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 static int t4_tom_stop(struct adapter *);
93 static int t4_tom_restart(struct adapter *);
94
95 static struct uld_info tom_uld_info = {
96 .uld_activate = t4_tom_activate,
97 .uld_deactivate = t4_tom_deactivate,
98 .uld_stop = t4_tom_stop,
99 .uld_restart = t4_tom_restart,
100 };
101
102 static void release_offload_resources(struct toepcb *);
103 static void done_with_toepcb(struct toepcb *);
104 static int alloc_tid_tabs(struct adapter *);
105 static void free_tid_tabs(struct adapter *);
106 static void free_tom_data(struct adapter *, struct tom_data *);
107 static void reclaim_wr_resources(void *, int);
108 static void cleanup_stranded_tids(void *, int);
109
110 struct toepcb *
alloc_toepcb(struct vi_info * vi,int flags)111 alloc_toepcb(struct vi_info *vi, int flags)
112 {
113 struct port_info *pi = vi->pi;
114 struct adapter *sc = pi->adapter;
115 struct toepcb *toep;
116 int tx_credits, txsd_total, len;
117
118 /*
119 * The firmware counts tx work request credits in units of 16 bytes
120 * each. Reserve room for an ABORT_REQ so the driver never has to worry
121 * about tx credits if it wants to abort a connection.
122 */
123 tx_credits = sc->params.ofldq_wr_cred;
124 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
125
126 /*
127 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
128 * immediate payload, and firmware counts tx work request credits in
129 * units of 16 byte. Calculate the maximum work requests possible.
130 */
131 txsd_total = tx_credits /
132 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
133
134 len = offsetof(struct toepcb, txsd) +
135 txsd_total * sizeof(struct ofld_tx_sdesc);
136
137 toep = malloc(len, M_CXGBE, M_ZERO | flags);
138 if (toep == NULL)
139 return (NULL);
140
141 refcount_init(&toep->refcount, 1);
142 toep->td = sc->tom_softc;
143 toep->incarnation = sc->incarnation;
144 toep->vi = vi;
145 toep->tid = -1;
146 toep->tx_total = tx_credits;
147 toep->tx_credits = tx_credits;
148 mbufq_init(&toep->ulp_pduq, INT_MAX);
149 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
150 toep->txsd_total = txsd_total;
151 toep->txsd_avail = txsd_total;
152 toep->txsd_pidx = 0;
153 toep->txsd_cidx = 0;
154 aiotx_init_toep(toep);
155
156 return (toep);
157 }
158
159 /*
160 * Initialize a toepcb after its params have been filled out.
161 */
162 int
init_toepcb(struct vi_info * vi,struct toepcb * toep)163 init_toepcb(struct vi_info *vi, struct toepcb *toep)
164 {
165 struct conn_params *cp = &toep->params;
166 struct port_info *pi = vi->pi;
167 struct adapter *sc = pi->adapter;
168 struct tx_cl_rl_params *tc;
169
170 if (cp->tc_idx >= 0 && cp->tc_idx < sc->params.nsched_cls) {
171 tc = &pi->sched_params->cl_rl[cp->tc_idx];
172 mtx_lock(&sc->tc_lock);
173 if (tc->state != CS_HW_CONFIGURED) {
174 CH_ERR(vi, "tid %d cannot be bound to traffic class %d "
175 "because it is not configured (its state is %d)\n",
176 toep->tid, cp->tc_idx, tc->state);
177 cp->tc_idx = -1;
178 } else {
179 tc->refcount++;
180 }
181 mtx_unlock(&sc->tc_lock);
182 }
183 toep->ofld_txq = &sc->sge.ofld_txq[cp->txq_idx];
184 toep->ofld_rxq = &sc->sge.ofld_rxq[cp->rxq_idx];
185 toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
186
187 tls_init_toep(toep);
188 MPASS(ulp_mode(toep) != ULP_MODE_TCPDDP);
189
190 toep->flags |= TPF_INITIALIZED;
191
192 return (0);
193 }
194
195 struct toepcb *
hold_toepcb(struct toepcb * toep)196 hold_toepcb(struct toepcb *toep)
197 {
198
199 refcount_acquire(&toep->refcount);
200 return (toep);
201 }
202
203 void
free_toepcb(struct toepcb * toep)204 free_toepcb(struct toepcb *toep)
205 {
206
207 if (refcount_release(&toep->refcount) == 0)
208 return;
209
210 KASSERT(!(toep->flags & TPF_ATTACHED),
211 ("%s: attached to an inpcb", __func__));
212 KASSERT(!(toep->flags & TPF_CPL_PENDING),
213 ("%s: CPL pending", __func__));
214
215 if (toep->flags & TPF_INITIALIZED) {
216 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
217 ddp_uninit_toep(toep);
218 tls_uninit_toep(toep);
219 }
220 free(toep, M_CXGBE);
221 }
222
223 /*
224 * Set up the socket for TCP offload.
225 */
226 void
offload_socket(struct socket * so,struct toepcb * toep)227 offload_socket(struct socket *so, struct toepcb *toep)
228 {
229 struct tom_data *td = toep->td;
230 struct inpcb *inp = sotoinpcb(so);
231 struct tcpcb *tp = intotcpcb(inp);
232 struct sockbuf *sb;
233
234 INP_WLOCK_ASSERT(inp);
235
236 /* Update socket */
237 sb = &so->so_snd;
238 SOCKBUF_LOCK(sb);
239 sb->sb_flags |= SB_NOCOALESCE;
240 SOCKBUF_UNLOCK(sb);
241 sb = &so->so_rcv;
242 SOCKBUF_LOCK(sb);
243 sb->sb_flags |= SB_NOCOALESCE;
244 if (inp->inp_vflag & INP_IPV6)
245 so->so_proto = &toe6_protosw;
246 else
247 so->so_proto = &toe_protosw;
248 SOCKBUF_UNLOCK(sb);
249
250 /* Update TCP PCB */
251 tp->tod = &td->tod;
252 tp->t_toe = toep;
253 tp->t_flags |= TF_TOE;
254
255 /* Install an extra hold on inp */
256 toep->inp = inp;
257 toep->flags |= TPF_ATTACHED;
258 in_pcbref(inp);
259 }
260
261 void
restore_so_proto(struct socket * so,bool v6)262 restore_so_proto(struct socket *so, bool v6)
263 {
264 if (v6)
265 so->so_proto = &tcp6_protosw;
266 else
267 so->so_proto = &tcp_protosw;
268 }
269
270 /* This is _not_ the normal way to "unoffload" a socket. */
271 void
undo_offload_socket(struct socket * so)272 undo_offload_socket(struct socket *so)
273 {
274 struct inpcb *inp = sotoinpcb(so);
275 struct tcpcb *tp = intotcpcb(inp);
276 struct toepcb *toep = tp->t_toe;
277 struct 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
301 static void
release_offload_resources(struct toepcb * toep)302 release_offload_resources(struct toepcb *toep)
303 {
304 struct tom_data *td = toep->td;
305 struct adapter *sc = td_adapter(td);
306 int tid = toep->tid;
307
308 KASSERT(!(toep->flags & TPF_CPL_PENDING),
309 ("%s: %p has CPL pending.", __func__, toep));
310
311 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
312 __func__, toep, tid, toep->l2te, toep->ce);
313
314 if (toep->l2te) {
315 t4_l2t_release(toep->l2te);
316 toep->l2te = NULL;
317 }
318 if (tid >= 0) {
319 remove_tid(sc, tid, toep->ce ? 2 : 1);
320 release_tid(sc, tid, toep->ctrlq);
321 toep->tid = -1;
322 mtx_lock(&td->toep_list_lock);
323 if (toep->flags & TPF_IN_TOEP_LIST) {
324 toep->flags &= ~TPF_IN_TOEP_LIST;
325 TAILQ_REMOVE(&td->toep_list, toep, link);
326 }
327 mtx_unlock(&td->toep_list_lock);
328 }
329 if (toep->ce) {
330 t4_release_clip_entry(sc, toep->ce);
331 toep->ce = NULL;
332 }
333 if (toep->params.tc_idx != -1)
334 t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx);
335 }
336
337 /*
338 * Both the driver and kernel are done with the toepcb.
339 */
340 static void
done_with_toepcb(struct toepcb * toep)341 done_with_toepcb(struct toepcb *toep)
342 {
343 KASSERT(!(toep->flags & TPF_CPL_PENDING),
344 ("%s: %p has CPL pending.", __func__, toep));
345 KASSERT(!(toep->flags & TPF_ATTACHED),
346 ("%s: %p is still attached.", __func__, toep));
347
348 CTR(KTR_CXGBE, "%s: toep %p (0x%x)", __func__, toep, toep->flags);
349
350 /*
351 * These queues should have been emptied at approximately the same time
352 * that a normal connection's socket's so_snd would have been purged or
353 * drained. Do _not_ clean up here.
354 */
355 MPASS(mbufq_empty(&toep->ulp_pduq));
356 MPASS(mbufq_empty(&toep->ulp_pdu_reclaimq));
357 #ifdef INVARIANTS
358 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
359 ddp_assert_empty(toep);
360 #endif
361 MPASS(TAILQ_EMPTY(&toep->aiotx_jobq));
362 MPASS(toep->tid == -1);
363 MPASS(toep->l2te == NULL);
364 MPASS(toep->ce == NULL);
365 MPASS((toep->flags & TPF_IN_TOEP_LIST) == 0);
366
367 free_toepcb(toep);
368 }
369
370 /*
371 * The kernel is done with the TCP PCB and this is our opportunity to unhook the
372 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no
373 * pending CPL) then it is time to release all resources tied to the toepcb.
374 *
375 * Also gets called when an offloaded active open fails and the TOM wants the
376 * kernel to take the TCP PCB back.
