opt2 |= F_T5_ISS;		/* ISS provided in CPL */
1193 	}
1194 
1195 	MPASS(cp->sack == 0 || cp->sack == 1);
1196 	opt2 |= V_SACK_EN(cp->sack);
1197 
1198 	MPASS(cp->tstamp == 0 || cp->tstamp == 1);
1199 	opt2 |= V_TSTAMPS_EN(cp->tstamp);
1200 
1201 	if (cp->wscale > 0)
1202 		opt2 |= F_WND_SCALE_EN;
1203 
1204 	MPASS(cp->ecn == 0 || cp->ecn == 1);
1205 	opt2 |= V_CCTRL_ECN(cp->ecn);
1206 
1207 	/* XXX: F_RX_CHANNEL for multiple rx c-chan support goes here. */
1208 
1209 	opt2 |= V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]);
1210 	opt2 |= V_PACE(0);
1211 	opt2 |= F_RSS_QUEUE_VALID;
1212 	opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id);
1213 
1214 	MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL);
1215 	opt2 |= V_CONG_CNTRL(cp->cong_algo);
1216 
1217 	MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1);
1218 	if (cp->rx_coalesce == 1)
1219 		opt2 |= V_RX_COALESCE(M_RX_COALESCE);
1220 
1221 	opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
1222 #ifdef USE_DDP_RX_FLOW_CONTROL
1223 	if (cp->ulp_mode == ULP_MODE_TCPDDP)
1224 		opt2 |= F_RX_FC_DDP;
1225 #endif
1226 
1227 	return (htobe32(opt2));
1228 }
1229 
1230 uint64_t
1231 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
1232 {
1233 	struct adapter *sc = vi->adapter;
1234 	struct tp_params *tp = &sc->params.tp;
1235 	uint64_t ntuple = 0;
1236 
1237 	/*
1238 	 * Initialize each of the fields which we care about which are present
1239 	 * in the Compressed Filter Tuple.
1240 	 */
1241 	if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE)
1242 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
1243 
1244 	if (tp->port_shift >= 0)
1245 		ntuple |= (uint64_t)e->lport << tp->port_shift;
1246 
1247 	if (tp->protocol_shift >= 0)
1248 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
1249 
1250 	if (tp->vnic_shift >= 0 && tp->vnic_mode == FW_VNIC_MODE_PF_VF) {
1251 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) |
1252 		    V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) <<
1253 		    tp->vnic_shift;
1254 	}
1255 
1256 	if (is_t4(sc))
1257 		return (htobe32((uint32_t)ntuple));
1258 	else
1259 		return (htobe64(V_FILTER_TUPLE(ntuple)));
1260 }
1261 
1262 static int
1263 is_tls_sock(struct socket *so, struct adapter *sc)
1264 {
1265 	struct inpcb *inp = sotoinpcb(so);
1266 	int i, rc;
1267 
1268 	/* XXX: Eventually add a SO_WANT_TLS socket option perhaps? */
1269 	rc = 0;
1270 	ADAPTER_LOCK(sc);
1271 	for (i = 0; i < sc->tt.num_tls_rx_ports; i++) {
1272 		if (inp->inp_lport == htons(sc->tt.tls_rx_ports[i]) ||
1273 		    inp->inp_fport == htons(sc->tt.tls_rx_ports[i])) {
1274 			rc = 1;
1275 			break;
1276 		}
1277 	}
1278 	ADAPTER_UNLOCK(sc);
1279 	return (rc);
1280 }
1281 
1282 /*
1283  * Initialize various connection parameters.
