xref: /freebsd/sys/dev/cxgbe/tom/t4_tom.c (revision f81cdf24ba5436367377f7c8e8f51f6df2a75ca7)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2012 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: Navdeep Parhar <np@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_kern_tls.h"
34 #include "opt_ratelimit.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/systm.h>
39 #include <sys/kernel.h>
40 #include <sys/ktr.h>
41 #include <sys/lock.h>
42 #include <sys/limits.h>
43 #include <sys/module.h>
44 #include <sys/protosw.h>
45 #include <sys/domain.h>
46 #include <sys/refcount.h>
47 #include <sys/rmlock.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
51 #include <sys/taskqueue.h>
52 #include <net/if.h>
53 #include <net/if_var.h>
54 #include <net/if_types.h>
55 #include <net/if_vlan_var.h>
56 #include <netinet/in.h>
57 #include <netinet/in_pcb.h>
58 #include <netinet/in_var.h>
59 #include <netinet/ip.h>
60 #include <netinet/ip6.h>
61 #include <netinet6/scope6_var.h>
62 #define TCPSTATES
63 #include <netinet/tcp_fsm.h>
64 #include <netinet/tcp_seq.h>
65 #include <netinet/tcp_timer.h>
66 #include <netinet/tcp_var.h>
67 #include <netinet/toecore.h>
68 #include <netinet/cc/cc.h>
69 
70 #ifdef TCP_OFFLOAD
71 #include "common/common.h"
72 #include "common/t4_msg.h"
73 #include "common/t4_regs.h"
74 #include "common/t4_regs_values.h"
75 #include "common/t4_tcb.h"
76 #include "t4_clip.h"
77 #include "tom/t4_tom_l2t.h"
78 #include "tom/t4_tom.h"
79 #include "tom/t4_tls.h"
80 
81 static struct protosw toe_protosw;
82 static struct protosw toe6_protosw;
83 
84 /* Module ops */
85 static int t4_tom_mod_load(void);
86 static int t4_tom_mod_unload(void);
87 static int t4_tom_modevent(module_t, int, void *);
88 
89 /* ULD ops and helpers */
90 static int t4_tom_activate(struct adapter *);
91 static int t4_tom_deactivate(struct adapter *);
92 
93 static struct uld_info tom_uld_info = {
94 	.uld_id = ULD_TOM,
95 	.activate = t4_tom_activate,
96 	.deactivate = t4_tom_deactivate,
97 };
98 
99 static void release_offload_resources(struct toepcb *);
100 static int alloc_tid_tabs(struct tid_info *);
101 static void free_tid_tabs(struct tid_info *);
102 static void free_tom_data(struct adapter *, struct tom_data *);
103 static void reclaim_wr_resources(void *, int);
104 
105 struct toepcb *
106 alloc_toepcb(struct vi_info *vi, int flags)
107 {
108 	struct port_info *pi = vi->pi;
109 	struct adapter *sc = pi->adapter;
110 	struct toepcb *toep;
111 	int tx_credits, txsd_total, len;
112 
113 	/*
114 	 * The firmware counts tx work request credits in units of 16 bytes
115 	 * each.  Reserve room for an ABORT_REQ so the driver never has to worry
116 	 * about tx credits if it wants to abort a connection.
117 	 */
118 	tx_credits = sc->params.ofldq_wr_cred;
119 	tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
120 
121 	/*
122 	 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
123 	 * immediate payload, and firmware counts tx work request credits in
124 	 * units of 16 byte.  Calculate the maximum work requests possible.
125 	 */
126 	txsd_total = tx_credits /
127 	    howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
128 
129 	len = offsetof(struct toepcb, txsd) +
130 	    txsd_total * sizeof(struct ofld_tx_sdesc);
131 
132 	toep = malloc(len, M_CXGBE, M_ZERO | flags);
133 	if (toep == NULL)
134 		return (NULL);
135 
136 	refcount_init(&toep->refcount, 1);
137 	toep->td = sc->tom_softc;
138 	toep->vi = vi;
139 	toep->tid = -1;
140 	toep->tx_total = tx_credits;
141 	toep->tx_credits = tx_credits;
142 	mbufq_init(&toep->ulp_pduq, INT_MAX);
143 	mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
144 	toep->txsd_total = txsd_total;
145 	toep->txsd_avail = txsd_total;
146 	toep->txsd_pidx = 0;
147 	toep->txsd_cidx = 0;
148 	aiotx_init_toep(toep);
149 
150 	return (toep);
151 }
152 
153 /*
154  * Initialize a toepcb after its params have been filled out.
155  */
156 int
157 init_toepcb(struct vi_info *vi, struct toepcb *toep)
158 {
159 	struct conn_params *cp = &toep->params;
160 	struct port_info *pi = vi->pi;
161 	struct adapter *sc = pi->adapter;
162 	struct tx_cl_rl_params *tc;
163 
164 	if (cp->tc_idx >= 0 && cp->tc_idx < sc->params.nsched_cls) {
165 		tc = &pi->sched_params->cl_rl[cp->tc_idx];
166 		mtx_lock(&sc->tc_lock);
167 		if (tc->state != CS_HW_CONFIGURED) {
168 			CH_ERR(vi, "tid %d cannot be bound to traffic class %d "
169 			    "because it is not configured (its state is %d)\n",
170 			    toep->tid, cp->tc_idx, tc->state);
171 			cp->tc_idx = -1;
172 		} else {
173 			tc->refcount++;
174 		}
175 		mtx_unlock(&sc->tc_lock);
176 	}
177 	toep->ofld_txq = &sc->sge.ofld_txq[cp->txq_idx];
178 	toep->ofld_rxq = &sc->sge.ofld_rxq[cp->rxq_idx];
179 	toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
180 
181 	tls_init_toep(toep);
182 	MPASS(ulp_mode(toep) != ULP_MODE_TCPDDP);
183 
184 	toep->flags |= TPF_INITIALIZED;
185 
186 	return (0);
187 }
188 
189 struct toepcb *
190 hold_toepcb(struct toepcb *toep)
191 {
192 
193 	refcount_acquire(&toep->refcount);
194 	return (toep);
195 }
196 
197 void
198 free_toepcb(struct toepcb *toep)
199 {
200 
201 	if (refcount_release(&toep->refcount) == 0)
202 		return;
203 
204 	KASSERT(!(toep->flags & TPF_ATTACHED),
205 	    ("%s: attached to an inpcb", __func__));
206 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
207 	    ("%s: CPL pending", __func__));
208 
209 	if (toep->flags & TPF_INITIALIZED) {
210 		if (ulp_mode(toep) == ULP_MODE_TCPDDP)
211 			ddp_uninit_toep(toep);
212 		tls_uninit_toep(toep);
213 	}
214 	free(toep, M_CXGBE);
215 }
216 
217 /*
218  * Set up the socket for TCP offload.
219  */
220 void
221 offload_socket(struct socket *so, struct toepcb *toep)
222 {
223 	struct tom_data *td = toep->td;
224 	struct inpcb *inp = sotoinpcb(so);
225 	struct tcpcb *tp = intotcpcb(inp);
226 	struct sockbuf *sb;
227 
228 	INP_WLOCK_ASSERT(inp);
229 
230 	/* Update socket */
231 	sb = &so->so_snd;
232 	SOCKBUF_LOCK(sb);
233 	sb->sb_flags |= SB_NOCOALESCE;
234 	SOCKBUF_UNLOCK(sb);
235 	sb = &so->so_rcv;
236 	SOCKBUF_LOCK(sb);
237 	sb->sb_flags |= SB_NOCOALESCE;
238 	if (inp->inp_vflag & INP_IPV6)
239 		so->so_proto = &toe6_protosw;
240 	else
241 		so->so_proto = &toe_protosw;
242 	SOCKBUF_UNLOCK(sb);
243 
244 	/* Update TCP PCB */
245 	tp->tod = &td->tod;
246 	tp->t_toe = toep;
247 	tp->t_flags |= TF_TOE;
248 
249 	/* Install an extra hold on inp */
250 	toep->inp = inp;
251 	toep->flags |= TPF_ATTACHED;
252 	in_pcbref(inp);
253 
254 	/* Add the TOE PCB to the active list */
255 	mtx_lock(&td->toep_list_lock);
256 	TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
257 	mtx_unlock(&td->toep_list_lock);
258 }
259 
260 void
261 restore_so_proto(struct socket *so, bool v6)
262 {
263 	if (v6)
264 		so->so_proto = &tcp6_protosw;
265 	else
266 		so->so_proto = &tcp_protosw;
267 }
268 
269 /* This is _not_ the normal way to "unoffload" a socket. */
270 void
271 undo_offload_socket(struct socket *so)
272 {
273 	struct inpcb *inp = sotoinpcb(so);
274 	struct tcpcb *tp = intotcpcb(inp);
275 	struct toepcb *toep = tp->t_toe;
276 	struct tom_data *td = toep->td;
277 	struct sockbuf *sb;
278 
279 	INP_WLOCK_ASSERT(inp);
280 
281 	sb = &so->so_snd;
282 	SOCKBUF_LOCK(sb);
283 	sb->sb_flags &= ~SB_NOCOALESCE;
284 	SOCKBUF_UNLOCK(sb);
285 	sb = &so->so_rcv;
286 	SOCKBUF_LOCK(sb);
287 	sb->sb_flags &= ~SB_NOCOALESCE;
288 	restore_so_proto(so, inp->inp_vflag & INP_IPV6);
289 	SOCKBUF_UNLOCK(sb);
290 
291 	tp->tod = NULL;
292 	tp->t_toe = NULL;
293 	tp->t_flags &= ~TF_TOE;
294 
295 	toep->inp = NULL;
296 	toep->flags &= ~TPF_ATTACHED;
297 	if (in_pcbrele_wlocked(inp))
298 		panic("%s: inp freed.", __func__);
299 
300 	mtx_lock(&td->toep_list_lock);
301 	TAILQ_REMOVE(&td->toep_list, toep, link);
302 	mtx_unlock(&td->toep_list_lock);
303 }
304 
305 static void
306 release_offload_resources(struct toepcb *toep)
307 {
308 	struct tom_data *td = toep->td;
309 	struct adapter *sc = td_adapter(td);
310 	int tid = toep->tid;
311 
312 	KASSERT(!(toep->flags & TPF_CPL_PENDING),
313 	    ("%s: %p has CPL pending.", __func__, toep));
314 	KASSERT(!(toep->flags & TPF_ATTACHED),
315 	    ("%s: %p is still attached.", __func__, toep));
316 
317 	CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
318 	    __func__, toep, tid, toep->l2te, toep->ce);
319 
320 	/*
321 	 * These queues should have been emptied at approximately the same time
322 	 * that a normal connection's socket's so_snd would have been purged or
323 	 * drained.  Do _not_ clean up here.
