1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2012 Chelsio Communications, Inc.
5 * All rights reserved.
6 * Written by: Navdeep Parhar <[email protected]>
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 #include "opt_ratelimit.h"
36
37 #include <sys/param.h>
38 #include <sys/types.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/ktr.h>
42 #include <sys/lock.h>
43 #include <sys/limits.h>
44 #include <sys/module.h>
45 #include <sys/protosw.h>
46 #include <sys/domain.h>
47 #include <sys/refcount.h>
48 #include <sys/rmlock.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/taskqueue.h>
53 #include <net/if.h>
54 #include <net/if_var.h>
55 #include <net/if_types.h>
56 #include <net/if_vlan_var.h>
57 #include <netinet/in.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/in_var.h>
60 #include <netinet/ip.h>
61 #include <netinet/ip6.h>
62 #include <netinet6/scope6_var.h>
63 #define TCPSTATES
64 #include <netinet/tcp_fsm.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 pr_usrreqs toe_usrreqs;
83
84 static struct protosw toe6_protosw;
85 static struct pr_usrreqs toe6_usrreqs;
86
87 /* Module ops */
88 static int t4_tom_mod_load(void);
89 static int t4_tom_mod_unload(void);
90 static int t4_tom_modevent(module_t, int, void *);
91
92 /* ULD ops and helpers */
93 static int t4_tom_activate(struct adapter *);
94 static int t4_tom_deactivate(struct adapter *);
95
96 static struct uld_info tom_uld_info = {
97 .uld_id = ULD_TOM,
98 .activate = t4_tom_activate,
99 .deactivate = t4_tom_deactivate,
100 };
101
102 static void release_offload_resources(struct toepcb *);
103 static int alloc_tid_tabs(struct tid_info *);
104 static void free_tid_tabs(struct tid_info *);
105 static void free_tom_data(struct adapter *, struct tom_data *);
106 static void reclaim_wr_resources(void *, int);
107
108 struct toepcb *
alloc_toepcb(struct vi_info * vi,int flags)109 alloc_toepcb(struct vi_info *vi, int flags)
110 {
111 struct port_info *pi = vi->pi;
112 struct adapter *sc = pi->adapter;
113 struct toepcb *toep;
114 int tx_credits, txsd_total, len;
115
116 /*
117 * The firmware counts tx work request credits in units of 16 bytes
118 * each. Reserve room for an ABORT_REQ so the driver never has to worry
119 * about tx credits if it wants to abort a connection.
120 */
121 tx_credits = sc->params.ofldq_wr_cred;
122 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16);
123
124 /*
125 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte
126 * immediate payload, and firmware counts tx work request credits in
127 * units of 16 byte. Calculate the maximum work requests possible.
128 */
129 txsd_total = tx_credits /
130 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16);
131
132 len = offsetof(struct toepcb, txsd) +
133 txsd_total * sizeof(struct ofld_tx_sdesc);
134
135 toep = malloc(len, M_CXGBE, M_ZERO | flags);
136 if (toep == NULL)
137 return (NULL);
138
139 refcount_init(&toep->refcount, 1);
140 toep->td = sc->tom_softc;
141 toep->vi = vi;
142 toep->tid = -1;
143 toep->tx_total = tx_credits;
144 toep->tx_credits = tx_credits;
145 mbufq_init(&toep->ulp_pduq, INT_MAX);
146 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX);
147 toep->txsd_total = txsd_total;
148 toep->txsd_avail = txsd_total;
149 toep->txsd_pidx = 0;
150 toep->txsd_cidx = 0;
151 aiotx_init_toep(toep);
152
153 return (toep);
154 }
155
156 /*
157 * Initialize a toepcb after its params have been filled out.