1284  */
1285 void
1286 init_conn_params(struct vi_info *vi , struct offload_settings *s,
1287     struct in_conninfo *inc, struct socket *so,
1288     const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp)
1289 {
1290 	struct port_info *pi = vi->pi;
1291 	struct adapter *sc = pi->adapter;
1292 	struct tom_tunables *tt = &sc->tt;
1293 	struct inpcb *inp = sotoinpcb(so);
1294 	struct tcpcb *tp = intotcpcb(inp);
1295 	u_long wnd;
1296 	u_int q_idx;
1297 
1298 	MPASS(s->offload != 0);
1299 
1300 	/* Congestion control algorithm */
1301 	if (s->cong_algo >= 0)
1302 		cp->cong_algo = s->cong_algo & M_CONG_CNTRL;
1303 	else if (sc->tt.cong_algorithm >= 0)
1304 		cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL;
1305 	else {
1306 		struct cc_algo *cc = CC_ALGO(tp);
1307 
1308 		if (strcasecmp(cc->name, "reno") == 0)
1309 			cp->cong_algo = CONG_ALG_RENO;
1310 		else if (strcasecmp(cc->name, "tahoe") == 0)
1311 			cp->cong_algo = CONG_ALG_TAHOE;
1312 		if (strcasecmp(cc->name, "newreno") == 0)
1313 			cp->cong_algo = CONG_ALG_NEWRENO;
1314 		if (strcasecmp(cc->name, "highspeed") == 0)
1315 			cp->cong_algo = CONG_ALG_HIGHSPEED;
1316 		else {
1317 			/*
1318 			 * Use newreno in case the algorithm selected by the
1319 			 * host stack is not supported by the hardware.
1320 			 */
1321 			cp->cong_algo = CONG_ALG_NEWRENO;
1322 		}
1323 	}
1324 
1325 	/* Tx traffic scheduling class. */
1326 	if (s->sched_class >= 0 && s->sched_class < sc->params.nsched_cls)
1327 		cp->tc_idx = s->sched_class;
1328 	else
1329 		cp->tc_idx = -1;
1330 
1331 	/* Nagle's algorithm. */
1332 	if (s->nagle >= 0)
1333 		cp->nagle = s->nagle > 0 ? 1 : 0;
1334 	else
1335 		cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
1336 
1337 	/* TCP Keepalive. */
1338 	if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE)
1339 		cp->keepalive = 1;
1340 	else
1341 		cp->keepalive = 0;
1342 
1343 	/* Optimization that's specific to T5 @ 40G. */
1344 	if (tt->tx_align >= 0)
1345 		cp->tx_align =  tt->tx_align > 0 ? 1 : 0;
1346 	else if (chip_id(sc) == CHELSIO_T5 &&
1347 	    (port_top_speed(pi) > 10 || sc->params.nports > 2))
1348 		cp->tx_align = 1;
1349 	else
1350 		cp->tx_align = 0;
1351 
1352 	/* ULP mode. */
1353 	if (can_tls_offload(sc) &&
1354 	    (s->tls > 0 || (s->tls < 0 && is_tls_sock(so, sc))))
1355 		cp->ulp_mode = ULP_MODE_TLS;
1356 	else if (s->ddp > 0 ||
1357 	    (s->ddp < 0 && sc->tt.ddp && (so_options_get(so) & SO_NO_DDP) == 0))
1358 		cp->ulp_mode = ULP_MODE_TCPDDP;
1359 	else
1360 		cp->ulp_mode = ULP_MODE_NONE;
1361 
1362 	/* Rx coalescing. */
1363 	if (s->rx_coalesce >= 0)
1364 		cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0;
1365 	else if (cp->ulp_mode == ULP_MODE_TLS)
1366 		cp->rx_coalesce = 0;
1367 	else if (tt->rx_coalesce >= 0)
1368 		cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0;
1369 	else
1370 		cp->rx_coalesce = 1;	/* default */
1371 
1372 	/*
1373 	 * Index in the PMTU table.  This controls the MSS that we announce in
1374 	 * our SYN initially, but after ESTABLISHED it controls the MSS that we
1375 	 * use to send data.