377 */
378 void
t4_pcb_detach(struct toedev * tod __unused,struct tcpcb * tp)379 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
380 {
381 #if defined(KTR) || defined(INVARIANTS)
382 struct inpcb *inp = tptoinpcb(tp);
383 #endif
384 struct toepcb *toep = tp->t_toe;
385
386 INP_WLOCK_ASSERT(inp);
387
388 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
389 KASSERT(toep->flags & TPF_ATTACHED,
390 ("%s: not attached", __func__));
391
392 #ifdef KTR
393 if (tp->t_state == TCPS_SYN_SENT) {
394 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
395 __func__, toep->tid, toep, toep->flags, inp,
396 inp->inp_flags);
397 } else {
398 CTR6(KTR_CXGBE,
399 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
400 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
401 inp->inp_flags);
402 }
403 #endif
404
405 tp->tod = NULL;
406 tp->t_toe = NULL;
407 tp->t_flags &= ~TF_TOE;
408 toep->flags &= ~TPF_ATTACHED;
409
410 if (!(toep->flags & TPF_CPL_PENDING))
411 done_with_toepcb(toep);
412 }
413
414 /*
415 * setsockopt handler.
416 */
417 static void
t4_ctloutput(struct toedev * tod,struct tcpcb * tp,int dir,int name)418 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
419 {
420 struct adapter *sc = tod->tod_softc;
421 struct toepcb *toep = tp->t_toe;
422
423 if (dir == SOPT_GET)
424 return;
425
426 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
427
428 switch (name) {
429 case TCP_NODELAY:
430 if (tp->t_state != TCPS_ESTABLISHED)
431 break;
432 toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
433 t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
434 V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0);
435 break;
436 default:
437 break;
438 }
439 }
440
441 static inline uint64_t
get_tcb_tflags(const uint64_t * tcb)442 get_tcb_tflags(const uint64_t *tcb)
443 {
444
445 return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32));
446 }
447
448 static inline uint32_t
get_tcb_field(const uint64_t * tcb,u_int word,uint32_t mask,u_int shift)449 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift)
450 {
451 #define LAST_WORD ((TCB_SIZE / 4) - 1)
452 uint64_t t1, t2;
453 int flit_idx;
454
455 MPASS(mask != 0);
456 MPASS(word <= LAST_WORD);
457 MPASS(shift < 32);
458
459 flit_idx = (LAST_WORD - word) / 2;
460 if (word & 0x1)
461 shift += 32;
462 t1 = be64toh(tcb[flit_idx]) >> shift;
463 t2 = 0;
464 if (fls(mask) > 64 - shift) {
465 /*
466 * Will spill over into the next logical flit, which is the flit
467 * before this one. The flit_idx before this one must be valid.
468 */
469 MPASS(flit_idx > 0);
470 t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift);
471 }
472 return ((t2 | t1) & mask);
473 #undef LAST_WORD
474 }
475 #define GET_TCB_FIELD(tcb, F) \
476 get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F)
477
478 /*
479 * Issues a CPL_GET_TCB to read the entire TCB for the tid.
480 */
481 static int
send_get_tcb(struct adapter * sc,u_int tid)482 send_get_tcb(struct adapter *sc, u_int tid)
483 {
484 struct cpl_get_tcb *cpl;
485 struct wrq_cookie cookie;
486
487 MPASS(tid >= sc->tids.tid_base);
488 MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
489
490 cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16),
491 &cookie);
492 if (__predict_false(cpl == NULL))
493 return (ENOMEM);
494 bzero(cpl, sizeof(*cpl));
495 INIT_TP_WR(cpl, tid);
496 OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid));
497 cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) |
498 V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id));
499 cpl->cookie = 0xff;
500 commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie);
501
502 return (0);
503 }
504
505 static struct tcb_histent *
alloc_tcb_histent(struct adapter * sc,u_int tid,int flags)506 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags)
507 {
508 struct tcb_histent *te;
509
510 MPASS(flags == M_NOWAIT || flags == M_WAITOK);
511
512 te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags);
513 if (te == NULL)
514 return (NULL);
515 mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF);
516 callout_init_mtx(&te->te_callout, &te->te_lock, 0);
517 te->te_adapter = sc;
518 te->te_tid = tid;
519
520 return (te);
521 }
522
523 static void
free_tcb_histent(struct tcb_histent * te)524 free_tcb_histent(struct tcb_histent *te)
525 {
526
527 mtx_destroy(&te->te_lock);
528 free(te, M_CXGBE);
529 }
530
531 /*
532 * Start tracking the tid in the TCB history.
533 */
534 int
add_tid_to_history(struct adapter * sc,u_int tid)535 add_tid_to_history(struct adapter *sc, u_int tid)
536 {
537 struct tcb_histent *te = NULL;
538 struct tom_data *td = sc->tom_softc;
539 int rc;
540
541 MPASS(tid >= sc->tids.tid_base);
542 MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
543
544 if (td->tcb_history == NULL)
545 return (ENXIO);
546
547 rw_wlock(&td->tcb_history_lock);
548 if (td->tcb_history[tid] != NULL) {
549 rc = EEXIST;
550 goto done;
551 }
552 te = alloc_tcb_histent(sc, tid, M_NOWAIT);
553 if (te == NULL) {
554 rc = ENOMEM;
555 goto done;
556 }
557 mtx_lock(&te->te_lock);
558 rc = send_get_tcb(sc, tid);
559 if (rc == 0) {
560 te->te_flags |= TE_RPL_PENDING;
561 td->tcb_history[tid] = te;
562 } else {
563 free(te, M_CXGBE);
564 }
565 mtx_unlock(&te->te_lock);
566 done:
567 rw_wunlock(&td->tcb_history_lock);
568 return (rc);
569 }
570
571 static void
remove_tcb_histent(struct tcb_histent * te)572 remove_tcb_histent(struct tcb_histent *te)
573 {
574 struct adapter *sc = te->te_adapter;
575 struct tom_data *td = sc->tom_softc;
576
577 rw_assert(&td->tcb_history_lock, RA_WLOCKED);
578 mtx_assert(&te->te_lock, MA_OWNED);
579 MPASS(td->tcb_history[te->te_tid] == te);
580
581 td->tcb_history[te->te_tid] = NULL;
582 free_tcb_histent(te);
583 rw_wunlock(&td->tcb_history_lock);
584 }
585
586 static inline struct tcb_histent *
lookup_tcb_histent(struct adapter * sc,u_int tid,bool addrem)587 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem)
588 {
589 struct tcb_histent *te;
590 struct tom_data *td = sc->tom_softc;
591
592 MPASS(tid >= sc->tids.tid_base);
593 MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
594
595 if (td->tcb_history == NULL)
596 return (NULL);
597
598 if (addrem)
599 rw_wlock(&td->tcb_history_lock);
600 else
601 rw_rlock(&td->tcb_history_lock);
602 te = td->tcb_history[tid];
603 if (te != NULL) {
604 mtx_lock(&te->te_lock);
605 return (te); /* with both locks held */
606 }
607 if (addrem)
608 rw_wunlock(&td->tcb_history_lock);
609 else
610 rw_runlock(&td->tcb_history_lock);
611
612 return (te);
613 }
614
615 static inline void
release_tcb_histent(struct tcb_histent * te)616 release_tcb_histent(struct tcb_histent *te)
617 {
618 struct adapter *sc = te->te_adapter;
619 struct tom_data *td = sc->tom_softc;
620
621 mtx_assert(&te->te_lock, MA_OWNED);
622 mtx_unlock(&te->te_lock);
623 rw_assert(&td->tcb_history_lock, RA_RLOCKED);
624 rw_runlock(&td->tcb_history_lock);
625 }
626
627 static void
request_tcb(void * arg)628 request_tcb(void *arg)
629 {
630 struct tcb_histent *te = arg;
631
632 mtx_assert(&te->te_lock, MA_OWNED);
633
634 /* Noone else is supposed to update the histent. */
635 MPASS(!(te->te_flags & TE_RPL_PENDING));
636 if (send_get_tcb(te->te_adapter, te->te_tid) == 0)
637 te->te_flags |= TE_RPL_PENDING;
638 else
639 callout_schedule(&te->te_callout, hz / 100);
640 }
641
642 static void
update_tcb_histent(struct tcb_histent * te,const uint64_t * tcb)643 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb)
644 {
645 struct tom_data *td = te->te_adapter->tom_softc;
646 uint64_t tflags = get_tcb_tflags(tcb);
647 uint8_t sample = 0;
648
649 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) {
650 if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0)
651 sample |= TS_RTO;
652 if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0)
653 sample |= TS_DUPACKS;
654 if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold)
655 sample |= TS_FASTREXMT;
656 }
657
658 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) {
659 uint32_t snd_wnd;
660
661 sample |= TS_SND_BACKLOGGED; /* for whatever reason. */
662
663 snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
664 if (tflags & V_TF_RECV_SCALE(1))
665 snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE);
666 if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd)
667 sample |= TS_CWND_LIMITED; /* maybe due to CWND */
668 }
669
670 if (tflags & V_TF_CCTRL_ECN(1)) {
671
672 /*
673 * CE marker on incoming IP hdr, echoing ECE back in the TCP
674 * hdr. Indicates congestion somewhere on the way from the peer
675 * to this node.
676 */
677 if (tflags & V_TF_CCTRL_ECE(1))
678 sample |= TS_ECN_ECE;
679
680 /*
681 * ECE seen and CWR sent (or about to be sent). Might indicate
682 * congestion on the way to the peer. This node is reducing its
683 * congestion window in response.