158 */
159 int
init_toepcb(struct vi_info * vi,struct toepcb * toep)160 init_toepcb(struct vi_info *vi, struct toepcb *toep)
161 {
162 struct conn_params *cp = &toep->params;
163 struct port_info *pi = vi->pi;
164 struct adapter *sc = pi->adapter;
165 struct tx_cl_rl_params *tc;
166
167 if (cp->tc_idx >= 0 && cp->tc_idx < sc->chip_params->nsched_cls) {
168 tc = &pi->sched_params->cl_rl[cp->tc_idx];
169 mtx_lock(&sc->tc_lock);
170 if (tc->flags & CLRL_ERR) {
171 log(LOG_ERR,
172 "%s: failed to associate traffic class %u with tid %u\n",
173 device_get_nameunit(vi->dev), cp->tc_idx,
174 toep->tid);
175 cp->tc_idx = -1;
176 } else {
177 tc->refcount++;
178 }
179 mtx_unlock(&sc->tc_lock);
180 }
181 toep->ofld_txq = &sc->sge.ofld_txq[cp->txq_idx];
182 toep->ofld_rxq = &sc->sge.ofld_rxq[cp->rxq_idx];
183 toep->ctrlq = &sc->sge.ctrlq[pi->port_id];
184
185 tls_init_toep(toep);
186 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
187 ddp_init_toep(toep);
188
189 return (0);
190 }
191
192 struct toepcb *
hold_toepcb(struct toepcb * toep)193 hold_toepcb(struct toepcb *toep)
194 {
195
196 refcount_acquire(&toep->refcount);
197 return (toep);
198 }
199
200 void
free_toepcb(struct toepcb * toep)201 free_toepcb(struct toepcb *toep)
202 {
203
204 if (refcount_release(&toep->refcount) == 0)
205 return;
206
207 KASSERT(!(toep->flags & TPF_ATTACHED),
208 ("%s: attached to an inpcb", __func__));
209 KASSERT(!(toep->flags & TPF_CPL_PENDING),
210 ("%s: CPL pending", __func__));
211
212 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
213 ddp_uninit_toep(toep);
214 tls_uninit_toep(toep);
215 free(toep, M_CXGBE);
216 }
217
218 /*
219 * Set up the socket for TCP offload.
220 */
221 void
offload_socket(struct socket * so,struct toepcb * toep)222 offload_socket(struct socket *so, struct toepcb *toep)
223 {
224 struct tom_data *td = toep->td;
225 struct inpcb *inp = sotoinpcb(so);
226 struct tcpcb *tp = intotcpcb(inp);
227 struct sockbuf *sb;
228
229 INP_WLOCK_ASSERT(inp);
230
231 /* Update socket */
232 sb = &so->so_snd;
233 SOCKBUF_LOCK(sb);
234 sb->sb_flags |= SB_NOCOALESCE;
235 SOCKBUF_UNLOCK(sb);
236 sb = &so->so_rcv;
237 SOCKBUF_LOCK(sb);
238 sb->sb_flags |= SB_NOCOALESCE;
239 if (inp->inp_vflag & INP_IPV6)
240 so->so_proto = &toe6_protosw;
241 else
242 so->so_proto = &toe_protosw;
243 SOCKBUF_UNLOCK(sb);
244
245 /* Update TCP PCB */
246 tp->tod = &td->tod;
247 tp->t_toe = toep;
248 tp->t_flags |= TF_TOE;
249
250 /* Install an extra hold on inp */
251 toep->inp = inp;
252 toep->flags |= TPF_ATTACHED;
253 in_pcbref(inp);
254
255 /* Add the TOE PCB to the active list */
256 mtx_lock(&td->toep_list_lock);
257 TAILQ_INSERT_HEAD(&td->toep_list, toep, link);
258 mtx_unlock(&td->toep_list_lock);
259 }
260
261 /* This is _not_ the normal way to "unoffload" a socket. */
262 void
undo_offload_socket(struct socket * so)263 undo_offload_socket(struct socket *so)
264 {
265 struct inpcb *inp = sotoinpcb(so);
266 struct tcpcb *tp = intotcpcb(inp);
267 struct toepcb *toep = tp->t_toe;
268 struct tom_data *td = toep->td;
269 struct sockbuf *sb;
270
271 INP_WLOCK_ASSERT(inp);
272
273 sb = &so->so_snd;
274 SOCKBUF_LOCK(sb);
275 sb->sb_flags &= ~SB_NOCOALESCE;
276 SOCKBUF_UNLOCK(sb);
277 sb = &so->so_rcv;
278 SOCKBUF_LOCK(sb);