324 	 */
325 	MPASS(mbufq_empty(&toep->ulp_pduq));
326 	MPASS(mbufq_empty(&toep->ulp_pdu_reclaimq));
327 #ifdef INVARIANTS
328 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
329 		ddp_assert_empty(toep);
330 #endif
331 	MPASS(TAILQ_EMPTY(&toep->aiotx_jobq));
332 
333 	if (toep->l2te)
334 		t4_l2t_release(toep->l2te);
335 
336 	if (tid >= 0) {
337 		remove_tid(sc, tid, toep->ce ? 2 : 1);
338 		release_tid(sc, tid, toep->ctrlq);
339 	}
340 
341 	if (toep->ce)
342 		t4_release_clip_entry(sc, toep->ce);
343 
344 	if (toep->params.tc_idx != -1)
345 		t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx);
346 
347 	mtx_lock(&td->toep_list_lock);
348 	TAILQ_REMOVE(&td->toep_list, toep, link);
349 	mtx_unlock(&td->toep_list_lock);
350 
351 	free_toepcb(toep);
352 }
353 
354 /*
355  * The kernel is done with the TCP PCB and this is our opportunity to unhook the
356  * toepcb hanging off of it.  If the TOE driver is also done with the toepcb (no
357  * pending CPL) then it is time to release all resources tied to the toepcb.
358  *
359  * Also gets called when an offloaded active open fails and the TOM wants the
360  * kernel to take the TCP PCB back.
361  */
362 static void
363 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
364 {
365 #if defined(KTR) || defined(INVARIANTS)
366 	struct inpcb *inp = tptoinpcb(tp);
367 #endif
368 	struct toepcb *toep = tp->t_toe;
369 
370 	INP_WLOCK_ASSERT(inp);
371 
372 	KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
373 	KASSERT(toep->flags & TPF_ATTACHED,
374 	    ("%s: not attached", __func__));
375 
376 #ifdef KTR
377 	if (tp->t_state == TCPS_SYN_SENT) {
378 		CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
379 		    __func__, toep->tid, toep, toep->flags, inp,
380 		    inp->inp_flags);
381 	} else {
382 		CTR6(KTR_CXGBE,
383 		    "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
384 		    toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
385 		    inp->inp_flags);
386 	}
387 #endif
388 
389 	tp->tod = NULL;
390 	tp->t_toe = NULL;
391 	tp->t_flags &= ~TF_TOE;
392 	toep->flags &= ~TPF_ATTACHED;
393 
394 	if (!(toep->flags & TPF_CPL_PENDING))
395 		release_offload_resources(toep);
396 }
397 
398 /*
399  * setsockopt handler.
400  */
401 static void
402 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
403 {
404 	struct adapter *sc = tod->tod_softc;
405 	struct toepcb *toep = tp->t_toe;
406 
407 	if (dir == SOPT_GET)
408 		return;
409 
410 	CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
411 
412 	switch (name) {
413 	case TCP_NODELAY:
414 		if (tp->t_state != TCPS_ESTABLISHED)
415 			break;
416 		toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
417 		t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
418 		    V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0);
419 		break;
420 	default:
421 		break;
422 	}
423 }
424 
425 static inline uint64_t
426 get_tcb_tflags(const uint64_t *tcb)
427 {
428 
429 	return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32));
430 }
431 
432 static inline uint32_t
433 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift)
434 {
435 #define LAST_WORD ((TCB_SIZE / 4) - 1)
436 	uint64_t t1, t2;
437 	int flit_idx;
438 
439 	MPASS(mask != 0);
440 	MPASS(word <= LAST_WORD);
441 	MPASS(shift < 32);
442 
443 	flit_idx = (LAST_WORD - word) / 2;
444 	if (word & 0x1)
445 		shift += 32;
446 	t1 = be64toh(tcb[flit_idx]) >> shift;
447 	t2 = 0;
448 	if (fls(mask) > 64 - shift) {
449 		/*
450 		 * Will spill over into the next logical flit, which is the flit
451 		 * before this one.  The flit_idx before this one must be valid.
452 		 */
453 		MPASS(flit_idx > 0);
454 		t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift);
455 	}
456 	return ((t2 | t1) & mask);
457 #undef LAST_WORD
458 }
459 #define GET_TCB_FIELD(tcb, F) \
460     get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F)
461 
462 /*
463  * Issues a CPL_GET_TCB to read the entire TCB for the tid.
464  */
465 static int
466 send_get_tcb(struct adapter *sc, u_int tid)
467 {
468 	struct cpl_get_tcb *cpl;
469 	struct wrq_cookie cookie;
470 
471 	MPASS(tid >= sc->tids.tid_base);
472 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
473 
474 	cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16),
475 	    &cookie);
476 	if (__predict_false(cpl == NULL))
477 		return (ENOMEM);
478 	bzero(cpl, sizeof(*cpl));
479 	INIT_TP_WR(cpl, tid);
480 	OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid));
481 	cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) |
482 	    V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id));
483 	cpl->cookie = 0xff;
484 	commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie);
485 
486 	return (0);
487 }
488 
489 static struct tcb_histent *
490 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags)
491 {
492 	struct tcb_histent *te;
493 
494 	MPASS(flags == M_NOWAIT || flags == M_WAITOK);
495 
496 	te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags);
497 	if (te == NULL)
498 		return (NULL);
499 	mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF);
500 	callout_init_mtx(&te->te_callout, &te->te_lock, 0);
501 	te->te_adapter = sc;
502 	te->te_tid = tid;
503 
504 	return (te);
505 }
506 
507 static void
508 free_tcb_histent(struct tcb_histent *te)
509 {
510 
511 	mtx_destroy(&te->te_lock);
512 	free(te, M_CXGBE);
513 }
514 
515 /*
516  * Start tracking the tid in the TCB history.