684 */
685 if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1)))
686 sample |= TS_ECN_CWR;
687 }
688
689 te->te_sample[te->te_pidx] = sample;
690 if (++te->te_pidx == nitems(te->te_sample))
691 te->te_pidx = 0;
692 memcpy(te->te_tcb, tcb, TCB_SIZE);
693 te->te_flags |= TE_ACTIVE;
694 }
695
696 static int
do_get_tcb_rpl(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)697 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
698 {
699 struct adapter *sc = iq->adapter;
700 const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *);
701 const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1);
702 struct tcb_histent *te;
703 const u_int tid = GET_TID(cpl);
704 bool remove;
705
706 remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED;
707 te = lookup_tcb_histent(sc, tid, remove);
708 if (te == NULL) {
709 /* Not in the history. Who issued the GET_TCB for this? */
710 device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, "
711 "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid,
712 (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE),
713 GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE),
714 GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie);
715 goto done;
716 }
717
718 MPASS(te->te_flags & TE_RPL_PENDING);
719 te->te_flags &= ~TE_RPL_PENDING;
720 if (remove) {
721 remove_tcb_histent(te);
722 } else {
723 update_tcb_histent(te, tcb);
724 callout_reset(&te->te_callout, hz / 10, request_tcb, te);
725 release_tcb_histent(te);
726 }
727 done:
728 m_freem(m);
729 return (0);
730 }
731
732 static void
fill_tcp_info_from_tcb(struct adapter * sc,uint64_t * tcb,struct tcp_info * ti)733 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti)
734 {
735 uint32_t v;
736
737 ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE);
738
739 v = GET_TCB_FIELD(tcb, T_SRTT);
740 ti->tcpi_rtt = tcp_ticks_to_us(sc, v);
741
742 v = GET_TCB_FIELD(tcb, T_RTTVAR);
743 ti->tcpi_rttvar = tcp_ticks_to_us(sc, v);
744
745 ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH);
746 ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND);
747 ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT);
748 ti->tcpi_rcv_adv = GET_TCB_FIELD(tcb, RCV_ADV);
749 ti->tcpi_dupacks = GET_TCB_FIELD(tcb, T_DUPACKS);
750
751 v = GET_TCB_FIELD(tcb, TX_MAX);
752 ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW);
753 ti->tcpi_snd_una = v - GET_TCB_FIELD(tcb, SND_UNA_RAW);
754 ti->tcpi_snd_max = v - GET_TCB_FIELD(tcb, SND_MAX_RAW);
755
756 /* Receive window being advertised by us. */
757 ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE); /* Yes, SND. */
758 ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND);
759
760 /* Send window */
761 ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE); /* Yes, RCV. */
762 ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
763 if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1))
764 ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale;
765 else
766 ti->tcpi_snd_wscale = 0;
767
768 }
769
770 static void
fill_tcp_info_from_history(struct adapter * sc,struct tcb_histent * te,struct tcp_info * ti)771 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te,
772 struct tcp_info *ti)
773 {
774
775 fill_tcp_info_from_tcb(sc, te->te_tcb, ti);
776 }
777
778 /*
779 * Reads the TCB for the given tid using a memory window and copies it to 'buf'
780 * in the same format as CPL_GET_TCB_RPL.
781 */
782 static void
read_tcb_using_memwin(struct adapter * sc,u_int tid,uint64_t * buf)783 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf)
784 {
785 int i, j, k, rc;
786 uint32_t addr;
787 u_char *tcb, tmp;
788
789 MPASS(tid >= sc->tids.tid_base);
790 MPASS(tid - sc->tids.tid_base < sc->tids.ntids);
791
792 addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE;
793 rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE);
794 if (rc != 0)
795 return;
796
797 tcb = (u_char *)buf;
798 for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) {
799 for (k = 0; k < 16; k++) {
800 tmp = tcb[i + k];
801 tcb[i + k] = tcb[j + k];
802 tcb[j + k] = tmp;
803 }
804 }
805 }
806
807 static void
fill_tcp_info(struct adapter * sc,u_int tid,struct tcp_info * ti)808 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti)
809 {
810 uint64_t tcb[TCB_SIZE / sizeof(uint64_t)];
811 struct tcb_histent *te;
812
813 ti->tcpi_toe_tid = tid;
814 te = lookup_tcb_histent(sc, tid, false);
815 if (te != NULL) {
816 fill_tcp_info_from_history(sc, te, ti);
817 release_tcb_histent(te);
818 } else {
819 if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) {
820 /* XXX: tell firmware to flush TCB cache. */
821 }
822 read_tcb_using_memwin(sc, tid, tcb);
823 fill_tcp_info_from_tcb(sc, tcb, ti);
824 }
825 }
826
827 /*
828 * Called by the kernel to allow the TOE driver to "refine" values filled up in
829 * the tcp_info for an offloaded connection.
830 */
831 static void
t4_tcp_info(struct toedev * tod,const struct tcpcb * tp,struct tcp_info * ti)832 t4_tcp_info(struct toedev *tod, const struct tcpcb *tp, struct tcp_info *ti)
833 {
834 struct adapter *sc = tod->tod_softc;
835 struct toepcb *toep = tp->t_toe;
836
837 INP_LOCK_ASSERT(tptoinpcb(tp));
838 MPASS(ti != NULL);
839
840 fill_tcp_info(sc, toep->tid, ti);
841 }
842
843 #ifdef KERN_TLS
844 static int
t4_alloc_tls_session(struct toedev * tod,struct tcpcb * tp,struct ktls_session * tls,int direction)845 t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp,
846 struct ktls_session *tls, int direction)
847 {
848 struct toepcb *toep = tp->t_toe;
849
850 INP_WLOCK_ASSERT(tptoinpcb(tp));
851 MPASS(tls != NULL);
852
853 return (tls_alloc_ktls(toep, tls, direction));
854 }
855 #endif
856
857 static void
send_mss_flowc_wr(struct adapter * sc,struct toepcb * toep)858 send_mss_flowc_wr(struct adapter *sc, struct toepcb *toep)
859 {
860 struct wrq_cookie cookie;
861 struct fw_flowc_wr *flowc;
862 struct ofld_tx_sdesc *txsd;
863 const int flowclen = sizeof(*flowc) + sizeof(struct fw_flowc_mnemval);
864 const int flowclen16 = howmany(flowclen, 16);
865
866 if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0) {
867 CH_ERR(sc, "%s: tid %u out of tx credits (%d, %d).\n", __func__,
868 toep->tid, toep->tx_credits, toep->txsd_avail);
869 return;
870 }
871
872 flowc = start_wrq_wr(&toep->ofld_txq->wrq, flowclen16, &cookie);
873 if (__predict_false(flowc == NULL)) {
874 CH_ERR(sc, "ENOMEM in %s for tid %u.\n", __func__, toep->tid);
875 return;
876 }
877 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) |
878 V_FW_FLOWC_WR_NPARAMS(1));
879 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) |
880 V_FW_WR_FLOWID(toep->tid));
881 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_MSS;
882 flowc->mnemval[0].val = htobe32(toep->params.emss);
883
884 txsd = &toep->txsd[toep->txsd_pidx];
885 txsd->tx_credits = flowclen16;
886 txsd->plen = 0;
887 toep->tx_credits -= txsd->tx_credits;
888 if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
889 toep->txsd_pidx = 0;
890 toep->txsd_avail--;
891 commit_wrq_wr(&toep->ofld_txq->wrq, flowc, &cookie);
892 }
893
894 static void
t4_pmtu_update(struct toedev * tod,struct tcpcb * tp,tcp_seq seq,int mtu)895 t4_pmtu_update(struct toedev *tod, struct tcpcb *tp, tcp_seq seq, int mtu)
896 {
897 struct work_request_hdr *wrh;
898 struct ulp_txpkt *ulpmc;
899 int idx, len;
900 struct wrq_cookie cookie;
901 struct inpcb *inp = tptoinpcb(tp);
902 struct toepcb *toep = tp->t_toe;
903 struct adapter *sc = td_adapter(toep->td);
904 unsigned short *mtus = &sc->params.mtus[0];
905
906 INP_WLOCK_ASSERT(inp);
907 MPASS(mtu > 0); /* kernel is supposed to provide something usable. */
908
909 /* tp->snd_una and snd_max are in host byte order too. */
910 seq = be32toh(seq);
911
912 CTR6(KTR_CXGBE, "%s: tid %d, seq 0x%08x, mtu %u, mtu_idx %u (%d)",
913 __func__, toep->tid, seq, mtu, toep->params.mtu_idx,
914 mtus[toep->params.mtu_idx]);
915
916 if (ulp_mode(toep) == ULP_MODE_NONE && /* XXX: Read TCB otherwise? */
917 (SEQ_LT(seq, tp->snd_una) || SEQ_GEQ(seq, tp->snd_max))) {
918 CTR5(KTR_CXGBE,
919 "%s: tid %d, seq 0x%08x not in range [0x%08x, 0x%08x).",
920 __func__, toep->tid, seq, tp->snd_una, tp->snd_max);
921 return;
922 }
923
924 /* Find the best mtu_idx for the suggested MTU. */
925 for (idx = 0; idx < NMTUS - 1 && mtus[idx + 1] <= mtu; idx++)
926 continue;
927 if (idx >= toep->params.mtu_idx)
928 return; /* Never increase the PMTU (just like the kernel). */
929
930 /*
931 * We'll send a compound work request with 2 SET_TCB_FIELDs -- the first
932 * one updates the mtu_idx and the second one triggers a retransmit.
933 */
934 len = sizeof(*wrh) + 2 * roundup2(LEN__SET_TCB_FIELD_ULP, 16);
935 wrh = start_wrq_wr(toep->ctrlq, howmany(len, 16), &cookie);
936 if (wrh == NULL) {
937 CH_ERR(sc, "failed to change mtu_idx of tid %d (%u -> %u).\n",
938 toep->tid, toep->params.mtu_idx, idx);
939 return;
940 }
941 INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */
942 ulpmc = (struct ulp_txpkt *)(wrh + 1);
943 ulpmc = mk_set_tcb_field_ulp(sc, ulpmc, toep->tid, W_TCB_T_MAXSEG,
944 V_TCB_T_MAXSEG(M_TCB_T_MAXSEG), V_TCB_T_MAXSEG(idx));
945 ulpmc = mk_set_tcb_field_ulp(sc, ulpmc, toep->tid, W_TCB_TIMESTAMP,
946 V_TCB_TIMESTAMP(0x7FFFFULL << 11), 0);
947 commit_wrq_wr(toep->ctrlq, wrh, &cookie);
948
949 /* Update the software toepcb and tcpcb. */
950 toep->params.mtu_idx = idx;
951 tp->t_maxseg = mtus[toep->params.mtu_idx];
952 if (inp->inp_inc.inc_flags & INC_ISIPV6)
953 tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
954 else
955 tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr);
956 toep->params.emss = tp->t_maxseg;
957 if (tp->t_flags & TF_RCVD_TSTMP)
958 toep->params.emss -= TCPOLEN_TSTAMP_APPA;
959
960 /* Update the firmware flowc. */
961 send_mss_flowc_wr(sc, toep);
962
963 /* Update the MTU in the kernel's hostcache. */
964 if (sc->tt.update_hc_on_pmtu_change != 0) {
965 struct in_conninfo inc = {0};
966
967 inc.inc_fibnum = inp->inp_inc.inc_fibnum;
968 if (inp->inp_inc.inc_flags & INC_ISIPV6) {
969 inc.inc_flags |= INC_ISIPV6;
970 inc.inc6_faddr = inp->inp_inc.inc6_faddr;
971 } else {
972 inc.inc_faddr = inp->inp_inc.inc_faddr;
973 }
974 tcp_hc_updatemtu(&inc, mtu);
975 }
976
977 CTR6(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u), t_maxseg %u, emss %u",
978 __func__, toep->tid, toep->params.mtu_idx,
979 mtus[toep->params.mtu_idx], tp->t_maxseg, toep->params.emss);
980 }
981
982 /*
983 * The TOE driver will not receive any more CPLs for the tid associated with the
984 * toepcb; release the hold on the inpcb.