279 sb->sb_flags &= ~SB_NOCOALESCE;
280 SOCKBUF_UNLOCK(sb);
281
282 tp->tod = NULL;
283 tp->t_toe = NULL;
284 tp->t_flags &= ~TF_TOE;
285
286 toep->inp = NULL;
287 toep->flags &= ~TPF_ATTACHED;
288 if (in_pcbrele_wlocked(inp))
289 panic("%s: inp freed.", __func__);
290
291 mtx_lock(&td->toep_list_lock);
292 TAILQ_REMOVE(&td->toep_list, toep, link);
293 mtx_unlock(&td->toep_list_lock);
294 }
295
296 static void
release_offload_resources(struct toepcb * toep)297 release_offload_resources(struct toepcb *toep)
298 {
299 struct tom_data *td = toep->td;
300 struct adapter *sc = td_adapter(td);
301 int tid = toep->tid;
302
303 KASSERT(!(toep->flags & TPF_CPL_PENDING),
304 ("%s: %p has CPL pending.", __func__, toep));
305 KASSERT(!(toep->flags & TPF_ATTACHED),
306 ("%s: %p is still attached.", __func__, toep));
307
308 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)",
309 __func__, toep, tid, toep->l2te, toep->ce);
310
311 /*
312 * These queues should have been emptied at approximately the same time
313 * that a normal connection's socket's so_snd would have been purged or
314 * drained. Do _not_ clean up here.
315 */
316 MPASS(mbufq_len(&toep->ulp_pduq) == 0);
317 MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0);
318 #ifdef INVARIANTS
319 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
320 ddp_assert_empty(toep);
321 #endif
322 MPASS(TAILQ_EMPTY(&toep->aiotx_jobq));
323
324 if (toep->l2te)
325 t4_l2t_release(toep->l2te);
326
327 if (tid >= 0) {
328 remove_tid(sc, tid, toep->ce ? 2 : 1);
329 release_tid(sc, tid, toep->ctrlq);
330 }
331
332 if (toep->ce)
333 t4_release_lip(sc, toep->ce);
334
335 if (toep->params.tc_idx != -1)
336 t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx);
337
338 mtx_lock(&td->toep_list_lock);
339 TAILQ_REMOVE(&td->toep_list, toep, link);
340 mtx_unlock(&td->toep_list_lock);
341
342 free_toepcb(toep);
343 }
344
345 /*
346 * The kernel is done with the TCP PCB and this is our opportunity to unhook the
347 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no
348 * pending CPL) then it is time to release all resources tied to the toepcb.
349 *
350 * Also gets called when an offloaded active open fails and the TOM wants the
351 * kernel to take the TCP PCB back.
352 */
353 static void
t4_pcb_detach(struct toedev * tod __unused,struct tcpcb * tp)354 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp)
355 {
356 #if defined(KTR) || defined(INVARIANTS)
357 struct inpcb *inp = tp->t_inpcb;
358 #endif
359 struct toepcb *toep = tp->t_toe;
360
361 INP_WLOCK_ASSERT(inp);
362
363 KASSERT(toep != NULL, ("%s: toep is NULL", __func__));
364 KASSERT(toep->flags & TPF_ATTACHED,
365 ("%s: not attached", __func__));
366
367 #ifdef KTR
368 if (tp->t_state == TCPS_SYN_SENT) {
369 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)",
370 __func__, toep->tid, toep, toep->flags, inp,
371 inp->inp_flags);
372 } else {
373 CTR6(KTR_CXGBE,
374 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)",
375 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp,
376 inp->inp_flags);
377 }
378 #endif
379
380 tp->t_toe = NULL;
381 tp->t_flags &= ~TF_TOE;
382 toep->flags &= ~TPF_ATTACHED;
383
384 if (!(toep->flags & TPF_CPL_PENDING))
385 release_offload_resources(toep);
386 }
387
388 /*
389 * setsockopt handler.