517  */
518 int
519 add_tid_to_history(struct adapter *sc, u_int tid)
520 {
521 	struct tcb_histent *te = NULL;
522 	struct tom_data *td = sc->tom_softc;
523 	int rc;
524 
525 	MPASS(tid >= sc->tids.tid_base);
526 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
527 
528 	if (td->tcb_history == NULL)
529 		return (ENXIO);
530 
531 	rw_wlock(&td->tcb_history_lock);
532 	if (td->tcb_history[tid] != NULL) {
533 		rc = EEXIST;
534 		goto done;
535 	}
536 	te = alloc_tcb_histent(sc, tid, M_NOWAIT);
537 	if (te == NULL) {
538 		rc = ENOMEM;
539 		goto done;
540 	}
541 	mtx_lock(&te->te_lock);
542 	rc = send_get_tcb(sc, tid);
543 	if (rc == 0) {
544 		te->te_flags |= TE_RPL_PENDING;
545 		td->tcb_history[tid] = te;
546 	} else {
547 		free(te, M_CXGBE);
548 	}
549 	mtx_unlock(&te->te_lock);
550 done:
551 	rw_wunlock(&td->tcb_history_lock);
552 	return (rc);
553 }
554 
555 static void
556 remove_tcb_histent(struct tcb_histent *te)
557 {
558 	struct adapter *sc = te->te_adapter;
559 	struct tom_data *td = sc->tom_softc;
560 
561 	rw_assert(&td->tcb_history_lock, RA_WLOCKED);
562 	mtx_assert(&te->te_lock, MA_OWNED);
563 	MPASS(td->tcb_history[te->te_tid] == te);
564 
565 	td->tcb_history[te->te_tid] = NULL;
566 	free_tcb_histent(te);
567 	rw_wunlock(&td->tcb_history_lock);
568 }
569 
570 static inline struct tcb_histent *
571 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem)
572 {
573 	struct tcb_histent *te;
574 	struct tom_data *td = sc->tom_softc;
575 
576 	MPASS(tid >= sc->tids.tid_base);
577 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
578 
579 	if (td->tcb_history == NULL)
580 		return (NULL);
581 
582 	if (addrem)
583 		rw_wlock(&td->tcb_history_lock);
584 	else
585 		rw_rlock(&td->tcb_history_lock);
586 	te = td->tcb_history[tid];
587 	if (te != NULL) {
588 		mtx_lock(&te->te_lock);
589 		return (te);	/* with both locks held */
590 	}
591 	if (addrem)
592 		rw_wunlock(&td->tcb_history_lock);
593 	else
594 		rw_runlock(&td->tcb_history_lock);
595 
596 	return (te);
597 }
598 
599 static inline void
600 release_tcb_histent(struct tcb_histent *te)
601 {
602 	struct adapter *sc = te->te_adapter;
603 	struct tom_data *td = sc->tom_softc;
604 
605 	mtx_assert(&te->te_lock, MA_OWNED);
606 	mtx_unlock(&te->te_lock);
607 	rw_assert(&td->tcb_history_lock, RA_RLOCKED);
608 	rw_runlock(&td->tcb_history_lock);
609 }
610 
611 static void
612 request_tcb(void *arg)
613 {
614 	struct tcb_histent *te = arg;
615 
616 	mtx_assert(&te->te_lock, MA_OWNED);
617 
618 	/* Noone else is supposed to update the histent. */
619 	MPASS(!(te->te_flags & TE_RPL_PENDING));
620 	if (send_get_tcb(te->te_adapter, te->te_tid) == 0)
621 		te->te_flags |= TE_RPL_PENDING;
622 	else
623 		callout_schedule(&te->te_callout, hz / 100);
624 }
625 
626 static void
627 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb)
628 {
629 	struct tom_data *td = te->te_adapter->tom_softc;
630 	uint64_t tflags = get_tcb_tflags(tcb);
631 	uint8_t sample = 0;
632 
633 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) {
634 		if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0)
635 			sample |= TS_RTO;
636 		if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0)
637 			sample |= TS_DUPACKS;
638 		if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold)
639 			sample |= TS_FASTREXMT;
640 	}
641 
642 	if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) {
643 		uint32_t snd_wnd;
644 
645 		sample |= TS_SND_BACKLOGGED;	/* for whatever reason. */
646 
647 		snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
648 		if (tflags & V_TF_RECV_SCALE(1))
649 			snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE);
650 		if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd)
651 			sample |= TS_CWND_LIMITED;	/* maybe due to CWND */
652 	}
653 
654 	if (tflags & V_TF_CCTRL_ECN(1)) {
655 
656 		/*
657 		 * CE marker on incoming IP hdr, echoing ECE back in the TCP
658 		 * hdr.  Indicates congestion somewhere on the way from the peer
659 		 * to this node.
660 		 */
661 		if (tflags & V_TF_CCTRL_ECE(1))
662 			sample |= TS_ECN_ECE;
663 
664 		/*
665 		 * ECE seen and CWR sent (or about to be sent).  Might indicate
666 		 * congestion on the way to the peer.  This node is reducing its
667 		 * congestion window in response.
668 		 */
669 		if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1)))
670 			sample |= TS_ECN_CWR;
671 	}
672 
673 	te->te_sample[te->te_pidx] = sample;
674 	if (++te->te_pidx == nitems(te->te_sample))
675 		te->te_pidx = 0;
676 	memcpy(te->te_tcb, tcb, TCB_SIZE);
677 	te->te_flags |= TE_ACTIVE;
678 }
679 
680 static int
681 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
682 {
683 	struct adapter *sc = iq->adapter;
684 	const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *);
685 	const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1);
686 	struct tcb_histent *te;
687 	const u_int tid = GET_TID(cpl);
688 	bool remove;
689 
690 	remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED;
691 	te = lookup_tcb_histent(sc, tid, remove);
692 	if (te == NULL) {
693 		/* Not in the history.  Who issued the GET_TCB for this? */
694 		device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, "
695 		    "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid,
696 		    (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE),
697 		    GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE),
698 		    GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie);
699 		goto done;
700 	}
701 
702 	MPASS(te->te_flags & TE_RPL_PENDING);
703 	te->te_flags &= ~TE_RPL_PENDING;
704 	if (remove) {
705 		remove_tcb_histent(te);
706 	} else {
707 		update_tcb_histent(te, tcb);
708 		callout_reset(&te->te_callout, hz / 10, request_tcb, te);
709 		release_tcb_histent(te);
710 	}
711 done:
712 	m_freem(m);
713 	return (0);
714 }
715 
716 static void
717 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti)
718 {
719 	uint32_t v;
720 
721 	ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE);
722 
723 	v = GET_TCB_FIELD(tcb, T_SRTT);
724 	ti->tcpi_rtt = tcp_ticks_to_us(sc, v);
725 
726 	v = GET_TCB_FIELD(tcb, T_RTTVAR);
727 	ti->tcpi_rttvar = tcp_ticks_to_us(sc, v);
728 
729 	ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH);
730 	ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND);
731 	ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT);
732 	ti->tcpi_rcv_adv = GET_TCB_FIELD(tcb, RCV_ADV);
733 	ti->tcpi_dupacks = GET_TCB_FIELD(tcb, T_DUPACKS);
734 
735 	v = GET_TCB_FIELD(tcb, TX_MAX);
736 	ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW);
737 	ti->tcpi_snd_una = v - GET_TCB_FIELD(tcb, SND_UNA_RAW);
738 	ti->tcpi_snd_max = v - GET_TCB_FIELD(tcb, SND_MAX_RAW);
739 
740 	/* Receive window being advertised by us. */
741 	ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE);	/* Yes, SND. */
742 	ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND);
743 
744 	/* Send window */
745 	ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE);	/* Yes, RCV. */
746 	ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
747 	if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1))
748 		ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale;
749 	else
750 		ti->tcpi_snd_wscale = 0;
751 
752 }
753 
754 static void
755 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te,
756     struct tcp_info *ti)
757 {
758 
759 	fill_tcp_info_from_tcb(sc, te->te_tcb, ti);
760 }
761 
762 /*
763  * Reads the TCB for the given tid using a memory window and copies it to 'buf'
764  * in the same format as CPL_GET_TCB_RPL.
765  */
766 static void
767 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf)
768 {
769 	int i, j, k, rc;
770 	uint32_t addr;
771 	u_char *tcb, tmp;
772 
773 	MPASS(tid >= sc->tids.tid_base);
774 	MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
775 
776 	addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE;
777 	rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE);
778 	if (rc != 0)
779 		return;
780 
781 	tcb = (u_char *)buf;
782 	for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) {
783 		for (k = 0; k < 16; k++) {
784 			tmp = tcb[i + k];
785 			tcb[i + k] = tcb[j + k];
786 			tcb[j + k] = tmp;
787 		}
788 	}
789 }
790 
791 static void
792 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti)
793 {
794 	uint64_t tcb[TCB_SIZE / sizeof(uint64_t)];
795 	struct tcb_histent *te;
796 
797 	ti->tcpi_toe_tid = tid;
798 	te = lookup_tcb_histent(sc, tid, false);
799 	if (te != NULL) {
800 		fill_tcp_info_from_history(sc, te, ti);
801 		release_tcb_histent(te);
802 	} else {
803 		if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) {
804 			/* XXX: tell firmware to flush TCB cache. */
805 		}
806 		read_tcb_using_memwin(sc, tid, tcb);
807 		fill_tcp_info_from_tcb(sc, tcb, ti);
808 	}
809 }
810 
811 /*
812  * Called by the kernel to allow the TOE driver to "refine" values filled up in
813  * the tcp_info for an offloaded connection.