985 */
986 void
final_cpl_received(struct toepcb * toep)987 final_cpl_received(struct toepcb *toep)
988 {
989 struct inpcb *inp = toep->inp;
990 bool need_wakeup;
991
992 KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
993 INP_WLOCK_ASSERT(inp);
994 KASSERT(toep->flags & TPF_CPL_PENDING,
995 ("%s: CPL not pending already?", __func__));
996
997 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
998 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
999
1000 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
1001 release_ddp_resources(toep);
1002 toep->inp = NULL;
1003 need_wakeup = (toep->flags & TPF_WAITING_FOR_FINAL) != 0;
1004 toep->flags &= ~(TPF_CPL_PENDING | TPF_WAITING_FOR_FINAL);
1005 mbufq_drain(&toep->ulp_pduq);
1006 mbufq_drain(&toep->ulp_pdu_reclaimq);
1007 release_offload_resources(toep);
1008 if (!(toep->flags & TPF_ATTACHED))
1009 done_with_toepcb(toep);
1010
1011 if (!in_pcbrele_wlocked(inp))
1012 INP_WUNLOCK(inp);
1013
1014 if (need_wakeup) {
1015 struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep);
1016
1017 mtx_lock(lock);
1018 wakeup(toep);
1019 mtx_unlock(lock);
1020 }
1021 }
1022
1023 void
insert_tid(struct adapter * sc,int tid,void * ctx,int ntids)1024 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
1025 {
1026 struct tid_info *t = &sc->tids;
1027
1028 MPASS(tid >= t->tid_base);
1029 MPASS(tid - t->tid_base < t->ntids);
1030
1031 t->tid_tab[tid - t->tid_base] = ctx;
1032 atomic_add_int(&t->tids_in_use, ntids);
1033 }
1034
1035 void *
lookup_tid(struct adapter * sc,int tid)1036 lookup_tid(struct adapter *sc, int tid)
1037 {
1038 struct tid_info *t = &sc->tids;
1039
1040 return (t->tid_tab[tid - t->tid_base]);
1041 }
1042
1043 void
update_tid(struct adapter * sc,int tid,void * ctx)1044 update_tid(struct adapter *sc, int tid, void *ctx)
1045 {
1046 struct tid_info *t = &sc->tids;
1047
1048 t->tid_tab[tid - t->tid_base] = ctx;
1049 }
1050
1051 void
remove_tid(struct adapter * sc,int tid,int ntids)1052 remove_tid(struct adapter *sc, int tid, int ntids)
1053 {
1054 struct tid_info *t = &sc->tids;
1055
1056 t->tid_tab[tid - t->tid_base] = NULL;
1057 atomic_subtract_int(&t->tids_in_use, ntids);
1058 }
1059
1060 /*
1061 * What mtu_idx to use, given a 4-tuple. Note that both s->mss and tcp_mssopt
1062 * have the MSS that we should advertise in our SYN. Advertised MSS doesn't
1063 * account for any TCP options so the effective MSS (only payload, no headers or
1064 * options) could be different.
1065 */
1066 static int
find_best_mtu_idx(struct adapter * sc,struct in_conninfo * inc,struct offload_settings * s)1067 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc,
1068 struct offload_settings *s)
1069 {
1070 unsigned short *mtus = &sc->params.mtus[0];
1071 int i, mss, mtu;
1072
1073 MPASS(inc != NULL);
1074
1075 mss = s->mss > 0 ? s->mss : tcp_mssopt(inc);
1076 if (inc->inc_flags & INC_ISIPV6)
1077 mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
1078 else
1079 mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr);
1080
1081 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++)
1082 continue;
1083
1084 return (i);
1085 }
1086
1087 /*
1088 * Determine the receive window size for a socket.
1089 */
1090 u_long
select_rcv_wnd(struct socket * so)1091 select_rcv_wnd(struct socket *so)
1092 {
1093 unsigned long wnd;
1094
1095 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
1096
1097 wnd = sbspace(&so->so_rcv);
1098 if (wnd < MIN_RCV_WND)
1099 wnd = MIN_RCV_WND;
1100
1101 return min(wnd, MAX_RCV_WND);
1102 }
1103
1104 int
select_rcv_wscale(void)1105 select_rcv_wscale(void)
1106 {
1107 int wscale = 0;
1108 unsigned long space = sb_max;
1109
1110 if (space > MAX_RCV_WND)
1111 space = MAX_RCV_WND;
1112
1113 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
1114 wscale++;
1115
1116 return (wscale);
1117 }
1118
1119 __be64
calc_options0(struct vi_info * vi,struct conn_params * cp)1120 calc_options0(struct vi_info *vi, struct conn_params *cp)
1121 {
1122 uint64_t opt0 = 0;
1123
1124 opt0 |= F_TCAM_BYPASS;
1125
1126 MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE);
1127 opt0 |= V_WND_SCALE(cp->wscale);
1128
1129 MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS);
1130 opt0 |= V_MSS_IDX(cp->mtu_idx);
1131
1132 MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE);
1133 opt0 |= V_ULP_MODE(cp->ulp_mode);
1134
1135 MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ);
1136 opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize);
1137
1138 MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->adapter->vres.l2t.size);
1139 opt0 |= V_L2T_IDX(cp->l2t_idx);
1140
1141 opt0 |= V_SMAC_SEL(vi->smt_idx);
1142 opt0 |= V_TX_CHAN(vi->pi->tx_chan);
1143
1144 MPASS(cp->keepalive == 0 || cp->keepalive == 1);
1145 opt0 |= V_KEEP_ALIVE(cp->keepalive);
1146
1147 MPASS(cp->nagle == 0 || cp->nagle == 1);
1148 opt0 |= V_NAGLE(cp->nagle);
1149
1150 return (htobe64(opt0));
1151 }
1152
1153 __be32
calc_options2(struct vi_info * vi,struct conn_params * cp)1154 calc_options2(struct vi_info *vi, struct conn_params *cp)
1155 {
1156 uint32_t opt2 = 0;
1157 struct port_info *pi = vi->pi;
1158 struct adapter *sc = pi->adapter;
1159
1160 /*
1161 * rx flow control, rx coalesce, congestion control, and tx pace are all
1162 * explicitly set by the driver. On T5+ the ISS is also set by the
1163 * driver to the value picked by the kernel.
1164 */
1165 if (is_t4(sc)) {
1166 opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID;
1167 opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID;
1168 } else {
1169 opt2 |= F_T5_OPT_2_VALID; /* all 4 valid */
1170 opt2 |= F_T5_ISS; /* ISS provided in CPL */
1171 }
1172
1173 MPASS(cp->sack == 0 || cp->sack == 1);
1174 opt2 |= V_SACK_EN(cp->sack);
1175
1176 MPASS(cp->tstamp == 0 || cp->tstamp == 1);
1177 opt2 |= V_TSTAMPS_EN(cp->tstamp);
1178
1179 if (cp->wscale > 0)
1180 opt2 |= F_WND_SCALE_EN;
1181
1182 MPASS(cp->ecn == 0 || cp->ecn == 1);
1183 opt2 |= V_CCTRL_ECN(cp->ecn);
1184
1185 opt2 |= V_TX_QUEUE(TX_MODQ(pi->tx_chan));
1186 opt2 |= V_PACE(0);
1187 opt2 |= F_RSS_QUEUE_VALID;
1188 opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id);
1189 if (chip_id(sc) <= CHELSIO_T6) {
1190 MPASS(pi->rx_chan == 0 || pi->rx_chan == 1);
1191 opt2 |= V_RX_CHANNEL(pi->rx_chan);
1192 }
1193
1194 MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL);
1195 opt2 |= V_CONG_CNTRL(cp->cong_algo);
1196
1197 MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1);
1198 if (cp->rx_coalesce == 1)
1199 opt2 |= V_RX_COALESCE(M_RX_COALESCE);
1200
1201 opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
1202 MPASS(cp->ulp_mode != ULP_MODE_TCPDDP);
1203
1204 return (htobe32(opt2));
1205 }
1206
1207 uint64_t
select_ntuple(struct vi_info * vi,struct l2t_entry * e)1208 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
1209 {
1210 struct adapter *sc = vi->adapter;
1211 struct tp_params *tp = &sc->params.tp;
1212 uint64_t ntuple = 0;
1213
1214 /*
1215 * Initialize each of the fields which we care about which are present
1216 * in the Compressed Filter Tuple.
1217 */
1218 if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE)
1219 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
1220
1221 if (tp->port_shift >= 0)
1222 ntuple |= (uint64_t)e->lport << tp->port_shift;
1223
1224 if (tp->protocol_shift >= 0)
1225 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
1226
1227 if (tp->vnic_shift >= 0 && tp->vnic_mode == FW_VNIC_MODE_PF_VF) {
1228 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) |
1229 V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) <<
1230 tp->vnic_shift;
1231 }
1232
1233 if (is_t4(sc))
1234 return (htobe32((uint32_t)ntuple));
1235 else
1236 return (htobe64(V_FILTER_TUPLE(ntuple)));
1237 }
1238
1239 /*
1240 * Initialize various connection parameters.