1376 	 */
1377 	cp->mtu_idx = find_best_mtu_idx(sc, inc, s);
1378 
1379 	/* Tx queue for this connection. */
1380 	if (s->txq == QUEUE_RANDOM)
1381 		q_idx = arc4random();
1382 	else if (s->txq == QUEUE_ROUNDROBIN)
1383 		q_idx = atomic_fetchadd_int(&vi->txq_rr, 1);
1384 	else
1385 		q_idx = s->txq;
1386 	cp->txq_idx = vi->first_ofld_txq + q_idx % vi->nofldtxq;
1387 
1388 	/* Rx queue for this connection. */
1389 	if (s->rxq == QUEUE_RANDOM)
1390 		q_idx = arc4random();
1391 	else if (s->rxq == QUEUE_ROUNDROBIN)
1392 		q_idx = atomic_fetchadd_int(&vi->rxq_rr, 1);
1393 	else
1394 		q_idx = s->rxq;
1395 	cp->rxq_idx = vi->first_ofld_rxq + q_idx % vi->nofldrxq;
1396 
1397 	if (SOLISTENING(so)) {
1398 		/* Passive open */
1399 		MPASS(tcpopt != NULL);
1400 
1401 		/* TCP timestamp option */
1402 		if (tcpopt->tstamp &&
1403 		    (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323)))
1404 			cp->tstamp = 1;
1405 		else
1406 			cp->tstamp = 0;
1407 
1408 		/* SACK */
1409 		if (tcpopt->sack &&
1410 		    (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack)))
1411 			cp->sack = 1;
1412 		else
1413 			cp->sack = 0;
1414 
1415 		/* Receive window scaling. */
1416 		if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323)
1417 			cp->wscale = select_rcv_wscale();
1418 		else
1419 			cp->wscale = 0;
1420 
1421 		/* ECN */
1422 		if (tcpopt->ecn &&	/* XXX: review. */
1423 		    (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn)))
1424 			cp->ecn = 1;
1425 		else
1426 			cp->ecn = 0;
1427 
1428 		wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND);
1429 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1430 
1431 		if (tt->sndbuf > 0)
1432 			cp->sndbuf = tt->sndbuf;
1433 		else if (so->sol_sbsnd_flags & SB_AUTOSIZE &&
1434 		    V_tcp_do_autosndbuf)
1435 			cp->sndbuf = 256 * 1024;
1436 		else
1437 			cp->sndbuf = so->sol_sbsnd_hiwat;
1438 	} else {
1439 		/* Active open */
1440 
1441 		/* TCP timestamp option */
1442 		if (s->tstamp > 0 ||
1443 		    (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP)))
1444 			cp->tstamp = 1;
1445 		else
1446 			cp->tstamp = 0;
1447 
1448 		/* SACK */
1449 		if (s->sack > 0 ||
1450 		    (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT)))
1451 			cp->sack = 1;
1452 		else
1453 			cp->sack = 0;
1454 
1455 		/* Receive window scaling */
1456 		if (tp->t_flags & TF_REQ_SCALE)
1457 			cp->wscale = select_rcv_wscale();
1458 		else
1459 			cp->wscale = 0;
1460 
1461 		/* ECN */
1462 		if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1))
1463 			cp->ecn = 1;
1464 		else
1465 			cp->ecn = 0;
1466 
1467 		SOCKBUF_LOCK(&so->so_rcv);
1468 		wnd = max(select_rcv_wnd(so), MIN_RCV_WND);
1469 		SOCKBUF_UNLOCK(&so->so_rcv);
1470 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1471 
1472 		if (tt->sndbuf > 0)
1473 			cp->sndbuf = tt->sndbuf;
1474 		else {
1475 			SOCKBUF_LOCK(&so->so_snd);
1476 			if (so->so_snd.sb_flags & SB_AUTOSIZE &&
1477 			    V_tcp_do_autosndbuf)
1478 				cp->sndbuf = 256 * 1024;
1479 			else
1480 				cp->sndbuf = so->so_snd.sb_hiwat;
1481 			SOCKBUF_UNLOCK(&so->so_snd);
1482 		}
1483 	}
1484 
1485 	cp->l2t_idx = l2t_idx;
1486 
1487 	/* This will be initialized on ESTABLISHED. */
1488 	cp->emss = 0;
1489 }
1490 
1491 int
1492 negative_advice(int status)
1493 {
1494 
1495 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
1496 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
1497 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
1498 }
1499 
1500 static int
1501 alloc_tid_tab(struct tid_info *t, int flags)
1502 {
1503 
1504 	MPASS(t->ntids > 0);
1505 	MPASS(t->tid_tab == NULL);
1506 