814  */
815 static void
816 t4_tcp_info(struct toedev *tod, const struct tcpcb *tp, struct tcp_info *ti)
817 {
818 	struct adapter *sc = tod->tod_softc;
819 	struct toepcb *toep = tp->t_toe;
820 
821 	INP_LOCK_ASSERT(tptoinpcb(tp));
822 	MPASS(ti != NULL);
823 
824 	fill_tcp_info(sc, toep->tid, ti);
825 }
826 
827 #ifdef KERN_TLS
828 static int
829 t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp,
830     struct ktls_session *tls, int direction)
831 {
832 	struct toepcb *toep = tp->t_toe;
833 
834 	INP_WLOCK_ASSERT(tptoinpcb(tp));
835 	MPASS(tls != NULL);
836 
837 	return (tls_alloc_ktls(toep, tls, direction));
838 }
839 #endif
840 
841 /* SET_TCB_FIELD sent as a ULP command looks like this */
842 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \
843     sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core))
844 
845 static void *
846 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, uint64_t word, uint64_t mask,
847 		uint64_t val, uint32_t tid)
848 {
849 	struct ulptx_idata *ulpsc;
850 	struct cpl_set_tcb_field_core *req;
851 
852 	ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0));
853 	ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16));
854 
855 	ulpsc = (struct ulptx_idata *)(ulpmc + 1);
856 	ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
857 	ulpsc->len = htobe32(sizeof(*req));
858 
859 	req = (struct cpl_set_tcb_field_core *)(ulpsc + 1);
860 	OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, tid));
861 	req->reply_ctrl = htobe16(V_NO_REPLY(1));
862 	req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0));
863 	req->mask = htobe64(mask);
864 	req->val = htobe64(val);
865 
866 	ulpsc = (struct ulptx_idata *)(req + 1);
867 	if (LEN__SET_TCB_FIELD_ULP % 16) {
868 		ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
869 		ulpsc->len = htobe32(0);
870 		return (ulpsc + 1);
871 	}
872 	return (ulpsc);
873 }
874 
875 static void
876 send_mss_flowc_wr(struct adapter *sc, struct toepcb *toep)
877 {
878 	struct wrq_cookie cookie;
879 	struct fw_flowc_wr *flowc;
880 	struct ofld_tx_sdesc *txsd;
881 	const int flowclen = sizeof(*flowc) + sizeof(struct fw_flowc_mnemval);
882 	const int flowclen16 = howmany(flowclen, 16);
883 
884 	if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0) {
885 		CH_ERR(sc, "%s: tid %u out of tx credits (%d, %d).\n", __func__,
886 		    toep->tid, toep->tx_credits, toep->txsd_avail);
887 		return;
888 	}
889 
890 	flowc = start_wrq_wr(&toep->ofld_txq->wrq, flowclen16, &cookie);
891 	if (__predict_false(flowc == NULL)) {
892 		CH_ERR(sc, "ENOMEM in %s for tid %u.\n", __func__, toep->tid);
893 		return;
894 	}
895 	flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
896 	    V_FW_FLOWC_WR_NPARAMS(1));
897 	flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
898 	    V_FW_WR_FLOWID(toep->tid));
899 	flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_MSS;
900 	flowc->mnemval[0].val = htobe32(toep->params.emss);
901 
902 	txsd = &toep->txsd[toep->txsd_pidx];
903 	txsd->tx_credits = flowclen16;
904 	txsd->plen = 0;
905 	toep->tx_credits -= txsd->tx_credits;
906 	if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
907 		toep->txsd_pidx = 0;
908 	toep->txsd_avail--;
909 	commit_wrq_wr(&toep->ofld_txq->wrq, flowc, &cookie);
910 }
911 
912 static void
913 t4_pmtu_update(struct toedev *tod, struct tcpcb *tp, tcp_seq seq, int mtu)
914 {
915 	struct work_request_hdr *wrh;
916 	struct ulp_txpkt *ulpmc;
917 	int idx, len;
918 	struct wrq_cookie cookie;
919 	struct inpcb *inp = tptoinpcb(tp);
920 	struct toepcb *toep = tp->t_toe;
921 	struct adapter *sc = td_adapter(toep->td);
922 	unsigned short *mtus = &sc->params.mtus[0];
923 
924 	INP_WLOCK_ASSERT(inp);
925 	MPASS(mtu > 0);	/* kernel is supposed to provide something usable. */
926 
927 	/* tp->snd_una and snd_max are in host byte order too. */
928 	seq = be32toh(seq);
929 
930 	CTR6(KTR_CXGBE, "%s: tid %d, seq 0x%08x, mtu %u, mtu_idx %u (%d)",
931 	    __func__, toep->tid, seq, mtu, toep->params.mtu_idx,
932 	    mtus[toep->params.mtu_idx]);
933 
934 	if (ulp_mode(toep) == ULP_MODE_NONE &&	/* XXX: Read TCB otherwise? */
935 	    (SEQ_LT(seq, tp->snd_una) || SEQ_GEQ(seq, tp->snd_max))) {
936 		CTR5(KTR_CXGBE,
937 		    "%s: tid %d, seq 0x%08x not in range [0x%08x, 0x%08x).",
938 		    __func__, toep->tid, seq, tp->snd_una, tp->snd_max);
939 		return;
940 	}
941 
942 	/* Find the best mtu_idx for the suggested MTU. */
943 	for (idx = 0; idx < NMTUS - 1 && mtus[idx + 1] <= mtu; idx++)
944 		continue;
945 	if (idx >= toep->params.mtu_idx)
946 		return;	/* Never increase the PMTU (just like the kernel). */
947 
948 	/*
949 	 * We'll send a compound work request with 2 SET_TCB_FIELDs -- the first
950 	 * one updates the mtu_idx and the second one triggers a retransmit.
951 	 */
952 	len = sizeof(*wrh) + 2 * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
953 	wrh = start_wrq_wr(toep->ctrlq, howmany(len, 16), &cookie);
954 	if (wrh == NULL) {
955 		CH_ERR(sc, "failed to change mtu_idx of tid %d (%u -> %u).\n",
956 		    toep->tid, toep->params.mtu_idx, idx);
957 		return;
958 	}
959 	INIT_ULPTX_WRH(wrh, len, 1, 0);	/* atomic */
960 	ulpmc = (struct ulp_txpkt *)(wrh + 1);
961 	ulpmc = mk_set_tcb_field_ulp(ulpmc, W_TCB_T_MAXSEG,
962 	    V_TCB_T_MAXSEG(M_TCB_T_MAXSEG), V_TCB_T_MAXSEG(idx), toep->tid);
963 	ulpmc = mk_set_tcb_field_ulp(ulpmc, W_TCB_TIMESTAMP,
964 	    V_TCB_TIMESTAMP(0x7FFFFULL << 11), 0, toep->tid);
965 	commit_wrq_wr(toep->ctrlq, wrh, &cookie);
966 
967 	/* Update the software toepcb and tcpcb. */
968 	toep->params.mtu_idx = idx;
969 	tp->t_maxseg = mtus[toep->params.mtu_idx];
970 	if (inp->inp_inc.inc_flags & INC_ISIPV6)
971 		tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
972 	else
973 		tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
974 	toep->params.emss = tp->t_maxseg;
975 	if (tp->t_flags & TF_RCVD_TSTMP)
976 		toep->params.emss -= TCPOLEN_TSTAMP_APPA;
977 
978 	/* Update the firmware flowc. */
979 	send_mss_flowc_wr(sc, toep);
980 
981 	/* Update the MTU in the kernel's hostcache. */
982 	if (sc->tt.update_hc_on_pmtu_change != 0) {
983 		struct in_conninfo inc = {0};
984 
985 		inc.inc_fibnum = inp->inp_inc.inc_fibnum;
986 		if (inp->inp_inc.inc_flags & INC_ISIPV6) {
987 			inc.inc_flags |= INC_ISIPV6;
988 			inc.inc6_faddr = inp->inp_inc.inc6_faddr;
989 		} else {
990 			inc.inc_faddr = inp->inp_inc.inc_faddr;
991 		}
992 		tcp_hc_updatemtu(&inc, mtu);
993 	}
994 
995 	CTR6(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u), t_maxseg %u, emss %u",
996 	    __func__, toep->tid, toep->params.mtu_idx,
997 	    mtus[toep->params.mtu_idx], tp->t_maxseg, toep->params.emss);
998 }
999 
1000 /*
1001  * The TOE driver will not receive any more CPLs for the tid associated with the
1002  * toepcb; release the hold on the inpcb.
1003  */
1004 void
1005 final_cpl_received(struct toepcb *toep)
1006 {
1007 	struct inpcb *inp = toep->inp;
1008 	bool need_wakeup;
1009 
1010 	KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
1011 	INP_WLOCK_ASSERT(inp);
1012 	KASSERT(toep->flags & TPF_CPL_PENDING,
1013 	    ("%s: CPL not pending already?", __func__));
1014 
1015 	CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
1016 	    __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
1017 
1018 	if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1019 		release_ddp_resources(toep);
1020 	toep->inp = NULL;
1021 	need_wakeup = (toep->flags & TPF_WAITING_FOR_FINAL) != 0;
1022 	toep->flags &= ~(TPF_CPL_PENDING | TPF_WAITING_FOR_FINAL);
1023 	mbufq_drain(&toep->ulp_pduq);
1024 	mbufq_drain(&toep->ulp_pdu_reclaimq);
1025 
1026 	if (!(toep->flags & TPF_ATTACHED))
1027 		release_offload_resources(toep);
1028 
1029 	if (!in_pcbrele_wlocked(inp))
1030 		INP_WUNLOCK(inp);
1031 
1032 	if (need_wakeup) {
1033 		struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep);
1034 
1035 		mtx_lock(lock);
1036 		wakeup(toep);
1037 		mtx_unlock(lock);
1038 	}
1039 }
1040 
1041 void
1042 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
1043 {
1044 	struct tid_info *t = &sc->tids;
1045 
1046 	MPASS(tid >= t->tid_base);
1047 	MPASS(tid - t->tid_base < t->ntids);
1048 
1049 	t->tid_tab[tid - t->tid_base] = ctx;
1050 	atomic_add_int(&t->tids_in_use, ntids);
1051 }
1052 
1053 void *
1054 lookup_tid(struct adapter *sc, int tid)
1055 {
1056 	struct tid_info *t = &sc->tids;
1057 
1058 	return (t->tid_tab[tid - t->tid_base]);
1059 }
1060 
1061 void
1062 update_tid(struct adapter *sc, int tid, void *ctx)
1063 {
1064 	struct tid_info *t = &sc->tids;
1065 
1066 	t->tid_tab[tid - t->tid_base] = ctx;
1067 }
1068 
1069 void
1070 remove_tid(struct adapter *sc, int tid, int ntids)
1071 {
1072 	struct tid_info *t = &sc->tids;
1073 
1074 	t->tid_tab[tid - t->tid_base] = NULL;
1075 	atomic_subtract_int(&t->tids_in_use, ntids);
1076 }
1077 
1078 /*
1079  * What mtu_idx to use, given a 4-tuple.  Note that both s->mss and tcp_mssopt
1080  * have the MSS that we should advertise in our SYN.  Advertised MSS doesn't
1081  * account for any TCP options so the effective MSS (only payload, no headers or
1082  * options) could be different.