1241 */
1242 void
init_conn_params(struct vi_info * vi,struct offload_settings * s,struct in_conninfo * inc,struct socket * so,const struct tcp_options * tcpopt,int16_t l2t_idx,struct conn_params * cp)1243 init_conn_params(struct vi_info *vi , struct offload_settings *s,
1244 struct in_conninfo *inc, struct socket *so,
1245 const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp)
1246 {
1247 struct port_info *pi = vi->pi;
1248 struct adapter *sc = pi->adapter;
1249 struct tom_tunables *tt = &sc->tt;
1250 struct inpcb *inp = sotoinpcb(so);
1251 struct tcpcb *tp = intotcpcb(inp);
1252 u_long wnd;
1253 u_int q_idx;
1254
1255 MPASS(s->offload != 0);
1256
1257 /* Congestion control algorithm */
1258 if (s->cong_algo >= 0)
1259 cp->cong_algo = s->cong_algo & M_CONG_CNTRL;
1260 else if (sc->tt.cong_algorithm >= 0)
1261 cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL;
1262 else {
1263 struct cc_algo *cc = CC_ALGO(tp);
1264
1265 if (strcasecmp(cc->name, "reno") == 0)
1266 cp->cong_algo = CONG_ALG_RENO;
1267 else if (strcasecmp(cc->name, "tahoe") == 0)
1268 cp->cong_algo = CONG_ALG_TAHOE;
1269 if (strcasecmp(cc->name, "newreno") == 0)
1270 cp->cong_algo = CONG_ALG_NEWRENO;
1271 if (strcasecmp(cc->name, "highspeed") == 0)
1272 cp->cong_algo = CONG_ALG_HIGHSPEED;
1273 else {
1274 /*
1275 * Use newreno in case the algorithm selected by the
1276 * host stack is not supported by the hardware.
1277 */
1278 cp->cong_algo = CONG_ALG_NEWRENO;
1279 }
1280 }
1281
1282 /* Tx traffic scheduling class. */
1283 if (s->sched_class >= 0 && s->sched_class < sc->params.nsched_cls)
1284 cp->tc_idx = s->sched_class;
1285 else
1286 cp->tc_idx = -1;
1287
1288 /* Nagle's algorithm. */
1289 if (s->nagle >= 0)
1290 cp->nagle = s->nagle > 0 ? 1 : 0;
1291 else
1292 cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
1293
1294 /* TCP Keepalive. */
1295 if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE)
1296 cp->keepalive = 1;
1297 else
1298 cp->keepalive = 0;
1299
1300 /* Optimization that's specific to T5 @ 40G. */
1301 if (tt->tx_align >= 0)
1302 cp->tx_align = tt->tx_align > 0 ? 1 : 0;
1303 else if (chip_id(sc) == CHELSIO_T5 &&
1304 (port_top_speed(pi) > 10 || sc->params.nports > 2))
1305 cp->tx_align = 1;
1306 else
1307 cp->tx_align = 0;
1308
1309 /* ULP mode. */
1310 cp->ulp_mode = ULP_MODE_NONE;
1311
1312 /* Rx coalescing. */
1313 if (s->rx_coalesce >= 0)
1314 cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0;
1315 else if (tt->rx_coalesce >= 0)
1316 cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0;
1317 else
1318 cp->rx_coalesce = 1; /* default */
1319
1320 /*
1321 * Index in the PMTU table. This controls the MSS that we announce in
1322 * our SYN initially, but after ESTABLISHED it controls the MSS that we
1323 * use to send data.
1324 */
1325 cp->mtu_idx = find_best_mtu_idx(sc, inc, s);
1326
1327 /* Tx queue for this connection. */
1328 if (s->txq == QUEUE_RANDOM)
1329 q_idx = arc4random();
1330 else if (s->txq == QUEUE_ROUNDROBIN)
1331 q_idx = atomic_fetchadd_int(&vi->txq_rr, 1);
1332 else
1333 q_idx = s->txq;
1334 cp->txq_idx = vi->first_ofld_txq + q_idx % vi->nofldtxq;
1335
1336 /* Rx queue for this connection. */
1337 if (s->rxq == QUEUE_RANDOM)
1338 q_idx = arc4random();
1339 else if (s->rxq == QUEUE_ROUNDROBIN)
1340 q_idx = atomic_fetchadd_int(&vi->rxq_rr, 1);
1341 else
1342 q_idx = s->rxq;
1343 cp->rxq_idx = vi->first_ofld_rxq + q_idx % vi->nofldrxq;
1344
1345 if (SOLISTENING(so)) {
1346 /* Passive open */
1347 MPASS(tcpopt != NULL);
1348
1349 /* TCP timestamp option */
1350 if (tcpopt->tstamp &&
1351 (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323)))
1352 cp->tstamp = 1;
1353 else
1354 cp->tstamp = 0;
1355
1356 /* SACK */
1357 if (tcpopt->sack &&
1358 (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack)))
1359 cp->sack = 1;
1360 else
1361 cp->sack = 0;
1362
1363 /* Receive window scaling. */
1364 if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323)
1365 cp->wscale = select_rcv_wscale();
1366 else
1367 cp->wscale = 0;
1368
1369 /* ECN */
1370 if (tcpopt->ecn && /* XXX: review. */
1371 (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn)))
1372 cp->ecn = 1;
1373 else
1374 cp->ecn = 0;
1375
1376 wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND);
1377 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1378
1379 if (tt->sndbuf > 0)
1380 cp->sndbuf = tt->sndbuf;
1381 else if (so->sol_sbsnd_flags & SB_AUTOSIZE &&
1382 V_tcp_do_autosndbuf)
1383 cp->sndbuf = 256 * 1024;
1384 else
1385 cp->sndbuf = so->sol_sbsnd_hiwat;
1386 } else {
1387 /* Active open */
1388
1389 /* TCP timestamp option */
1390 if (s->tstamp > 0 ||
1391 (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP)))
1392 cp->tstamp = 1;
1393 else
1394 cp->tstamp = 0;
1395
1396 /* SACK */
1397 if (s->sack > 0 ||
1398 (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT)))
1399 cp->sack = 1;
1400 else
1401 cp->sack = 0;
1402
1403 /* Receive window scaling */
1404 if (tp->t_flags & TF_REQ_SCALE)
1405 cp->wscale = select_rcv_wscale();
1406 else
1407 cp->wscale = 0;
1408
1409 /* ECN */
1410 if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1))
1411 cp->ecn = 1;
1412 else
1413 cp->ecn = 0;
1414
1415 SOCKBUF_LOCK(&so->so_rcv);
1416 wnd = max(select_rcv_wnd(so), MIN_RCV_WND);
1417 SOCKBUF_UNLOCK(&so->so_rcv);
1418 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1419
1420 if (tt->sndbuf > 0)
1421 cp->sndbuf = tt->sndbuf;
1422 else {
1423 SOCKBUF_LOCK(&so->so_snd);
1424 if (so->so_snd.sb_flags & SB_AUTOSIZE &&
1425 V_tcp_do_autosndbuf)
1426 cp->sndbuf = 256 * 1024;
1427 else
1428 cp->sndbuf = so->so_snd.sb_hiwat;
1429 SOCKBUF_UNLOCK(&so->so_snd);
1430 }
1431 }
1432
1433 cp->l2t_idx = l2t_idx;
1434
1435 /* This will be initialized on ESTABLISHED. */
1436 cp->emss = 0;
1437 }
1438
1439 int
negative_advice(int status)1440 negative_advice(int status)
1441 {
1442
1443 return (status == CPL_ERR_RTX_NEG_ADVICE ||
1444 status == CPL_ERR_PERSIST_NEG_ADVICE ||
1445 status == CPL_ERR_KEEPALV_NEG_ADVICE);
1446 }
1447
1448 static int
alloc_tid_tab(struct adapter * sc)1449 alloc_tid_tab(struct adapter *sc)
1450 {
1451 struct tid_info *t = &sc->tids;
1452
1453 MPASS(t->ntids > 0);
1454 MPASS(t->tid_tab == NULL);
1455
1456 t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE,
1457 M_ZERO | M_NOWAIT);
1458 if (t->tid_tab == NULL)
1459 return (ENOMEM);
1460 atomic_store_rel_int(&t->tids_in_use, 0);
1461
1462 return (0);
1463 }
1464
1465 static void
free_tid_tab(struct adapter * sc)1466 free_tid_tab(struct adapter *sc)
1467 {
1468 struct tid_info *t = &sc->tids;
1469
1470 KASSERT(t->tids_in_use == 0,
1471 ("%s: %d tids still in use.", __func__, t->tids_in_use));
1472
1473 free(t->tid_tab, M_CXGBE);
1474 t->tid_tab = NULL;
1475 }
1476
1477 static void
free_tid_tabs(struct adapter * sc)1478 free_tid_tabs(struct adapter *sc)
1479 {
1480 free_tid_tab(sc);
1481 free_stid_tab(sc);
1482 }
1483
1484 static int
alloc_tid_tabs(struct adapter * sc)1485 alloc_tid_tabs(struct adapter *sc)
1486 {
1487 int rc;
1488
1489 rc = alloc_tid_tab(sc);
1490 if (rc != 0)
1491 goto failed;
1492
1493 rc = alloc_stid_tab(sc);
1494 if (rc != 0)
1495 goto failed;
1496
1497 return (0);
1498 failed:
1499 free_tid_tabs(sc);
1500 return (rc);
1501 }
1502
1503 static inline void
alloc_tcb_history(struct adapter * sc,struct tom_data * td)1504 alloc_tcb_history(struct adapter *sc, struct tom_data *td)
1505 {
1506
1507 if (sc->tids.ntids == 0 || sc->tids.ntids > 1024)
1508 return;
1509 rw_init(&td->tcb_history_lock, "TCB history");
1510 td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history),
1511 M_CXGBE, M_ZERO | M_NOWAIT);
1512 td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0));
1513 }
1514
1515 static inline void
free_tcb_history(struct adapter * sc,struct tom_data * td)1516 free_tcb_history(struct adapter *sc, struct tom_data *td)
1517 {
1518 #ifdef INVARIANTS