1507 	t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE,
1508 	    M_ZERO | flags);
1509 	if (t->tid_tab == NULL)
1510 		return (ENOMEM);
1511 	atomic_store_rel_int(&t->tids_in_use, 0);
1512 
1513 	return (0);
1514 }
1515 
1516 static void
1517 free_tid_tab(struct tid_info *t)
1518 {
1519 
1520 	KASSERT(t->tids_in_use == 0,
1521 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
1522 
1523 	free(t->tid_tab, M_CXGBE);
1524 	t->tid_tab = NULL;
1525 }
1526 
1527 static int
1528 alloc_stid_tab(struct tid_info *t, int flags)
1529 {
1530 
1531 	MPASS(t->nstids > 0);
1532 	MPASS(t->stid_tab == NULL);
1533 
1534 	t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE,
1535 	    M_ZERO | flags);
1536 	if (t->stid_tab == NULL)
1537 		return (ENOMEM);
1538 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
1539 	t->stids_in_use = 0;
1540 	TAILQ_INIT(&t->stids);
1541 	t->nstids_free_head = t->nstids;
1542 
1543 	return (0);
1544 }
1545 
1546 static void
1547 free_stid_tab(struct tid_info *t)
1548 {
1549 
1550 	KASSERT(t->stids_in_use == 0,
1551 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
1552 
1553 	if (mtx_initialized(&t->stid_lock))
1554 		mtx_destroy(&t->stid_lock);
1555 	free(t->stid_tab, M_CXGBE);
1556 	t->stid_tab = NULL;
1557 }
1558 
1559 static void
1560 free_tid_tabs(struct tid_info *t)
1561 {
1562 
1563 	free_tid_tab(t);
1564 	free_stid_tab(t);
1565 }
1566 
1567 static int
1568 alloc_tid_tabs(struct tid_info *t)
1569 {
1570 	int rc;
1571 
1572 	rc = alloc_tid_tab(t, M_NOWAIT);
1573 	if (rc != 0)
1574 		goto failed;
1575 
1576 	rc = alloc_stid_tab(t, M_NOWAIT);
1577 	if (rc != 0)
1578 		goto failed;
1579 
1580 	return (0);
1581 failed:
1582 	free_tid_tabs(t);
1583 	return (rc);
1584 }
1585 
1586 static inline void
1587 alloc_tcb_history(struct adapter *sc, struct tom_data *td)
1588 {
1589 
1590 	if (sc->tids.ntids == 0 || sc->tids.ntids > 1024)
1591 		return;
1592 	rw_init(&td->tcb_history_lock, "TCB history");
1593 	td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history),
1594 	    M_CXGBE, M_ZERO | M_NOWAIT);
1595 	td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0));
1596 }
1597 
1598 static inline void
1599 free_tcb_history(struct adapter *sc, struct tom_data *td)
1600 {
1601 #ifdef INVARIANTS
1602 	int i;
1603 
1604 	if (td->tcb_history != NULL) {
1605 		for (i = 0; i < sc->tids.ntids; i++) {
1606 			MPASS(td->tcb_history[i] == NULL);
1607 		}
1608 	}
1609 #endif
1610 	free(td->tcb_history, M_CXGBE);
1611 	if (rw_initialized(&td->tcb_history_lock))
1612 		rw_destroy(&td->tcb_history_lock);
1613 }
1614 
1615 static void
1616 free_tom_data(struct adapter *sc, struct tom_data *td)
1617 {
1618 
1619 	ASSERT_SYNCHRONIZED_OP(sc);
1620 
1621 	KASSERT(TAILQ_EMPTY(&td->toep_list),
1622 	    ("%s: TOE PCB list is not empty.", __func__));
1623 	KASSERT(td->lctx_count == 0,
1624 	    ("%s: lctx hash table is not empty.", __func__));
1625 
1626 	t4_free_ppod_region(&td->pr);
1627 
1628 	if (td->listen_mask != 0)
1629 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
1630 
1631 	if (mtx_initialized(&td->unsent_wr_lock))
1632 		mtx_destroy(&td->unsent_wr_lock);
1633 	if (mtx_initialized(&td->lctx_hash_lock))
1634 		mtx_destroy(&td->lctx_hash_lock);
1635 	if (mtx_initialized(&td->toep_list_lock))
1636 		mtx_destroy(&td->toep_list_lock);
1637 
1638 	free_tcb_history(sc, td);
1639 	free_tid_tabs(&sc->tids);
1640 	free(td, M_CXGBE);
1641 }
1642 
1643 static char *
1644 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen,
1645     int *buflen)
1646 {
1647 	char *pkt;
1648 	struct tcphdr *th;
1649 	int ipv6, len;
1650 	const int maxlen =
1651 	    max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) +