1083  */
1084 static int
1085 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc,
1086     struct offload_settings *s)
1087 {
1088 	unsigned short *mtus = &sc->params.mtus[0];
1089 	int i, mss, mtu;
1090 
1091 	MPASS(inc != NULL);
1092 
1093 	mss = s->mss > 0 ? s->mss : tcp_mssopt(inc);
1094 	if (inc->inc_flags & INC_ISIPV6)
1095 		mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1096 	else
1097 		mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr);
1098 
1099 	for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++)
1100 		continue;
1101 
1102 	return (i);
1103 }
1104 
1105 /*
1106  * Determine the receive window size for a socket.
1107  */
1108 u_long
1109 select_rcv_wnd(struct socket *so)
1110 {
1111 	unsigned long wnd;
1112 
1113 	SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1114 
1115 	wnd = sbspace(&so->so_rcv);
1116 	if (wnd < MIN_RCV_WND)
1117 		wnd = MIN_RCV_WND;
1118 
1119 	return min(wnd, MAX_RCV_WND);
1120 }
1121 
1122 int
1123 select_rcv_wscale(void)
1124 {
1125 	int wscale = 0;
1126 	unsigned long space = sb_max;
1127 
1128 	if (space > MAX_RCV_WND)
1129 		space = MAX_RCV_WND;
1130 
1131 	while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
1132 		wscale++;
1133 
1134 	return (wscale);
1135 }
1136 
1137 __be64
1138 calc_options0(struct vi_info *vi, struct conn_params *cp)
1139 {
1140 	uint64_t opt0 = 0;
1141 
1142 	opt0 |= F_TCAM_BYPASS;
1143 
1144 	MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE);
1145 	opt0 |= V_WND_SCALE(cp->wscale);
1146 
1147 	MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS);
1148 	opt0 |= V_MSS_IDX(cp->mtu_idx);
1149 
1150 	MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE);
1151 	opt0 |= V_ULP_MODE(cp->ulp_mode);
1152 
1153 	MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ);
1154 	opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize);
1155 
1156 	MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->adapter->vres.l2t.size);
1157 	opt0 |= V_L2T_IDX(cp->l2t_idx);
1158 
1159 	opt0 |= V_SMAC_SEL(vi->smt_idx);
1160 	opt0 |= V_TX_CHAN(vi->pi->tx_chan);
1161 
1162 	MPASS(cp->keepalive == 0 || cp->keepalive == 1);
1163 	opt0 |= V_KEEP_ALIVE(cp->keepalive);
1164 
1165 	MPASS(cp->nagle == 0 || cp->nagle == 1);
1166 	opt0 |= V_NAGLE(cp->nagle);
1167 
1168 	return (htobe64(opt0));
1169 }
1170 
1171 __be32
1172 calc_options2(struct vi_info *vi, struct conn_params *cp)
1173 {
1174 	uint32_t opt2 = 0;
1175 	struct port_info *pi = vi->pi;
1176 	struct adapter *sc = pi->adapter;
1177 
1178 	/*
1179 	 * rx flow control, rx coalesce, congestion control, and tx pace are all
1180 	 * explicitly set by the driver.  On T5+ the ISS is also set by the
1181 	 * driver to the value picked by the kernel.
1182 	 */
1183 	if (is_t4(sc)) {
1184 		opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID;
1185 		opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID;
1186 	} else {
1187 		opt2 |= F_T5_OPT_2_VALID;	/* all 4 valid */
1188 		opt2 |= F_T5_ISS;		/* ISS provided in CPL */
1189 	}
1190 
1191 	MPASS(cp->sack == 0 || cp->sack == 1);
1192 	opt2 |= V_SACK_EN(cp->sack);
1193 
1194 	MPASS(cp->tstamp == 0 || cp->tstamp == 1);
1195 	opt2 |= V_TSTAMPS_EN(cp->tstamp);
1196 
1197 	if (cp->wscale > 0)
1198 		opt2 |= F_WND_SCALE_EN;
1199 
1200 	MPASS(cp->ecn == 0 || cp->ecn == 1);
1201 	opt2 |= V_CCTRL_ECN(cp->ecn);
1202 
1203 	/* XXX: F_RX_CHANNEL for multiple rx c-chan support goes here. */
1204 
1205 	opt2 |= V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]);
1206 	opt2 |= V_PACE(0);
1207 	opt2 |= F_RSS_QUEUE_VALID;
1208 	opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id);
1209 
1210 	MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL);
1211 	opt2 |= V_CONG_CNTRL(cp->cong_algo);
1212 
1213 	MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1);
1214 	if (cp->rx_coalesce == 1)
1215 		opt2 |= V_RX_COALESCE(M_RX_COALESCE);
1216 
1217 	opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
1218 	MPASS(cp->ulp_mode != ULP_MODE_TCPDDP);
1219 
1220 	return (htobe32(opt2));
1221 }
1222 
1223 uint64_t
1224 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
1225 {
1226 	struct adapter *sc = vi->adapter;
1227 	struct tp_params *tp = &sc->params.tp;
1228 	uint64_t ntuple = 0;
1229 
1230 	/*
1231 	 * Initialize each of the fields which we care about which are present
1232 	 * in the Compressed Filter Tuple.
1233 	 */
1234 	if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE)
1235 		ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
1236 
1237 	if (tp->port_shift >= 0)
1238 		ntuple |= (uint64_t)e->lport << tp->port_shift;
1239 
1240 	if (tp->protocol_shift >= 0)
1241 		ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
1242 
1243 	if (tp->vnic_shift >= 0 && tp->vnic_mode == FW_VNIC_MODE_PF_VF) {
1244 		ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) |
1245 		    V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) <<
1246 		    tp->vnic_shift;
1247 	}
1248 
1249 	if (is_t4(sc))
1250 		return (htobe32((uint32_t)ntuple));
1251 	else
1252 		return (htobe64(V_FILTER_TUPLE(ntuple)));
1253 }
1254 
1255 /*
1256  * Initialize various connection parameters.
1257  */
1258 void
1259 init_conn_params(struct vi_info *vi , struct offload_settings *s,
1260     struct in_conninfo *inc, struct socket *so,
1261     const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp)
1262 {
1263 	struct port_info *pi = vi->pi;
1264 	struct adapter *sc = pi->adapter;
1265 	struct tom_tunables *tt = &sc->tt;
1266 	struct inpcb *inp = sotoinpcb(so);
1267 	struct tcpcb *tp = intotcpcb(inp);
1268 	u_long wnd;
1269 	u_int q_idx;
1270 
1271 	MPASS(s->offload != 0);
1272 
1273 	/* Congestion control algorithm */
1274 	if (s->cong_algo >= 0)
1275 		cp->cong_algo = s->cong_algo & M_CONG_CNTRL;
1276 	else if (sc->tt.cong_algorithm >= 0)
1277 		cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL;
1278 	else {
1279 		struct cc_algo *cc = CC_ALGO(tp);
1280 
1281 		if (strcasecmp(cc->name, "reno") == 0)
1282 			cp->cong_algo = CONG_ALG_RENO;
1283 		else if (strcasecmp(cc->name, "tahoe") == 0)
1284 			cp->cong_algo = CONG_ALG_TAHOE;
1285 		if (strcasecmp(cc->name, "newreno") == 0)
1286 			cp->cong_algo = CONG_ALG_NEWRENO;
1287 		if (strcasecmp(cc->name, "highspeed") == 0)
1288 			cp->cong_algo = CONG_ALG_HIGHSPEED;
1289 		else {
1290 			/*
1291 			 * Use newreno in case the algorithm selected by the
1292 			 * host stack is not supported by the hardware.