390 */
391 static void
t4_ctloutput(struct toedev * tod,struct tcpcb * tp,int dir,int name)392 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name)
393 {
394 struct adapter *sc = tod->tod_softc;
395 struct toepcb *toep = tp->t_toe;
396
397 if (dir == SOPT_GET)
398 return;
399
400 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name);
401
402 switch (name) {
403 case TCP_NODELAY:
404 if (tp->t_state != TCPS_ESTABLISHED)
405 break;
406 toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
407 t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS,
408 V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0);
409 break;
410 default:
411 break;
412 }
413 }
414
415 static inline uint64_t
get_tcb_tflags(const uint64_t * tcb)416 get_tcb_tflags(const uint64_t *tcb)
417 {
418
419 return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32));
420 }
421
422 static inline uint32_t
get_tcb_field(const uint64_t * tcb,u_int word,uint32_t mask,u_int shift)423 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift)
424 {
425 #define LAST_WORD ((TCB_SIZE / 4) - 1)
426 uint64_t t1, t2;
427 int flit_idx;
428
429 MPASS(mask != 0);
430 MPASS(word <= LAST_WORD);
431 MPASS(shift < 32);
432
433 flit_idx = (LAST_WORD - word) / 2;
434 if (word & 0x1)
435 shift += 32;
436 t1 = be64toh(tcb[flit_idx]) >> shift;
437 t2 = 0;
438 if (fls(mask) > 64 - shift) {
439 /*
440 * Will spill over into the next logical flit, which is the flit
441 * before this one. The flit_idx before this one must be valid.
442 */
443 MPASS(flit_idx > 0);
444 t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift);
445 }
446 return ((t2 | t1) & mask);
447 #undef LAST_WORD
448 }
449 #define GET_TCB_FIELD(tcb, F) \
450 get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F)
451
452 /*
453 * Issues a CPL_GET_TCB to read the entire TCB for the tid.
454 */
455 static int
send_get_tcb(struct adapter * sc,u_int tid)456 send_get_tcb(struct adapter *sc, u_int tid)
457 {
458 struct cpl_get_tcb *cpl;
459 struct wrq_cookie cookie;
460
461 MPASS(tid < sc->tids.ntids);
462
463 cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16),
464 &cookie);
465 if (__predict_false(cpl == NULL))
466 return (ENOMEM);
467 bzero(cpl, sizeof(*cpl));
468 INIT_TP_WR(cpl, tid);
469 OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid));
470 cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) |
471 V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id));
472 cpl->cookie = 0xff;
473 commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie);
474
475 return (0);
476 }
477
478 static struct tcb_histent *
alloc_tcb_histent(struct adapter * sc,u_int tid,int flags)479 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags)
480 {
481 struct tcb_histent *te;
482
483 MPASS(flags == M_NOWAIT || flags == M_WAITOK);
484
485 te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags);
486 if (te == NULL)
487 return (NULL);
488 mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF);
489 callout_init_mtx(&te->te_callout, &te->te_lock, 0);
490 te->te_adapter = sc;
491 te->te_tid = tid;
492
493 return (te);
494 }
495
496 static void
free_tcb_histent(struct tcb_histent * te)497 free_tcb_histent(struct tcb_histent *te)
498 {
499
500 mtx_destroy(&te->te_lock);
501 free(te, M_CXGBE);
502 }
503
504 /*
505 * Start tracking the tid in the TCB history.