1652 	    max(sizeof(struct ip), sizeof(struct ip6_hdr)) +
1653 	    sizeof(struct tcphdr);
1654 
1655 	MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN);
1656 
1657 	pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT);
1658 	if (pkt == NULL)
1659 		return (NULL);
1660 
1661 	ipv6 = inp->inp_vflag & INP_IPV6;
1662 	len = 0;
1663 
1664 	if (EVL_VLANOFTAG(vtag) == 0xfff) {
1665 		struct ether_header *eh = (void *)pkt;
1666 
1667 		if (ipv6)
1668 			eh->ether_type = htons(ETHERTYPE_IPV6);
1669 		else
1670 			eh->ether_type = htons(ETHERTYPE_IP);
1671 
1672 		len += sizeof(*eh);
1673 	} else {
1674 		struct ether_vlan_header *evh = (void *)pkt;
1675 
1676 		evh->evl_encap_proto = htons(ETHERTYPE_VLAN);
1677 		evh->evl_tag = htons(vtag);
1678 		if (ipv6)
1679 			evh->evl_proto = htons(ETHERTYPE_IPV6);
1680 		else
1681 			evh->evl_proto = htons(ETHERTYPE_IP);
1682 
1683 		len += sizeof(*evh);
1684 	}
1685 
1686 	if (ipv6) {
1687 		struct ip6_hdr *ip6 = (void *)&pkt[len];
1688 
1689 		ip6->ip6_vfc = IPV6_VERSION;
1690 		ip6->ip6_plen = htons(sizeof(struct tcphdr));
1691 		ip6->ip6_nxt = IPPROTO_TCP;
1692 		if (open_type == OPEN_TYPE_ACTIVE) {
1693 			ip6->ip6_src = inp->in6p_laddr;
1694 			ip6->ip6_dst = inp->in6p_faddr;
1695 		} else if (open_type == OPEN_TYPE_LISTEN) {
1696 			ip6->ip6_src = inp->in6p_laddr;
1697 			ip6->ip6_dst = ip6->ip6_src;
1698 		}
1699 
1700 		len += sizeof(*ip6);
1701 	} else {
1702 		struct ip *ip = (void *)&pkt[len];
1703 
1704 		ip->ip_v = IPVERSION;
1705 		ip->ip_hl = sizeof(*ip) >> 2;
1706 		ip->ip_tos = inp->inp_ip_tos;
1707 		ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr));
1708 		ip->ip_ttl = inp->inp_ip_ttl;
1709 		ip->ip_p = IPPROTO_TCP;
1710 		if (open_type == OPEN_TYPE_ACTIVE) {
1711 			ip->ip_src = inp->inp_laddr;
1712 			ip->ip_dst = inp->inp_faddr;
1713 		} else if (open_type == OPEN_TYPE_LISTEN) {
1714 			ip->ip_src = inp->inp_laddr;
1715 			ip->ip_dst = ip->ip_src;
1716 		}
1717 
1718 		len += sizeof(*ip);
1719 	}
1720 
1721 	th = (void *)&pkt[len];
1722 	if (open_type == OPEN_TYPE_ACTIVE) {
1723 		th->th_sport = inp->inp_lport;	/* network byte order already */
1724 		th->th_dport = inp->inp_fport;	/* ditto */
1725 	} else if (open_type == OPEN_TYPE_LISTEN) {
1726 		th->th_sport = inp->inp_lport;	/* network byte order already */
1727 		th->th_dport = th->th_sport;
1728 	}
1729 	len += sizeof(th);
1730 
1731 	*pktlen = *buflen = len;
1732 	return (pkt);
1733 }
1734 
1735 const struct offload_settings *
1736 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m,
1737     uint16_t vtag, struct inpcb *inp)
1738 {
1739 	const struct t4_offload_policy *op;
1740 	char *pkt;
1741 	struct offload_rule *r;
1742 	int i, matched, pktlen, buflen;
1743 	static const struct offload_settings allow_offloading_settings = {
1744 		.offload = 1,
1745 		.rx_coalesce = -1,
1746 		.cong_algo = -1,
1747 		.sched_class = -1,
1748 		.tstamp = -1,
1749 		.sack = -1,
1750 		.nagle = -1,
1751 		.ecn = -1,
1752 		.ddp = -1,
1753 		.tls = -1,
1754 		.txq = QUEUE_RANDOM,
1755 		.rxq = QUEUE_RANDOM,
1756 		.mss = -1,
1757 	};
1758 	static const struct offload_settings disallow_offloading_settings = {
1759 		.offload = 0,
1760 		/* rest is irrelevant when offload is off. */
1761 	};
1762 
1763 	rw_assert(&sc->policy_lock, RA_LOCKED);
1764 
1765 	/*
1766 	 * If there's no Connection Offloading Policy attached to the device
1767 	 * then we need to return a default static policy.  If
1768 	 * "cop_managed_offloading" is true, then we need to disallow
1769 	 * offloading until a COP is attached to the device.  Otherwise we
1770 	 * allow offloading ...