1293 			 */
1294 			cp->cong_algo = CONG_ALG_NEWRENO;
1295 		}
1296 	}
1297 
1298 	/* Tx traffic scheduling class. */
1299 	if (s->sched_class >= 0 && s->sched_class < sc->params.nsched_cls)
1300 		cp->tc_idx = s->sched_class;
1301 	else
1302 		cp->tc_idx = -1;
1303 
1304 	/* Nagle's algorithm. */
1305 	if (s->nagle >= 0)
1306 		cp->nagle = s->nagle > 0 ? 1 : 0;
1307 	else
1308 		cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
1309 
1310 	/* TCP Keepalive. */
1311 	if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE)
1312 		cp->keepalive = 1;
1313 	else
1314 		cp->keepalive = 0;
1315 
1316 	/* Optimization that's specific to T5 @ 40G. */
1317 	if (tt->tx_align >= 0)
1318 		cp->tx_align =  tt->tx_align > 0 ? 1 : 0;
1319 	else if (chip_id(sc) == CHELSIO_T5 &&
1320 	    (port_top_speed(pi) > 10 || sc->params.nports > 2))
1321 		cp->tx_align = 1;
1322 	else
1323 		cp->tx_align = 0;
1324 
1325 	/* ULP mode. */
1326 	cp->ulp_mode = ULP_MODE_NONE;
1327 
1328 	/* Rx coalescing. */
1329 	if (s->rx_coalesce >= 0)
1330 		cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0;
1331 	else if (tt->rx_coalesce >= 0)
1332 		cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0;
1333 	else
1334 		cp->rx_coalesce = 1;	/* default */
1335 
1336 	/*
1337 	 * Index in the PMTU table.  This controls the MSS that we announce in
1338 	 * our SYN initially, but after ESTABLISHED it controls the MSS that we
1339 	 * use to send data.
1340 	 */
1341 	cp->mtu_idx = find_best_mtu_idx(sc, inc, s);
1342 
1343 	/* Tx queue for this connection. */
1344 	if (s->txq == QUEUE_RANDOM)
1345 		q_idx = arc4random();
1346 	else if (s->txq == QUEUE_ROUNDROBIN)
1347 		q_idx = atomic_fetchadd_int(&vi->txq_rr, 1);
1348 	else
1349 		q_idx = s->txq;
1350 	cp->txq_idx = vi->first_ofld_txq + q_idx % vi->nofldtxq;
1351 
1352 	/* Rx queue for this connection. */
1353 	if (s->rxq == QUEUE_RANDOM)
1354 		q_idx = arc4random();
1355 	else if (s->rxq == QUEUE_ROUNDROBIN)
1356 		q_idx = atomic_fetchadd_int(&vi->rxq_rr, 1);
1357 	else
1358 		q_idx = s->rxq;
1359 	cp->rxq_idx = vi->first_ofld_rxq + q_idx % vi->nofldrxq;
1360 
1361 	if (SOLISTENING(so)) {
1362 		/* Passive open */
1363 		MPASS(tcpopt != NULL);
1364 
1365 		/* TCP timestamp option */
1366 		if (tcpopt->tstamp &&
1367 		    (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323)))
1368 			cp->tstamp = 1;
1369 		else
1370 			cp->tstamp = 0;
1371 
1372 		/* SACK */
1373 		if (tcpopt->sack &&
1374 		    (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack)))
1375 			cp->sack = 1;
1376 		else
1377 			cp->sack = 0;
1378 
1379 		/* Receive window scaling. */
1380 		if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323)
1381 			cp->wscale = select_rcv_wscale();
1382 		else
1383 			cp->wscale = 0;
1384 
1385 		/* ECN */
1386 		if (tcpopt->ecn &&	/* XXX: review. */
1387 		    (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn)))
1388 			cp->ecn = 1;
1389 		else
1390 			cp->ecn = 0;
1391 
1392 		wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND);
1393 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1394 
1395 		if (tt->sndbuf > 0)
1396 			cp->sndbuf = tt->sndbuf;
1397 		else if (so->sol_sbsnd_flags & SB_AUTOSIZE &&
1398 		    V_tcp_do_autosndbuf)
1399 			cp->sndbuf = 256 * 1024;
1400 		else
1401 			cp->sndbuf = so->sol_sbsnd_hiwat;
1402 	} else {
1403 		/* Active open */
1404 
1405 		/* TCP timestamp option */
1406 		if (s->tstamp > 0 ||
1407 		    (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP)))
1408 			cp->tstamp = 1;
1409 		else
1410 			cp->tstamp = 0;
1411 
1412 		/* SACK */
1413 		if (s->sack > 0 ||
1414 		    (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT)))
1415 			cp->sack = 1;
1416 		else
1417 			cp->sack = 0;
1418 
1419 		/* Receive window scaling */
1420 		if (tp->t_flags & TF_REQ_SCALE)
1421 			cp->wscale = select_rcv_wscale();
1422 		else
1423 			cp->wscale = 0;
1424 
1425 		/* ECN */
1426 		if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1))
1427 			cp->ecn = 1;
1428 		else
1429 			cp->ecn = 0;
1430 
1431 		SOCKBUF_LOCK(&so->so_rcv);
1432 		wnd = max(select_rcv_wnd(so), MIN_RCV_WND);
1433 		SOCKBUF_UNLOCK(&so->so_rcv);
1434 		cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1435 
1436 		if (tt->sndbuf > 0)
1437 			cp->sndbuf = tt->sndbuf;
1438 		else {
1439 			SOCKBUF_LOCK(&so->so_snd);
1440 			if (so->so_snd.sb_flags & SB_AUTOSIZE &&
1441 			    V_tcp_do_autosndbuf)
1442 				cp->sndbuf = 256 * 1024;
1443 			else
1444 				cp->sndbuf = so->so_snd.sb_hiwat;
1445 			SOCKBUF_UNLOCK(&so->so_snd);
1446 		}
1447 	}
1448 
1449 	cp->l2t_idx = l2t_idx;
1450 
1451 	/* This will be initialized on ESTABLISHED. */
1452 	cp->emss = 0;
1453 }
1454 
1455 int
1456 negative_advice(int status)
1457 {
1458 
1459 	return (status == CPL_ERR_RTX_NEG_ADVICE ||
1460 	    status == CPL_ERR_PERSIST_NEG_ADVICE ||
1461 	    status == CPL_ERR_KEEPALV_NEG_ADVICE);
1462 }
1463 
1464 static int
1465 alloc_tid_tab(struct tid_info *t, int flags)
1466 {
1467 
1468 	MPASS(t->ntids > 0);
1469 	MPASS(t->tid_tab == NULL);
1470 
1471 	t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE,
1472 	    M_ZERO | flags);
1473 	if (t->tid_tab == NULL)
1474 		return (ENOMEM);
1475 	atomic_store_rel_int(&t->tids_in_use, 0);
1476 
1477 	return (0);
1478 }
1479 
1480 static void
1481 free_tid_tab(struct tid_info *t)
1482 {
1483 
1484 	KASSERT(t->tids_in_use == 0,
1485 	    ("%s: %d tids still in use.", __func__, t->tids_in_use));
1486 
1487 	free(t->tid_tab, M_CXGBE);
1488 	t->tid_tab = NULL;
1489 }
1490 
1491 static int
1492 alloc_stid_tab(struct tid_info *t, int flags)
1493 {
1494 
1495 	MPASS(t->nstids > 0);
1496 	MPASS(t->stid_tab == NULL);
1497 
1498 	t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE,
1499 	    M_ZERO | flags);
1500 	if (t->stid_tab == NULL)
1501 		return (ENOMEM);
1502 	mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
1503 	t->stids_in_use = 0;
1504 	TAILQ_INIT(&t->stids);
1505 	t->nstids_free_head = t->nstids;
1506 
1507 	return (0);
1508 }
1509 
1510 static void
1511 free_stid_tab(struct tid_info *t)
1512 {
1513 
1514 	KASSERT(t->stids_in_use == 0,
1515 	    ("%s: %d tids still in use.", __func__, t->stids_in_use));
1516 