506 */
507 int
add_tid_to_history(struct adapter * sc,u_int tid)508 add_tid_to_history(struct adapter *sc, u_int tid)
509 {
510 struct tcb_histent *te = NULL;
511 struct tom_data *td = sc->tom_softc;
512 int rc;
513
514 MPASS(tid < sc->tids.ntids);
515
516 if (td->tcb_history == NULL)
517 return (ENXIO);
518
519 rw_wlock(&td->tcb_history_lock);
520 if (td->tcb_history[tid] != NULL) {
521 rc = EEXIST;
522 goto done;
523 }
524 te = alloc_tcb_histent(sc, tid, M_NOWAIT);
525 if (te == NULL) {
526 rc = ENOMEM;
527 goto done;
528 }
529 mtx_lock(&te->te_lock);
530 rc = send_get_tcb(sc, tid);
531 if (rc == 0) {
532 te->te_flags |= TE_RPL_PENDING;
533 td->tcb_history[tid] = te;
534 } else {
535 free(te, M_CXGBE);
536 }
537 mtx_unlock(&te->te_lock);
538 done:
539 rw_wunlock(&td->tcb_history_lock);
540 return (rc);
541 }
542
543 static void
remove_tcb_histent(struct tcb_histent * te)544 remove_tcb_histent(struct tcb_histent *te)
545 {
546 struct adapter *sc = te->te_adapter;
547 struct tom_data *td = sc->tom_softc;
548
549 rw_assert(&td->tcb_history_lock, RA_WLOCKED);
550 mtx_assert(&te->te_lock, MA_OWNED);
551 MPASS(td->tcb_history[te->te_tid] == te);
552
553 td->tcb_history[te->te_tid] = NULL;
554 free_tcb_histent(te);
555 rw_wunlock(&td->tcb_history_lock);
556 }
557
558 static inline struct tcb_histent *
lookup_tcb_histent(struct adapter * sc,u_int tid,bool addrem)559 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem)
560 {
561 struct tcb_histent *te;
562 struct tom_data *td = sc->tom_softc;
563
564 MPASS(tid < sc->tids.ntids);
565
566 if (td->tcb_history == NULL)
567 return (NULL);
568
569 if (addrem)
570 rw_wlock(&td->tcb_history_lock);
571 else
572 rw_rlock(&td->tcb_history_lock);
573 te = td->tcb_history[tid];
574 if (te != NULL) {
575 mtx_lock(&te->te_lock);
576 return (te); /* with both locks held */
577 }
578 if (addrem)
579 rw_wunlock(&td->tcb_history_lock);
580 else
581 rw_runlock(&td->tcb_history_lock);
582
583 return (te);
584 }
585
586 static inline void
release_tcb_histent(struct tcb_histent * te)587 release_tcb_histent(struct tcb_histent *te)
588 {
589 struct adapter *sc = te->te_adapter;
590 struct tom_data *td = sc->tom_softc;
591
592 mtx_assert(&te->te_lock, MA_OWNED);
593 mtx_unlock(&te->te_lock);
594 rw_assert(&td->tcb_history_lock, RA_RLOCKED);
595 rw_runlock(&td->tcb_history_lock);
596 }
597
598 static void
request_tcb(void * arg)599 request_tcb(void *arg)
600 {
601 struct tcb_histent *te = arg;
602
603 mtx_assert(&te->te_lock, MA_OWNED);
604
605 /* Noone else is supposed to update the histent. */
606 MPASS(!(te->te_flags & TE_RPL_PENDING));
607 if (send_get_tcb(te->te_adapter, te->te_tid) == 0)
608 te->te_flags |= TE_RPL_PENDING;
609 else
610 callout_schedule(&te->te_callout, hz / 100);
611 }
612
613 static void
update_tcb_histent(struct tcb_histent * te,const uint64_t * tcb)614 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb)
615 {
616 struct tom_data *td = te->te_adapter->tom_softc;
617 uint64_t tflags = get_tcb_tflags(tcb);
618 uint8_t sample = 0;
619
620 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) {
621 if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0)
622 sample |= TS_RTO;
623 if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0)
624 sample |= TS_DUPACKS;
625 if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold)
626 sample |= TS_FASTREXMT;
627 }
628
629 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) {
630 uint32_t snd_wnd;
631
632 sample |= TS_SND_BACKLOGGED; /* for whatever reason. */
633
634 snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
635 if (tflags & V_TF_RECV_SCALE(1))
636 snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE);
637 if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd)
638 sample |= TS_CWND_LIMITED; /* maybe due to CWND */
639 }
640
641 if (tflags & V_TF_CCTRL_ECN(1)) {
642
643 /*
644 * CE marker on incoming IP hdr, echoing ECE back in the TCP
645 * hdr. Indicates congestion somewhere on the way from the peer
646 * to this node.