1771 	 */
1772 	op = sc->policy;
1773 	if (op == NULL) {
1774 		if (sc->tt.cop_managed_offloading)
1775 			return (&disallow_offloading_settings);
1776 		else
1777 			return (&allow_offloading_settings);
1778 	}
1779 
1780 	switch (open_type) {
1781 	case OPEN_TYPE_ACTIVE:
1782 	case OPEN_TYPE_LISTEN:
1783 		pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen);
1784 		break;
1785 	case OPEN_TYPE_PASSIVE:
1786 		MPASS(m != NULL);
1787 		pkt = mtod(m, char *);
1788 		MPASS(*pkt == CPL_PASS_ACCEPT_REQ);
1789 		pkt += sizeof(struct cpl_pass_accept_req);
1790 		pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req);
1791 		buflen = m->m_len - sizeof(struct cpl_pass_accept_req);
1792 		break;
1793 	default:
1794 		MPASS(0);
1795 		return (&disallow_offloading_settings);
1796 	}
1797 
1798 	if (pkt == NULL || pktlen == 0 || buflen == 0)
1799 		return (&disallow_offloading_settings);
1800 
1801 	matched = 0;
1802 	r = &op->rule[0];
1803 	for (i = 0; i < op->nrules; i++, r++) {
1804 		if (r->open_type != open_type &&
1805 		    r->open_type != OPEN_TYPE_DONTCARE) {
1806 			continue;
1807 		}
1808 		matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen);
1809 		if (matched)
1810 			break;
1811 	}
1812 
1813 	if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN)
1814 		free(pkt, M_CXGBE);
1815 
1816 	return (matched ? &r->settings : &disallow_offloading_settings);
1817 }
1818 
1819 static void
1820 reclaim_wr_resources(void *arg, int count)
1821 {
1822 	struct tom_data *td = arg;
1823 	STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
1824 	struct cpl_act_open_req *cpl;
1825 	u_int opcode, atid, tid;
1826 	struct wrqe *wr;
1827 	struct adapter *sc = td_adapter(td);
1828 
1829 	mtx_lock(&td->unsent_wr_lock);
1830 	STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
1831 	mtx_unlock(&td->unsent_wr_lock);
1832 
1833 	while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
1834 		STAILQ_REMOVE_HEAD(&twr_list, link);
1835 
1836 		cpl = wrtod(wr);
1837 		opcode = GET_OPCODE(cpl);
1838 
1839 		switch (opcode) {
1840 		case CPL_ACT_OPEN_REQ:
1841 		case CPL_ACT_OPEN_REQ6:
1842 			atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1843 			CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1844 			act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1845 			free(wr, M_CXGBE);
1846 			break;
1847 		case CPL_PASS_ACCEPT_RPL:
1848 			tid = GET_TID(cpl);
1849 			CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid);
1850 			synack_failure_cleanup(sc, tid);
1851 			free(wr, M_CXGBE);
1852 			break;
1853 		default:
1854 			log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1855 			    "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1856 			/* WR not freed here; go look at it with a debugger.  */
1857 		}
1858 	}
1859 }
1860 
1861 /*
1862  * Ground control to Major TOM
1863  * Commencing countdown, engines on
1864  */
1865 static int
1866 t4_tom_activate(struct adapter *sc)
1867 {
1868 	struct tom_data *td;
1869 	struct toedev *tod;
1870 	struct vi_info *vi;
1871 	int i, rc, v;
1872 
1873 	ASSERT_SYNCHRONIZED_OP(sc);
1874 
1875 	/* per-adapter softc for TOM */
1876 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1877 	if (td == NULL)
1878 		return (ENOMEM);
1879 
1880 	/* List of TOE PCBs and associated lock */
1881 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1882 	TAILQ_INIT(&td->toep_list);
1883 
1884 	/* Listen context */