1517 	if (mtx_initialized(&t->stid_lock))
1518 		mtx_destroy(&t->stid_lock);
1519 	free(t->stid_tab, M_CXGBE);
1520 	t->stid_tab = NULL;
1521 }
1522 
1523 static void
1524 free_tid_tabs(struct tid_info *t)
1525 {
1526 
1527 	free_tid_tab(t);
1528 	free_stid_tab(t);
1529 }
1530 
1531 static int
1532 alloc_tid_tabs(struct tid_info *t)
1533 {
1534 	int rc;
1535 
1536 	rc = alloc_tid_tab(t, M_NOWAIT);
1537 	if (rc != 0)
1538 		goto failed;
1539 
1540 	rc = alloc_stid_tab(t, M_NOWAIT);
1541 	if (rc != 0)
1542 		goto failed;
1543 
1544 	return (0);
1545 failed:
1546 	free_tid_tabs(t);
1547 	return (rc);
1548 }
1549 
1550 static inline void
1551 alloc_tcb_history(struct adapter *sc, struct tom_data *td)
1552 {
1553 
1554 	if (sc->tids.ntids == 0 || sc->tids.ntids > 1024)
1555 		return;
1556 	rw_init(&td->tcb_history_lock, "TCB history");
1557 	td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history),
1558 	    M_CXGBE, M_ZERO | M_NOWAIT);
1559 	td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0));
1560 }
1561 
1562 static inline void
1563 free_tcb_history(struct adapter *sc, struct tom_data *td)
1564 {
1565 #ifdef INVARIANTS
1566 	int i;
1567 
1568 	if (td->tcb_history != NULL) {
1569 		for (i = 0; i < sc->tids.ntids; i++) {
1570 			MPASS(td->tcb_history[i] == NULL);
1571 		}
1572 	}
1573 #endif
1574 	free(td->tcb_history, M_CXGBE);
1575 	if (rw_initialized(&td->tcb_history_lock))
1576 		rw_destroy(&td->tcb_history_lock);
1577 }
1578 
1579 static void
1580 free_tom_data(struct adapter *sc, struct tom_data *td)
1581 {
1582 
1583 	ASSERT_SYNCHRONIZED_OP(sc);
1584 
1585 	KASSERT(TAILQ_EMPTY(&td->toep_list),
1586 	    ("%s: TOE PCB list is not empty.", __func__));
1587 	KASSERT(td->lctx_count == 0,
1588 	    ("%s: lctx hash table is not empty.", __func__));
1589 
1590 	t4_free_ppod_region(&td->pr);
1591 
1592 	if (td->listen_mask != 0)
1593 		hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
1594 
1595 	if (mtx_initialized(&td->unsent_wr_lock))
1596 		mtx_destroy(&td->unsent_wr_lock);
1597 	if (mtx_initialized(&td->lctx_hash_lock))
1598 		mtx_destroy(&td->lctx_hash_lock);
1599 	if (mtx_initialized(&td->toep_list_lock))
1600 		mtx_destroy(&td->toep_list_lock);
1601 
1602 	free_tcb_history(sc, td);
1603 	free_tid_tabs(&sc->tids);
1604 	free(td, M_CXGBE);
1605 }
1606 
1607 static char *
1608 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen,
1609     int *buflen)
1610 {
1611 	char *pkt;
1612 	struct tcphdr *th;
1613 	int ipv6, len;
1614 	const int maxlen =
1615 	    max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) +
1616 	    max(sizeof(struct ip), sizeof(struct ip6_hdr)) +
1617 	    sizeof(struct tcphdr);
1618 
1619 	MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN);
1620 
1621 	pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT);
1622 	if (pkt == NULL)
1623 		return (NULL);
1624 
1625 	ipv6 = inp->inp_vflag & INP_IPV6;
1626 	len = 0;
1627 
1628 	if (EVL_VLANOFTAG(vtag) == 0xfff) {
1629 		struct ether_header *eh = (void *)pkt;
1630 
1631 		if (ipv6)
1632 			eh->ether_type = htons(ETHERTYPE_IPV6);
1633 		else
1634 			eh->ether_type = htons(ETHERTYPE_IP);
1635 
1636 		len += sizeof(*eh);
1637 	} else {
1638 		struct ether_vlan_header *evh = (void *)pkt;
1639 
1640 		evh->evl_encap_proto = htons(ETHERTYPE_VLAN);
1641 		evh->evl_tag = htons(vtag);
1642 		if (ipv6)
1643 			evh->evl_proto = htons(ETHERTYPE_IPV6);
1644 		else
1645 			evh->evl_proto = htons(ETHERTYPE_IP);
1646 
1647 		len += sizeof(*evh);
1648 	}
1649 
1650 	if (ipv6) {
1651 		struct ip6_hdr *ip6 = (void *)&pkt[len];
1652 
1653 		ip6->ip6_vfc = IPV6_VERSION;
1654 		ip6->ip6_plen = htons(sizeof(struct tcphdr));
1655 		ip6->ip6_nxt = IPPROTO_TCP;
1656 		if (open_type == OPEN_TYPE_ACTIVE) {
1657 			ip6->ip6_src = inp->in6p_laddr;
1658 			ip6->ip6_dst = inp->in6p_faddr;
1659 		} else if (open_type == OPEN_TYPE_LISTEN) {
1660 			ip6->ip6_src = inp->in6p_laddr;
1661 			ip6->ip6_dst = ip6->ip6_src;
1662 		}
1663 
1664 		len += sizeof(*ip6);
1665 	} else {
1666 		struct ip *ip = (void *)&pkt[len];
1667 
1668 		ip->ip_v = IPVERSION;
1669 		ip->ip_hl = sizeof(*ip) >> 2;
1670 		ip->ip_tos = inp->inp_ip_tos;
1671 		ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr));
1672 		ip->ip_ttl = inp->inp_ip_ttl;
1673 		ip->ip_p = IPPROTO_TCP;
1674 		if (open_type == OPEN_TYPE_ACTIVE) {
1675 			ip->ip_src = inp->inp_laddr;
1676 			ip->ip_dst = inp->inp_faddr;
1677 		} else if (open_type == OPEN_TYPE_LISTEN) {
1678 			ip->ip_src = inp->inp_laddr;
1679 			ip->ip_dst = ip->ip_src;
1680 		}
1681 
1682 		len += sizeof(*ip);
1683 	}
1684 
1685 	th = (void *)&pkt[len];
1686 	if (open_type == OPEN_TYPE_ACTIVE) {
1687 		th->th_sport = inp->inp_lport;	/* network byte order already */
1688 		th->th_dport = inp->inp_fport;	/* ditto */
1689 	} else if (open_type == OPEN_TYPE_LISTEN) {
1690 		th->th_sport = inp->inp_lport;	/* network byte order already */
1691 		th->th_dport = th->th_sport;
1692 	}
1693 	len += sizeof(th);
1694 
1695 	*pktlen = *buflen = len;
1696 	return (pkt);
1697 }
1698 
1699 const struct offload_settings *
1700 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m,
1701     uint16_t vtag, struct inpcb *inp)
1702 {
1703 	const struct t4_offload_policy *op;
1704 	char *pkt;
1705 	struct offload_rule *r;
1706 	int i, matched, pktlen, buflen;
1707 	static const struct offload_settings allow_offloading_settings = {
1708 		.offload = 1,
1709 		.rx_coalesce = -1,
1710 		.cong_algo = -1,
1711 		.sched_class = -1,
1712 		.tstamp = -1,
1713 		.sack = -1,
1714 		.nagle = -1,
1715 		.ecn = -1,
1716 		.ddp = -1,
1717 		.tls = -1,
1718 		.txq = QUEUE_RANDOM,
1719 		.rxq = QUEUE_RANDOM,
1720 		.mss = -1,
1721 	};
1722 	static const struct offload_settings disallow_offloading_settings = {
1723 		.offload = 0,
1724 		/* rest is irrelevant when offload is off. */
1725 	};
1726 
1727 	rw_assert(&sc->policy_lock, RA_LOCKED);
1728 
1729 	/*
1730 	 * If there's no Connection Offloading Policy attached to the device
1731 	 * then we need to return a default static policy.  If
1732 	 * "cop_managed_offloading" is true, then we need to disallow
1733 	 * offloading until a COP is attached to the device.  Otherwise we
1734 	 * allow offloading ...