647 */
648 if (tflags & V_TF_CCTRL_ECE(1))
649 sample |= TS_ECN_ECE;
650
651 /*
652 * ECE seen and CWR sent (or about to be sent). Might indicate
653 * congestion on the way to the peer. This node is reducing its
654 * congestion window in response.
655 */
656 if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1)))
657 sample |= TS_ECN_CWR;
658 }
659
660 te->te_sample[te->te_pidx] = sample;
661 if (++te->te_pidx == nitems(te->te_sample))
662 te->te_pidx = 0;
663 memcpy(te->te_tcb, tcb, TCB_SIZE);
664 te->te_flags |= TE_ACTIVE;
665 }
666
667 static int
do_get_tcb_rpl(struct sge_iq * iq,const struct rss_header * rss,struct mbuf * m)668 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
669 {
670 struct adapter *sc = iq->adapter;
671 const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *);
672 const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1);
673 struct tcb_histent *te;
674 const u_int tid = GET_TID(cpl);
675 bool remove;
676
677 remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED;
678 te = lookup_tcb_histent(sc, tid, remove);
679 if (te == NULL) {
680 /* Not in the history. Who issued the GET_TCB for this? */
681 device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, "
682 "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid,
683 (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE),
684 GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE),
685 GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie);
686 goto done;
687 }
688
689 MPASS(te->te_flags & TE_RPL_PENDING);
690 te->te_flags &= ~TE_RPL_PENDING;
691 if (remove) {
692 remove_tcb_histent(te);
693 } else {
694 update_tcb_histent(te, tcb);
695 callout_reset(&te->te_callout, hz / 10, request_tcb, te);
696 release_tcb_histent(te);
697 }
698 done:
699 m_freem(m);
700 return (0);
701 }
702
703 static void
fill_tcp_info_from_tcb(struct adapter * sc,uint64_t * tcb,struct tcp_info * ti)704 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti)
705 {
706 uint32_t v;
707
708 ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE);
709
710 v = GET_TCB_FIELD(tcb, T_SRTT);
711 ti->tcpi_rtt = tcp_ticks_to_us(sc, v);
712
713 v = GET_TCB_FIELD(tcb, T_RTTVAR);
714 ti->tcpi_rttvar = tcp_ticks_to_us(sc, v);
715
716 ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH);
717 ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND);
718 ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT);
719
720 v = GET_TCB_FIELD(tcb, TX_MAX);
721 ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW);
722
723 /* Receive window being advertised by us. */
724 ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE); /* Yes, SND. */
725 ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND);
726
727 /* Send window */
728 ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE); /* Yes, RCV. */
729 ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV);
730 if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1))
731 ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale;
732 else
733 ti->tcpi_snd_wscale = 0;
734
735 }
736
737 static void
fill_tcp_info_from_history(struct adapter * sc,struct tcb_histent * te,struct tcp_info * ti)738 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te,
739 struct tcp_info *ti)
740 {
741
742 fill_tcp_info_from_tcb(sc, te->te_tcb, ti);
743 }
744
745 /*
746 * Reads the TCB for the given tid using a memory window and copies it to 'buf'
747 * in the same format as CPL_GET_TCB_RPL.
748 */
749 static void
read_tcb_using_memwin(struct adapter * sc,u_int tid,uint64_t * buf)750 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf)
751 {
752 int i, j, k, rc;
753 uint32_t addr;
754 u_char *tcb, tmp;
755
756 MPASS(tid < sc->tids.ntids);
757
758 addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE;
759 rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE);
760 if (rc != 0)
761 return;
762
763 tcb = (u_char *)buf;
764 for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) {
765 for (k = 0; k < 16; k++) {
766 tmp = tcb[i + k];
767 tcb[i + k] = tcb[j + k];
768 tcb[j + k] = tmp;
769 }
770 }
771 }
772
773 static void
fill_tcp_info(struct adapter * sc,u_int tid,struct tcp_info * ti)774 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti)
775 {
776 uint64_t tcb[TCB_SIZE / sizeof(uint64_t)];
777 struct tcb_histent *te;
778
779 ti->tcpi_toe_tid = tid;
780 te = lookup_tcb_histent(sc, tid, false);
781 if (te != NULL) {
782 fill_tcp_info_from_history(sc, te, ti);
783 release_tcb_histent(te);
784 } else {
785 if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) {
786 /* XXX: tell firmware to flush TCB cache. */
787 }
788 read_tcb_using_memwin(sc, tid, tcb);
789 fill_tcp_info_from_tcb(sc, tcb, ti);
790 }
791 }
792
793 /*
794 * Called by the kernel to allow the TOE driver to "refine" values filled up in
795 * the tcp_info for an offloaded connection.