1885 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1886 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1887 	    &td->listen_mask, HASH_NOWAIT);
1888 
1889 	/* List of WRs for which L2 resolution failed */
1890 	mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1891 	STAILQ_INIT(&td->unsent_wr_list);
1892 	TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1893 
1894 	/* TID tables */
1895 	rc = alloc_tid_tabs(&sc->tids);
1896 	if (rc != 0)
1897 		goto done;
1898 
1899 	rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1900 	    t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1901 	if (rc != 0)
1902 		goto done;
1903 	t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1904 	    V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1905 
1906 	alloc_tcb_history(sc, td);
1907 
1908 	/* toedev ops */
1909 	tod = &td->tod;
1910 	init_toedev(tod);
1911 	tod->tod_softc = sc;
1912 	tod->tod_connect = t4_connect;
1913 	tod->tod_listen_start = t4_listen_start;
1914 	tod->tod_listen_stop = t4_listen_stop;
1915 	tod->tod_rcvd = t4_rcvd;
1916 	tod->tod_output = t4_tod_output;
1917 	tod->tod_send_rst = t4_send_rst;
1918 	tod->tod_send_fin = t4_send_fin;
1919 	tod->tod_pcb_detach = t4_pcb_detach;
1920 	tod->tod_l2_update = t4_l2_update;
1921 	tod->tod_syncache_added = t4_syncache_added;
1922 	tod->tod_syncache_removed = t4_syncache_removed;
1923 	tod->tod_syncache_respond = t4_syncache_respond;
1924 	tod->tod_offload_socket = t4_offload_socket;
1925 	tod->tod_ctloutput = t4_ctloutput;
1926 	tod->tod_tcp_info = t4_tcp_info;
1927 #ifdef KERN_TLS
1928 	tod->tod_alloc_tls_session = t4_alloc_tls_session;
1929 #endif
1930 	tod->tod_pmtu_update = t4_pmtu_update;
1931 
1932 	for_each_port(sc, i) {
1933 		for_each_vi(sc->port[i], v, vi) {
1934 			TOEDEV(vi->ifp) = &td->tod;
1935 		}
1936 	}
1937 
1938 	sc->tom_softc = td;
1939 	register_toedev(sc->tom_softc);
1940 
1941 done:
1942 	if (rc != 0)
1943 		free_tom_data(sc, td);
1944 	return (rc);
1945 }
1946 
1947 static int
1948 t4_tom_deactivate(struct adapter *sc)
1949 {
1950 	int rc = 0;
1951 	struct tom_data *td = sc->tom_softc;
1952 
1953 	ASSERT_SYNCHRONIZED_OP(sc);
1954 
1955 	if (td == NULL)
1956 		return (0);	/* XXX. KASSERT? */
1957 
1958 	if (sc->offload_map != 0)
1959 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1960 
1961 	if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1962 		return (EBUSY);	/* both iWARP and iSCSI rely on the TOE. */
1963 
1964 	mtx_lock(&td->toep_list_lock);
1965 	if (!TAILQ_EMPTY(&td->toep_list))
1966 		rc = EBUSY;
1967 	mtx_unlock(&td->toep_list_lock);
1968 
1969 	mtx_lock(&td->lctx_hash_lock);
1970 	if (td->lctx_count > 0)
1971 		rc = EBUSY;
1972 	mtx_unlock(&td->lctx_hash_lock);
1973 
1974 	taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1975 	mtx_lock(&td->unsent_wr_lock);
1976 	if (!STAILQ_EMPTY(&td->unsent_wr_list))
1977 		rc = EBUSY;
1978 	mtx_unlock(&td->unsent_wr_lock);
1979 
1980 	if (rc == 0) {
1981 		unregister_toedev(sc->tom_softc);
1982 		free_tom_data(sc, td);
1983 		sc->tom_softc = NULL;
1984 	}
1985 
1986 	return (rc);
1987 }
1988 
1989 static int
1990 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1991 {
1992 	struct tcpcb *tp = so_sototcpcb(so);
1993 	struct toepcb *toep = tp->t_toe;
1994 	int error;
1995 
1996 	/*
1997 	 * No lock is needed as TOE sockets never change between
1998 	 * active and passive.