1735 	 */
1736 	op = sc->policy;
1737 	if (op == NULL) {
1738 		if (sc->tt.cop_managed_offloading)
1739 			return (&disallow_offloading_settings);
1740 		else
1741 			return (&allow_offloading_settings);
1742 	}
1743 
1744 	switch (open_type) {
1745 	case OPEN_TYPE_ACTIVE:
1746 	case OPEN_TYPE_LISTEN:
1747 		pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen);
1748 		break;
1749 	case OPEN_TYPE_PASSIVE:
1750 		MPASS(m != NULL);
1751 		pkt = mtod(m, char *);
1752 		MPASS(*pkt == CPL_PASS_ACCEPT_REQ);
1753 		pkt += sizeof(struct cpl_pass_accept_req);
1754 		pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req);
1755 		buflen = m->m_len - sizeof(struct cpl_pass_accept_req);
1756 		break;
1757 	default:
1758 		MPASS(0);
1759 		return (&disallow_offloading_settings);
1760 	}
1761 
1762 	if (pkt == NULL || pktlen == 0 || buflen == 0)
1763 		return (&disallow_offloading_settings);
1764 
1765 	matched = 0;
1766 	r = &op->rule[0];
1767 	for (i = 0; i < op->nrules; i++, r++) {
1768 		if (r->open_type != open_type &&
1769 		    r->open_type != OPEN_TYPE_DONTCARE) {
1770 			continue;
1771 		}
1772 		matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen);
1773 		if (matched)
1774 			break;
1775 	}
1776 
1777 	if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN)
1778 		free(pkt, M_CXGBE);
1779 
1780 	return (matched ? &r->settings : &disallow_offloading_settings);
1781 }
1782 
1783 static void
1784 reclaim_wr_resources(void *arg, int count)
1785 {
1786 	struct tom_data *td = arg;
1787 	STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
1788 	struct cpl_act_open_req *cpl;
1789 	u_int opcode, atid, tid;
1790 	struct wrqe *wr;
1791 	struct adapter *sc = td_adapter(td);
1792 
1793 	mtx_lock(&td->unsent_wr_lock);
1794 	STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
1795 	mtx_unlock(&td->unsent_wr_lock);
1796 
1797 	while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
1798 		STAILQ_REMOVE_HEAD(&twr_list, link);
1799 
1800 		cpl = wrtod(wr);
1801 		opcode = GET_OPCODE(cpl);
1802 
1803 		switch (opcode) {
1804 		case CPL_ACT_OPEN_REQ:
1805 		case CPL_ACT_OPEN_REQ6:
1806 			atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1807 			CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1808 			act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1809 			free(wr, M_CXGBE);
1810 			break;
1811 		case CPL_PASS_ACCEPT_RPL:
1812 			tid = GET_TID(cpl);
1813 			CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid);
1814 			synack_failure_cleanup(sc, tid);
1815 			free(wr, M_CXGBE);
1816 			break;
1817 		default:
1818 			log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1819 			    "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1820 			/* WR not freed here; go look at it with a debugger.  */
1821 		}
1822 	}
1823 }
1824 
1825 /*
1826  * Ground control to Major TOM
1827  * Commencing countdown, engines on
1828  */
1829 static int
1830 t4_tom_activate(struct adapter *sc)
1831 {
1832 	struct tom_data *td;
1833 	struct toedev *tod;
1834 	struct vi_info *vi;
1835 	int i, rc, v;
1836 
1837 	ASSERT_SYNCHRONIZED_OP(sc);
1838 
1839 	/* per-adapter softc for TOM */
1840 	td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1841 	if (td == NULL)
1842 		return (ENOMEM);
1843 
1844 	/* List of TOE PCBs and associated lock */
1845 	mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1846 	TAILQ_INIT(&td->toep_list);
1847 
1848 	/* Listen context */
1849 	mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1850 	td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1851 	    &td->listen_mask, HASH_NOWAIT);
1852 
1853 	/* List of WRs for which L2 resolution failed */
1854 	mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1855 	STAILQ_INIT(&td->unsent_wr_list);
1856 	TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1857 
1858 	/* TID tables */
1859 	rc = alloc_tid_tabs(&sc->tids);
1860 	if (rc != 0)
1861 		goto done;
1862 
1863 	rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1864 	    t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1865 	if (rc != 0)
1866 		goto done;
1867 	t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1868 	    V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1869 
1870 	alloc_tcb_history(sc, td);
1871 
1872 	/* toedev ops */
1873 	tod = &td->tod;
1874 	init_toedev(tod);
1875 	tod->tod_softc = sc;
1876 	tod->tod_connect = t4_connect;
1877 	tod->tod_listen_start = t4_listen_start;
1878 	tod->tod_listen_stop = t4_listen_stop;
1879 	tod->tod_rcvd = t4_rcvd;
1880 	tod->tod_output = t4_tod_output;
1881 	tod->tod_send_rst = t4_send_rst;
1882 	tod->tod_send_fin = t4_send_fin;
1883 	tod->tod_pcb_detach = t4_pcb_detach;
1884 	tod->tod_l2_update = t4_l2_update;
1885 	tod->tod_syncache_added = t4_syncache_added;
1886 	tod->tod_syncache_removed = t4_syncache_removed;
1887 	tod->tod_syncache_respond = t4_syncache_respond;
1888 	tod->tod_offload_socket = t4_offload_socket;
1889 	tod->tod_ctloutput = t4_ctloutput;
1890 	tod->tod_tcp_info = t4_tcp_info;
1891 #ifdef KERN_TLS
1892 	tod->tod_alloc_tls_session = t4_alloc_tls_session;
1893 #endif
1894 	tod->tod_pmtu_update = t4_pmtu_update;
1895 
1896 	for_each_port(sc, i) {
1897 		for_each_vi(sc->port[i], v, vi) {
1898 			SETTOEDEV(vi->ifp, &td->tod);
1899 		}
1900 	}
1901 
1902 	sc->tom_softc = td;
1903 	register_toedev(sc->tom_softc);
1904 
1905 done:
1906 	if (rc != 0)
1907 		free_tom_data(sc, td);
1908 	return (rc);
1909 }
1910 
1911 static int
1912 t4_tom_deactivate(struct adapter *sc)
1913 {
1914 	int rc = 0;
1915 	struct tom_data *td = sc->tom_softc;
1916 
1917 	ASSERT_SYNCHRONIZED_OP(sc);
1918 
1919 	if (td == NULL)
1920 		return (0);	/* XXX. KASSERT? */
1921 
1922 	if (sc->offload_map != 0)
1923 		return (EBUSY);	/* at least one port has IFCAP_TOE enabled */
1924 
1925 	if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1926 		return (EBUSY);	/* both iWARP and iSCSI rely on the TOE. */
1927 
1928 	mtx_lock(&td->toep_list_lock);
1929 	if (!TAILQ_EMPTY(&td->toep_list))
1930 		rc = EBUSY;
1931 	mtx_unlock(&td->toep_list_lock);
1932 
1933 	mtx_lock(&td->lctx_hash_lock);
1934 	if (td->lctx_count > 0)
1935 		rc = EBUSY;
1936 	mtx_unlock(&td->lctx_hash_lock);
1937 
1938 	taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1939 	mtx_lock(&td->unsent_wr_lock);
1940 	if (!STAILQ_EMPTY(&td->unsent_wr_list))
1941 		rc = EBUSY;
1942 	mtx_unlock(&td->unsent_wr_lock);
1943 
1944 	if (rc == 0) {
1945 		unregister_toedev(sc->tom_softc);
1946 		free_tom_data(sc, td);
1947 		sc->tom_softc = NULL;
1948 	}
1949 
1950 	return (rc);
1951 }
1952 
1953 static int
1954 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1955 {
1956 	struct tcpcb *tp = sototcpcb(so);
1957 	struct toepcb *toep = tp->t_toe;
1958 	int error;
1959 
1960 	/*
1961 	 * No lock is needed as TOE sockets never change between
1962 	 * active and passive.
1963 	 */
1964 	if (SOLISTENING(so))
1965 		return (EINVAL);
1966 
1967 	if (ulp_mode(toep) == ULP_MODE_TCPDDP ||
1968 	    ulp_mode(toep) == ULP_MODE_NONE) {
1969 		error = t4_aio_queue_ddp(so, job);
1970 		if (error != EOPNOTSUPP)
1971 			return (error);
1972 	}
1973 
1974 	return (t4_aio_queue_aiotx(so, job));
1975 }
1976 
1977 static int
1978 t4_tom_mod_load(void)
1979 {
1980 	/* CPL handlers */
1981 	t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl);
1982 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2,
1983 	    CPL_COOKIE_TOM);
1984 	t4_init_connect_cpl_handlers();
1985 	t4_init_listen_cpl_handlers();
1986 	t4_init_cpl_io_handlers();
1987 
1988 	t4_ddp_mod_load();
1989 	t4_tls_mod_load();
1990 
1991 	bcopy(&tcp_protosw, &toe_protosw, sizeof(toe_protosw));
1992 	toe_protosw.pr_aio_queue = t4_aio_queue_tom;
1993 
1994 	bcopy(&tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
1995 	toe6_protosw.pr_aio_queue = t4_aio_queue_tom;
1996 
1997 	return (t4_register_uld(&tom_uld_info));
1998 }
1999 
2000 static void
2001 tom_uninit(struct adapter *sc, void *arg __unused)
2002 {
2003 	if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
2004 		return;
2005 
2006 	/* Try to free resources (works only if no port has IFCAP_TOE) */
2007 	if (uld_active(sc, ULD_TOM))
2008 		t4_deactivate_uld(sc, ULD_TOM);
2009 
2010 	end_synchronized_op(sc, 0);
2011 }
2012 
2013 static int
2014 t4_tom_mod_unload(void)
2015 {
2016 	t4_iterate(tom_uninit, NULL);
2017 
2018 	if (t4_unregister_uld(&tom_uld_info) == EBUSY)
2019 		return (EBUSY);
2020 
2021 	t4_tls_mod_unload();
2022 	t4_ddp_mod_unload();
2023 
2024 	t4_uninit_connect_cpl_handlers();
2025 	t4_uninit_listen_cpl_handlers();
2026 	t4_uninit_cpl_io_handlers();
2027 	t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM);
2028 	t4_register_cpl_handler(CPL_GET_TCB_RPL, NULL);
2029 
2030 	return (0);
2031 }
2032 #endif	/* TCP_OFFLOAD */
2033 
2034 static int
2035 t4_tom_modevent(module_t mod, int cmd, void *arg)
2036 {
2037 	int rc = 0;
2038 
2039 #ifdef TCP_OFFLOAD
2040 	switch (cmd) {
2041 	case MOD_LOAD:
2042 		rc = t4_tom_mod_load();
2043 		break;
2044 
2045 	case MOD_UNLOAD:
2046 		rc = t4_tom_mod_unload();
2047 		break;
2048 
2049 	default:
2050 		rc = EINVAL;
2051 	}
2052 #else
2053 	printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
2054 	rc = EOPNOTSUPP;
2055 #endif
2056 	return (rc);
2057 }
2058 
2059 static moduledata_t t4_tom_moddata= {
2060 	"t4_tom",
2061 	t4_tom_modevent,
2062 	0
2063 };
2064 
2065 MODULE_VERSION(t4_tom, 1);
2066 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
2067 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
2068 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
2069