1519 int i;
1520
1521 if (td->tcb_history != NULL) {
1522 for (i = 0; i < sc->tids.ntids; i++) {
1523 MPASS(td->tcb_history[i] == NULL);
1524 }
1525 }
1526 #endif
1527 free(td->tcb_history, M_CXGBE);
1528 if (rw_initialized(&td->tcb_history_lock))
1529 rw_destroy(&td->tcb_history_lock);
1530 }
1531
1532 static void
free_tom_data(struct adapter * sc,struct tom_data * td)1533 free_tom_data(struct adapter *sc, struct tom_data *td)
1534 {
1535
1536 ASSERT_SYNCHRONIZED_OP(sc);
1537
1538 KASSERT(TAILQ_EMPTY(&td->toep_list),
1539 ("%s: TOE PCB list is not empty.", __func__));
1540 KASSERT(td->lctx_count == 0,
1541 ("%s: lctx hash table is not empty.", __func__));
1542
1543 t4_free_ppod_region(&td->pr);
1544
1545 if (td->listen_mask != 0)
1546 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
1547
1548 if (mtx_initialized(&td->unsent_wr_lock))
1549 mtx_destroy(&td->unsent_wr_lock);
1550 if (mtx_initialized(&td->lctx_hash_lock))
1551 mtx_destroy(&td->lctx_hash_lock);
1552 if (mtx_initialized(&td->toep_list_lock))
1553 mtx_destroy(&td->toep_list_lock);
1554
1555 free_tcb_history(sc, td);
1556 free_tid_tabs(sc);
1557 free(td, M_CXGBE);
1558 }
1559
1560 static char *
prepare_pkt(int open_type,uint16_t vtag,struct inpcb * inp,int * pktlen,int * buflen)1561 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen,
1562 int *buflen)
1563 {
1564 char *pkt;
1565 struct tcphdr *th;
1566 int ipv6, len;
1567 const int maxlen =
1568 max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) +
1569 max(sizeof(struct ip), sizeof(struct ip6_hdr)) +
1570 sizeof(struct tcphdr);
1571
1572 MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN);
1573
1574 pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT);
1575 if (pkt == NULL)
1576 return (NULL);
1577
1578 ipv6 = inp->inp_vflag & INP_IPV6;
1579 len = 0;
1580
1581 if (EVL_VLANOFTAG(vtag) == 0xfff) {
1582 struct ether_header *eh = (void *)pkt;
1583
1584 if (ipv6)
1585 eh->ether_type = htons(ETHERTYPE_IPV6);
1586 else
1587 eh->ether_type = htons(ETHERTYPE_IP);
1588
1589 len += sizeof(*eh);
1590 } else {
1591 struct ether_vlan_header *evh = (void *)pkt;
1592
1593 evh->evl_encap_proto = htons(ETHERTYPE_VLAN);
1594 evh->evl_tag = htons(vtag);
1595 if (ipv6)
1596 evh->evl_proto = htons(ETHERTYPE_IPV6);
1597 else
1598 evh->evl_proto = htons(ETHERTYPE_IP);
1599
1600 len += sizeof(*evh);
1601 }
1602
1603 if (ipv6) {
1604 struct ip6_hdr *ip6 = (void *)&pkt[len];
1605
1606 ip6->ip6_vfc = IPV6_VERSION;
1607 ip6->ip6_plen = htons(sizeof(struct tcphdr));
1608 ip6->ip6_nxt = IPPROTO_TCP;
1609 if (open_type == OPEN_TYPE_ACTIVE) {
1610 ip6->ip6_src = inp->in6p_laddr;
1611 ip6->ip6_dst = inp->in6p_faddr;
1612 } else if (open_type == OPEN_TYPE_LISTEN) {
1613 ip6->ip6_src = inp->in6p_laddr;
1614 ip6->ip6_dst = ip6->ip6_src;
1615 }
1616
1617 len += sizeof(*ip6);
1618 } else {
1619 struct ip *ip = (void *)&pkt[len];
1620
1621 ip->ip_v = IPVERSION;
1622 ip->ip_hl = sizeof(*ip) >> 2;
1623 ip->ip_tos = inp->inp_ip_tos;
1624 ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr));
1625 ip->ip_ttl = inp->inp_ip_ttl;
1626 ip->ip_p = IPPROTO_TCP;
1627 if (open_type == OPEN_TYPE_ACTIVE) {
1628 ip->ip_src = inp->inp_laddr;
1629 ip->ip_dst = inp->inp_faddr;
1630 } else if (open_type == OPEN_TYPE_LISTEN) {
1631 ip->ip_src = inp->inp_laddr;
1632 ip->ip_dst = ip->ip_src;
1633 }
1634
1635 len += sizeof(*ip);
1636 }
1637
1638 th = (void *)&pkt[len];
1639 if (open_type == OPEN_TYPE_ACTIVE) {
1640 th->th_sport = inp->inp_lport; /* network byte order already */
1641 th->th_dport = inp->inp_fport; /* ditto */
1642 } else if (open_type == OPEN_TYPE_LISTEN) {
1643 th->th_sport = inp->inp_lport; /* network byte order already */
1644 th->th_dport = th->th_sport;
1645 }
1646 len += sizeof(th);
1647
1648 *pktlen = *buflen = len;
1649 return (pkt);
1650 }
1651
1652 const struct offload_settings *
lookup_offload_policy(struct adapter * sc,int open_type,struct mbuf * m,uint16_t vtag,struct inpcb * inp)1653 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m,
1654 uint16_t vtag, struct inpcb *inp)
1655 {
1656 const struct t4_offload_policy *op;
1657 char *pkt;
1658 struct offload_rule *r;
1659 int i, matched, pktlen, buflen;
1660 static const struct offload_settings allow_offloading_settings = {
1661 .offload = 1,
1662 .rx_coalesce = -1,
1663 .cong_algo = -1,
1664 .sched_class = -1,
1665 .tstamp = -1,
1666 .sack = -1,
1667 .nagle = -1,
1668 .ecn = -1,
1669 .ddp = -1,
1670 .tls = -1,
1671 .txq = QUEUE_RANDOM,
1672 .rxq = QUEUE_RANDOM,
1673 .mss = -1,
1674 };
1675 static const struct offload_settings disallow_offloading_settings = {
1676 .offload = 0,
1677 /* rest is irrelevant when offload is off. */
1678 };
1679
1680 rw_assert(&sc->policy_lock, RA_LOCKED);
1681
1682 /*
1683 * If there's no Connection Offloading Policy attached to the device
1684 * then we need to return a default static policy. If
1685 * "cop_managed_offloading" is true, then we need to disallow
1686 * offloading until a COP is attached to the device. Otherwise we
1687 * allow offloading ...
1688 */
1689 op = sc->policy;
1690 if (op == NULL) {
1691 if (sc->tt.cop_managed_offloading)
1692 return (&disallow_offloading_settings);
1693 else
1694 return (&allow_offloading_settings);
1695 }
1696
1697 switch (open_type) {
1698 case OPEN_TYPE_ACTIVE:
1699 case OPEN_TYPE_LISTEN:
1700 pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen);
1701 break;
1702 case OPEN_TYPE_PASSIVE:
1703 MPASS(m != NULL);
1704 pkt = mtod(m, char *);
1705 MPASS(*pkt == CPL_PASS_ACCEPT_REQ);
1706 pkt += sizeof(struct cpl_pass_accept_req);
1707 pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req);
1708 buflen = m->m_len - sizeof(struct cpl_pass_accept_req);
1709 break;
1710 default:
1711 MPASS(0);
1712 return (&disallow_offloading_settings);
1713 }
1714
1715 if (pkt == NULL || pktlen == 0 || buflen == 0)
1716 return (&disallow_offloading_settings);
1717
1718 matched = 0;
1719 r = &op->rule[0];
1720 for (i = 0; i < op->nrules; i++, r++) {
1721 if (r->open_type != open_type &&
1722 r->open_type != OPEN_TYPE_DONTCARE) {
1723 continue;
1724 }
1725 matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen);
1726 if (matched)
1727 break;
1728 }
1729
1730 if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN)
1731 free(pkt, M_CXGBE);
1732
1733 return (matched ? &r->settings : &disallow_offloading_settings);
1734 }
1735
1736 static void
reclaim_wr_resources(void * arg,int count)1737 reclaim_wr_resources(void *arg, int count)
1738 {
1739 struct tom_data *td = arg;
1740 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
1741 struct cpl_act_open_req *cpl;
1742 u_int opcode, atid, tid;
1743 struct wrqe *wr;
1744 struct adapter *sc = td_adapter(td);
1745
1746 mtx_lock(&td->unsent_wr_lock);
1747 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
1748 mtx_unlock(&td->unsent_wr_lock);
1749
1750 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
1751 STAILQ_REMOVE_HEAD(&twr_list, link);
1752
1753 cpl = wrtod(wr);
1754 opcode = GET_OPCODE(cpl);
1755
1756 switch (opcode) {
1757 case CPL_ACT_OPEN_REQ:
1758 case CPL_ACT_OPEN_REQ6:
1759 atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1760 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1761 act_open_failure_cleanup(sc, lookup_atid(sc, atid),
1762 EHOSTUNREACH);
1763 free(wr, M_CXGBE);
1764 break;
1765 case CPL_PASS_ACCEPT_RPL:
1766 tid = GET_TID(cpl);
1767 CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid);
1768 synack_failure_cleanup(sc, lookup_tid(sc, tid));
1769 free(wr, M_CXGBE);
1770 break;
1771 default:
1772 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1773 "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1774 /* WR not freed here; go look at it with a debugger. */
1775 }
1776 }
1777 }
1778
1779 /*
1780 * Based on do_abort_req. We treat an abrupt hardware stop as a connection
1781 * abort from the hardware.