1999 	 */
2000 	if (SOLISTENING(so))
2001 		return (EINVAL);
2002 
2003 	if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
2004 		error = t4_aio_queue_ddp(so, job);
2005 		if (error != EOPNOTSUPP)
2006 			return (error);
2007 	}
2008 
2009 	return (t4_aio_queue_aiotx(so, job));
2010 }
2011 
2012 static int
2013 t4_tom_mod_load(void)
2014 {
2015 	/* CPL handlers */
2016 	t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl);
2017 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2,
2018 	    CPL_COOKIE_TOM);
2019 	t4_init_connect_cpl_handlers();
2020 	t4_init_listen_cpl_handlers();
2021 	t4_init_cpl_io_handlers();
2022 
2023 	t4_ddp_mod_load();
2024 	t4_tls_mod_load();
2025 
2026 	bcopy(&tcp_protosw, &toe_protosw, sizeof(toe_protosw));
2027 	toe_protosw.pr_aio_queue = t4_aio_queue_tom;
2028 
2029 	bcopy(&tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
2030 	toe6_protosw.pr_aio_queue = t4_aio_queue_tom;
2031 
2032 	return (t4_register_uld(&tom_uld_info));
2033 }
2034 
2035 static void
2036 tom_uninit(struct adapter *sc, void *arg __unused)
2037 {
2038 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
2039 		return;
2040 
2041 	/* Try to free resources (works only if no port has IFCAP_TOE) */
2042 	if (uld_active(sc, ULD_TOM))
2043 		t4_deactivate_uld(sc, ULD_TOM);
2044 
2045 	end_synchronized_op(sc, 0);
2046 }
2047 
2048 static int
2049 t4_tom_mod_unload(void)
2050 {
2051 	t4_iterate(tom_uninit, NULL);
2052 
2053 	if (t4_unregister_uld(&tom_uld_info) == EBUSY)
2054 		return (EBUSY);
2055 
2056 	t4_tls_mod_unload();
2057 	t4_ddp_mod_unload();
2058 
2059 	t4_uninit_connect_cpl_handlers();
2060 	t4_uninit_listen_cpl_handlers();
2061 	t4_uninit_cpl_io_handlers();
2062 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM);
2063 	t4_register_cpl_handler(CPL_GET_TCB_RPL, NULL);
2064 
2065 	return (0);
2066 }
2067 #endif	/* TCP_OFFLOAD */
2068 
2069 static int
2070 t4_tom_modevent(module_t mod, int cmd, void *arg)
2071 {
2072 	int rc = 0;
2073 
2074 #ifdef TCP_OFFLOAD
2075 	switch (cmd) {
2076 	case MOD_LOAD:
2077 		rc = t4_tom_mod_load();
2078 		break;
2079 
2080 	case MOD_UNLOAD:
2081 		rc = t4_tom_mod_unload();
2082 		break;
2083 
2084 	default:
2085 		rc = EINVAL;
2086 	}
2087 #else
2088 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
2089 	rc = EOPNOTSUPP;
2090 #endif
2091 	return (rc);
2092 }
2093 
2094 static moduledata_t t4_tom_moddata= {
2095 	"t4_tom",
2096 	t4_tom_modevent,
2097 	0
2098 };
2099 
2100 MODULE_VERSION(t4_tom, 1);
2101 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
2102 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
2103 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
2104