1782 */
1783 static void
live_tid_failure_cleanup(struct adapter * sc,struct toepcb * toep,u_int status)1784 live_tid_failure_cleanup(struct adapter *sc, struct toepcb *toep, u_int status)
1785 {
1786 struct inpcb *inp;
1787 struct tcpcb *tp;
1788 struct epoch_tracker et;
1789
1790 MPASS(!(toep->flags & TPF_SYNQE));
1791
1792 inp = toep->inp;
1793 CURVNET_SET(toep->vnet);
1794 NET_EPOCH_ENTER(et); /* for tcp_close */
1795 INP_WLOCK(inp);
1796 tp = intotcpcb(inp);
1797 toep->flags |= TPF_ABORT_SHUTDOWN;
1798 if ((inp->inp_flags & INP_DROPPED) == 0) {
1799 struct socket *so = inp->inp_socket;
1800
1801 if (so != NULL)
1802 so_error_set(so, status);
1803 tp = tcp_close(tp);
1804 if (tp == NULL)
1805 INP_WLOCK(inp); /* re-acquire */
1806 }
1807 final_cpl_received(toep);
1808 NET_EPOCH_EXIT(et);
1809 CURVNET_RESTORE();
1810 }
1811
1812 static void
cleanup_stranded_tids(void * arg,int count)1813 cleanup_stranded_tids(void *arg, int count)
1814 {
1815 TAILQ_HEAD(, toepcb) tlist = TAILQ_HEAD_INITIALIZER(tlist);
1816 TAILQ_HEAD(, synq_entry) slist = TAILQ_HEAD_INITIALIZER(slist);
1817 struct tom_data *td = arg;
1818 struct adapter *sc = td_adapter(td);
1819 struct toepcb *toep;
1820 struct synq_entry *synqe;
1821
1822 /* Clean up synq entries. */
1823 mtx_lock(&td->toep_list_lock);
1824 TAILQ_SWAP(&td->stranded_synqe, &slist, synq_entry, link);
1825 mtx_unlock(&td->toep_list_lock);
1826 while ((synqe = TAILQ_FIRST(&slist)) != NULL) {
1827 TAILQ_REMOVE(&slist, synqe, link);
1828 MPASS(synqe->tid >= 0); /* stale, was kept around for debug */
1829 synqe->tid = -1;
1830 synack_failure_cleanup(sc, synqe);
1831 }
1832
1833 /* Clean up in-flight active opens. */
1834 mtx_lock(&td->toep_list_lock);
1835 TAILQ_SWAP(&td->stranded_atids, &tlist, toepcb, link);
1836 mtx_unlock(&td->toep_list_lock);
1837 while ((toep = TAILQ_FIRST(&tlist)) != NULL) {
1838 TAILQ_REMOVE(&tlist, toep, link);
1839 MPASS(toep->tid >= 0); /* stale, was kept around for debug */
1840 toep->tid = -1;
1841 act_open_failure_cleanup(sc, toep, EHOSTUNREACH);
1842 }
1843
1844 /* Clean up live connections. */
1845 mtx_lock(&td->toep_list_lock);
1846 TAILQ_SWAP(&td->stranded_tids, &tlist, toepcb, link);
1847 mtx_unlock(&td->toep_list_lock);
1848 while ((toep = TAILQ_FIRST(&tlist)) != NULL) {
1849 TAILQ_REMOVE(&tlist, toep, link);
1850 MPASS(toep->tid >= 0); /* stale, was kept around for debug */
1851 toep->tid = -1;
1852 live_tid_failure_cleanup(sc, toep, ECONNABORTED);
1853 }
1854 }
1855
1856 /*
1857 * Ground control to Major TOM
1858 * Commencing countdown, engines on
1859 */
1860 static int
t4_tom_activate(struct adapter * sc)1861 t4_tom_activate(struct adapter *sc)
1862 {
1863 struct tom_data *td;
1864 struct toedev *tod;
1865 struct vi_info *vi;
1866 int i, rc, v;
1867
1868 ASSERT_SYNCHRONIZED_OP(sc);
1869
1870 /* per-adapter softc for TOM */
1871 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1872 if (td == NULL)
1873 return (ENOMEM);
1874
1875 /* List of TOE PCBs and associated lock */
1876 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1877 TAILQ_INIT(&td->toep_list);
1878 TAILQ_INIT(&td->synqe_list);
1879 TAILQ_INIT(&td->stranded_atids);
1880 TAILQ_INIT(&td->stranded_tids);
1881 TASK_INIT(&td->cleanup_stranded_tids, 0, cleanup_stranded_tids, td);
1882
1883 /* Listen context */
1884 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1885 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1886 &td->listen_mask, HASH_NOWAIT);
1887
1888 /* List of WRs for which L2 resolution failed */
1889 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1890 STAILQ_INIT(&td->unsent_wr_list);
1891 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1892
1893 /* TID tables */
1894 rc = alloc_tid_tabs(sc);
1895 if (rc != 0)
1896 goto done;
1897
1898 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1899 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1900 if (rc != 0)
1901 goto done;
1902 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1903 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1904
1905 alloc_tcb_history(sc, td);
1906
1907 /* toedev ops */
1908 tod = &td->tod;
1909 init_toedev(tod);
1910 tod->tod_softc = sc;
1911 tod->tod_connect = t4_connect;
1912 tod->tod_listen_start = t4_listen_start;
1913 tod->tod_listen_stop = t4_listen_stop;
1914 tod->tod_rcvd = t4_rcvd;
1915 tod->tod_output = t4_tod_output;
1916 tod->tod_send_rst = t4_send_rst;
1917 tod->tod_send_fin = t4_send_fin;
1918 tod->tod_pcb_detach = t4_pcb_detach;
1919 tod->tod_l2_update = t4_l2_update;
1920 tod->tod_syncache_added = t4_syncache_added;
1921 tod->tod_syncache_removed = t4_syncache_removed;
1922 tod->tod_syncache_respond = t4_syncache_respond;
1923 tod->tod_offload_socket = t4_offload_socket;
1924 tod->tod_ctloutput = t4_ctloutput;
1925 tod->tod_tcp_info = t4_tcp_info;
1926 #ifdef KERN_TLS
1927 tod->tod_alloc_tls_session = t4_alloc_tls_session;
1928 #endif
1929 tod->tod_pmtu_update = t4_pmtu_update;
1930
1931 for_each_port(sc, i) {
1932 for_each_vi(sc->port[i], v, vi) {
1933 SETTOEDEV(vi->ifp, &td->tod);
1934 }
1935 }
1936
1937 sc->tom_softc = td;
1938 register_toedev(sc->tom_softc);
1939
1940 done:
1941 if (rc != 0)
1942 free_tom_data(sc, td);
1943 return (rc);
1944 }
1945
1946 static int
t4_tom_deactivate(struct adapter * sc)1947 t4_tom_deactivate(struct adapter *sc)
1948 {
1949 int rc = 0, i, v;
1950 struct tom_data *td = sc->tom_softc;
1951 struct vi_info *vi;
1952
1953 ASSERT_SYNCHRONIZED_OP(sc);
1954
1955 if (td == NULL)
1956 return (0); /* XXX. KASSERT? */
1957
1958 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1959 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */
1960
1961 if (sc->offload_map != 0) {
1962 for_each_port(sc, i) {
1963 for_each_vi(sc->port[i], v, vi) {
1964 toe_capability(vi, false);
1965 if_setcapenablebit(vi->ifp, 0, IFCAP_TOE);
1966 SETTOEDEV(vi->ifp, NULL);
1967 }
1968 }
1969 MPASS(sc->offload_map == 0);
1970 }
1971
1972 mtx_lock(&td->toep_list_lock);
1973 if (!TAILQ_EMPTY(&td->toep_list))
1974 rc = EBUSY;
1975 MPASS(TAILQ_EMPTY(&td->synqe_list));
1976 MPASS(TAILQ_EMPTY(&td->stranded_tids));
1977 mtx_unlock(&td->toep_list_lock);
1978
1979 mtx_lock(&td->lctx_hash_lock);
1980 if (td->lctx_count > 0)
1981 rc = EBUSY;
1982 mtx_unlock(&td->lctx_hash_lock);
1983
1984 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1985 taskqueue_drain(taskqueue_thread, &td->cleanup_stranded_tids);
1986 mtx_lock(&td->unsent_wr_lock);
1987 if (!STAILQ_EMPTY(&td->unsent_wr_list))
1988 rc = EBUSY;
1989 mtx_unlock(&td->unsent_wr_lock);
1990
1991 if (rc == 0) {
1992 unregister_toedev(sc->tom_softc);
1993 free_tom_data(sc, td);
1994 sc->tom_softc = NULL;
1995 }
1996
1997 return (rc);
1998 }
1999
2000 static void
stop_atids(struct adapter * sc)2001 stop_atids(struct adapter *sc)
2002 {
2003 struct tom_data *td = sc->tom_softc;
2004 struct tid_info *t = &sc->tids;
2005 struct toepcb *toep;
2006 int atid;
2007
2008 /*
2009 * Hashfilters and T6-KTLS are the only other users of atids but they're
2010 * both mutually exclusive with TOE. That means t4_tom owns all the
2011 * atids in the table.
2012 */
2013 MPASS(!is_hashfilter(sc));
2014 if (is_t6(sc))
2015 MPASS(!(sc->flags & KERN_TLS_ON));
2016
2017 /* New atids are not being allocated. */
2018 #ifdef INVARIANTS
2019 mtx_lock(&t->atid_lock);
2020 MPASS(t->atid_alloc_stopped == true);
2021 mtx_unlock(&t->atid_lock);
2022 #endif
2023
2024 /*
2025 * In-use atids fall in one of these two categories:
2026 * a) Those waiting for L2 resolution before being submitted to
2027 * hardware.
2028 * b) Those that have been submitted to hardware and are awaiting
2029 * replies that will never arrive because the LLD is stopped.
2030 */
2031 for (atid = 0; atid < t->natids; atid++) {
2032 toep = lookup_atid(sc, atid);
2033 if ((uintptr_t)toep >= (uintptr_t)&t->atid_tab[0] &&
2034 (uintptr_t)toep < (uintptr_t)&t->atid_tab[t->natids])
2035 continue;
2036 if (__predict_false(toep == NULL))
2037 continue;
2038 MPASS(toep->tid == atid);
2039 MPASS(toep->incarnation == sc->incarnation);
2040 /*
2041 * Take the atid out of the lookup table. toep->tid is stale
2042 * after this but useful for debug.
2043 */
2044 CTR(KTR_CXGBE, "%s: atid %d@%d STRANDED, removed from table",
2045 __func__, atid, toep->incarnation);
2046 free_atid(sc, toep->tid);
2047 #if 0
2048 toep->tid = -1;
2049 #endif
2050 mtx_lock(&td->toep_list_lock);
2051 toep->flags &= ~TPF_IN_TOEP_LIST;
2052 TAILQ_REMOVE(&td->toep_list, toep, link);
2053 TAILQ_INSERT_TAIL(&td->stranded_atids, toep, link);
2054 mtx_unlock(&td->toep_list_lock);
2055 }
2056 MPASS(atomic_load_int(&t->atids_in_use) == 0);
2057 }
2058
2059 static void
stop_tids(struct adapter * sc)2060 stop_tids(struct adapter *sc)
2061 {
2062 struct tom_data *td = sc->tom_softc;
2063 struct toepcb *toep;
2064 #ifdef INVARIANTS
2065 struct tid_info *t = &sc->tids;
2066 #endif
2067
2068 /*
2069 * The LLD's offload queues are stopped so do_act_establish and
2070 * do_pass_accept_req cannot run and insert tids in parallel with this
2071 * thread. stop_stid_tab has also run and removed the synq entries'
2072 * tids from the table. The only tids in the table are for connections
2073 * at or beyond ESTABLISHED that are still waiting for the final CPL.
2074 */
2075 mtx_lock(&td->toep_list_lock);
2076 TAILQ_FOREACH(toep, &td->toep_list, link) {
2077 MPASS(sc->incarnation == toep->incarnation);
2078 MPASS(toep->tid >= 0);
2079 MPASS(toep == lookup_tid(sc, toep->tid));
2080 /* Remove tid from the lookup table immediately. */
2081 CTR(KTR_CXGBE, "%s: tid %d@%d STRANDED, removed from table",
2082 __func__, toep->tid, toep->incarnation);
2083 remove_tid(sc, toep->tid, toep->ce ? 2 : 1);
2084 #if 0
2085 /* toep->tid is stale now but left alone for debug. */
2086 toep->tid = -1;
2087 #endif
2088 /* All toep in this list will get bulk moved to stranded_tids */
2089 toep->flags &= ~TPF_IN_TOEP_LIST;
2090 }
2091 MPASS(TAILQ_EMPTY(&td->stranded_tids));
2092 TAILQ_CONCAT(&td->stranded_tids, &td->toep_list, link);
2093 MPASS(TAILQ_EMPTY(&td->toep_list));
2094 mtx_unlock(&td->toep_list_lock);
2095
2096 MPASS(atomic_load_int(&t->tids_in_use) == 0);
2097 }
2098
2099 /*
2100 * L2T is stable because
2101 * 1. stop_lld stopped all new allocations.
2102 * 2. stop_lld also stopped the tx wrq so nothing is enqueueing new WRs to the
2103 * queue or to l2t_entry->wr_list.
2104 * 3. t4_l2t_update is ignoring all L2 updates.
2105 */
2106 static void
stop_tom_l2t(struct adapter * sc)2107 stop_tom_l2t(struct adapter *sc)
2108 {
2109 struct l2t_data *d = sc->l2t;
2110 struct tom_data *td = sc->tom_softc;
2111 struct l2t_entry *e;
2112 struct wrqe *wr;
2113 int i;
2114
2115 /*
2116 * This task cannot be enqueued because L2 state changes are not being
2117 * processed. But if it's already scheduled or running then we need to
2118 * wait for it to cleanup the atids in the unsent_wr_list.
2119 */
2120 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
2121 MPASS(STAILQ_EMPTY(&td->unsent_wr_list));
2122
2123 for (i = 0; i < d->l2t_size; i++) {
2124 e = &d->l2tab[i];
2125 mtx_lock(&e->lock);
2126 if (e->state == L2T_STATE_VALID || e->state == L2T_STATE_STALE)
2127 e->state = L2T_STATE_RESOLVING;
2128 /*
2129 * stop_atids is going to clean up _all_ atids in use, including
2130 * these that were pending L2 resolution. Just discard the WRs.
2131 */
2132 while ((wr = STAILQ_FIRST(&e->wr_list)) != NULL) {
2133 STAILQ_REMOVE_HEAD(&e->wr_list, link);
2134 free(wr, M_CXGBE);
2135 }
2136 mtx_unlock(&e->lock);
2137 }
2138 }
2139
2140 static int
t4_tom_stop(struct adapter * sc)2141 t4_tom_stop(struct adapter *sc)
2142 {
2143 struct tid_info *t = &sc->tids;
2144 struct tom_data *td = sc->tom_softc;
2145
2146 ASSERT_SYNCHRONIZED_OP(sc);
2147
2148 stop_tom_l2t(sc);
2149 if (atomic_load_int(&t->atids_in_use) > 0)
2150 stop_atids(sc);
2151 if (atomic_load_int(&t->stids_in_use) > 0)
2152 stop_stid_tab(sc);
2153 if (atomic_load_int(&t->tids_in_use) > 0)
2154 stop_tids(sc);
2155 taskqueue_enqueue(taskqueue_thread, &td->cleanup_stranded_tids);
2156
2157 /*
2158 * L2T and atid_tab are restarted before t4_tom_restart so this assert
2159 * is not valid in t4_tom_restart. This is the next best place for it.
2160 */
2161 MPASS(STAILQ_EMPTY(&td->unsent_wr_list));
2162
2163 return (0);
2164 }
2165
2166 static int
t4_tom_restart(struct adapter * sc)2167 t4_tom_restart(struct adapter *sc)
2168 {
2169 ASSERT_SYNCHRONIZED_OP(sc);
2170
2171 restart_stid_tab(sc);
2172
2173 return (0);
2174 }
2175
2176 static int
t4_ctloutput_tom(struct socket * so,struct sockopt * sopt)2177 t4_ctloutput_tom(struct socket *so, struct sockopt *sopt)
2178 {
2179 struct tcpcb *tp = sototcpcb(so);
2180 struct toepcb *toep = tp->t_toe;
2181 int error, optval;
2182
2183 if (sopt->sopt_level == IPPROTO_TCP && sopt->sopt_name == TCP_USE_DDP) {
2184 if (sopt->sopt_dir != SOPT_SET)
2185 return (EOPNOTSUPP);
2186
2187 if (sopt->sopt_td != NULL) {
2188 /* Only settable by the kernel. */
2189 return (EPERM);
2190 }
2191
2192 error = sooptcopyin(sopt, &optval, sizeof(optval),
2193 sizeof(optval));
2194 if (error != 0)
2195 return (error);
2196
2197 if (optval != 0)
2198 return (t4_enable_ddp_rcv(so, toep));
2199 else
2200 return (EOPNOTSUPP);
2201 }
2202 return (tcp_ctloutput(so, sopt));
2203 }
2204
2205 static int
t4_aio_queue_tom(struct socket * so,struct kaiocb * job)2206 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
2207 {
2208 struct tcpcb *tp = sototcpcb(so);
2209 struct toepcb *toep = tp->t_toe;
2210 int error;
2211
2212 /*
2213 * No lock is needed as TOE sockets never change between
2214 * active and passive.
2215 */
2216 if (SOLISTENING(so))
2217 return (EINVAL);
2218
2219 if (ulp_mode(toep) == ULP_MODE_TCPDDP ||
2220 ulp_mode(toep) == ULP_MODE_NONE) {
2221 error = t4_aio_queue_ddp(so, job);
2222 if (error != EOPNOTSUPP)
2223 return (error);
2224 }
2225
2226 return (t4_aio_queue_aiotx(so, job));
2227 }
2228
2229 static int
t4_tom_mod_load(void)2230 t4_tom_mod_load(void)
2231 {
2232 /* CPL handlers */
2233 t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl);
2234 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2,
2235 CPL_COOKIE_TOM);
2236 t4_init_connect_cpl_handlers();
2237 t4_init_listen_cpl_handlers();
2238 t4_init_cpl_io_handlers();
2239
2240 t4_ddp_mod_load();
2241 t4_tls_mod_load();
2242
2243 bcopy(&tcp_protosw, &toe_protosw, sizeof(toe_protosw));
2244 toe_protosw.pr_ctloutput = t4_ctloutput_tom;
2245 toe_protosw.pr_aio_queue = t4_aio_queue_tom;
2246
2247 bcopy(&tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
2248 toe6_protosw.pr_ctloutput = t4_ctloutput_tom;
2249 toe6_protosw.pr_aio_queue = t4_aio_queue_tom;
2250
2251 return (t4_register_uld(&tom_uld_info, ULD_TOM));
2252 }
2253
2254 static void
tom_uninit(struct adapter * sc,void * arg)2255 tom_uninit(struct adapter *sc, void *arg)
2256 {
2257 bool *ok_to_unload = arg;
2258
2259 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
2260 return;
2261
2262 /* Try to free resources (works only if no port has IFCAP_TOE) */
2263 if (uld_active(sc, ULD_TOM) && t4_deactivate_uld(sc, ULD_TOM) != 0)
2264 *ok_to_unload = false;
2265
2266 end_synchronized_op(sc, 0);
2267 }
2268
2269 static int
t4_tom_mod_unload(void)2270 t4_tom_mod_unload(void)
2271 {
2272 bool ok_to_unload = true;
2273
2274 t4_iterate(tom_uninit, &ok_to_unload);
2275 if (!ok_to_unload)
2276 return (EBUSY);
2277
2278 if (t4_unregister_uld(&tom_uld_info, ULD_TOM) == EBUSY)
2279 return (EBUSY);
2280
2281 t4_tls_mod_unload();
2282 t4_ddp_mod_unload();
2283
2284 t4_uninit_connect_cpl_handlers();
2285 t4_uninit_listen_cpl_handlers();
2286 t4_uninit_cpl_io_handlers();
2287 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM);
2288 t4_register_cpl_handler(CPL_GET_TCB_RPL, NULL);
2289
2290 return (0);
2291 }
2292 #endif /* TCP_OFFLOAD */
2293
2294 static int
t4_tom_modevent(module_t mod,int cmd,void * arg)2295 t4_tom_modevent(module_t mod, int cmd, void *arg)
2296 {
2297 int rc = 0;
2298
2299 #ifdef TCP_OFFLOAD
2300 switch (cmd) {
2301 case MOD_LOAD:
2302 rc = t4_tom_mod_load();
2303 break;
2304
2305 case MOD_UNLOAD:
2306 rc = t4_tom_mod_unload();
2307 break;
2308
2309 default:
2310 rc = EINVAL;
2311 }
2312 #else
2313 printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
2314 rc = EOPNOTSUPP;
2315 #endif
2316 return (rc);
2317 }
2318
2319 static moduledata_t t4_tom_moddata= {
2320 "t4_tom",
2321 t4_tom_modevent,
2322 0
2323 };
2324
2325 MODULE_VERSION(t4_tom, 1);
2326 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
2327 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
2328 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
2329