796 */
797 static void
t4_tcp_info(struct toedev * tod,struct tcpcb * tp,struct tcp_info * ti)798 t4_tcp_info(struct toedev *tod, struct tcpcb *tp, struct tcp_info *ti)
799 {
800 struct adapter *sc = tod->tod_softc;
801 struct toepcb *toep = tp->t_toe;
802
803 INP_WLOCK_ASSERT(tp->t_inpcb);
804 MPASS(ti != NULL);
805
806 fill_tcp_info(sc, toep->tid, ti);
807 }
808
809 /*
810 * The TOE driver will not receive any more CPLs for the tid associated with the
811 * toepcb; release the hold on the inpcb.
812 */
813 void
final_cpl_received(struct toepcb * toep)814 final_cpl_received(struct toepcb *toep)
815 {
816 struct inpcb *inp = toep->inp;
817
818 KASSERT(inp != NULL, ("%s: inp is NULL", __func__));
819 INP_WLOCK_ASSERT(inp);
820 KASSERT(toep->flags & TPF_CPL_PENDING,
821 ("%s: CPL not pending already?", __func__));
822
823 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)",
824 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags);
825
826 if (ulp_mode(toep) == ULP_MODE_TCPDDP)
827 release_ddp_resources(toep);
828 toep->inp = NULL;
829 toep->flags &= ~TPF_CPL_PENDING;
830 mbufq_drain(&toep->ulp_pdu_reclaimq);
831
832 if (!(toep->flags & TPF_ATTACHED))
833 release_offload_resources(toep);
834
835 if (!in_pcbrele_wlocked(inp))
836 INP_WUNLOCK(inp);
837 }
838
839 void
insert_tid(struct adapter * sc,int tid,void * ctx,int ntids)840 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids)
841 {
842 struct tid_info *t = &sc->tids;
843
844 MPASS(tid >= t->tid_base);
845 MPASS(tid - t->tid_base < t->ntids);
846
847 t->tid_tab[tid - t->tid_base] = ctx;
848 atomic_add_int(&t->tids_in_use, ntids);
849 }
850
851 void *
lookup_tid(struct adapter * sc,int tid)852 lookup_tid(struct adapter *sc, int tid)
853 {
854 struct tid_info *t = &sc->tids;
855
856 return (t->tid_tab[tid - t->tid_base]);
857 }
858
859 void
update_tid(struct adapter * sc,int tid,void * ctx)860 update_tid(struct adapter *sc, int tid, void *ctx)
861 {
862 struct tid_info *t = &sc->tids;
863
864 t->tid_tab[tid - t->tid_base] = ctx;
865 }
866
867 void
remove_tid(struct adapter * sc,int tid,int ntids)868 remove_tid(struct adapter *sc, int tid, int ntids)
869 {
870 struct tid_info *t = &sc->tids;
871
872 t->tid_tab[tid - t->tid_base] = NULL;
873 atomic_subtract_int(&t->tids_in_use, ntids);
874 }
875
876 /*
877 * What mtu_idx to use, given a 4-tuple. Note that both s->mss and tcp_mssopt
878 * have the MSS that we should advertise in our SYN. Advertised MSS doesn't
879 * account for any TCP options so the effective MSS (only payload, no headers or
880 * options) could be different.
881 */
882 static int
find_best_mtu_idx(struct adapter * sc,struct in_conninfo * inc,struct offload_settings * s)883 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc,
884 struct offload_settings *s)
885 {
886 unsigned short *mtus = &sc->params.mtus[0];
887 int i, mss, mtu;
888
889 MPASS(inc != NULL);
890
891 mss = s->mss > 0 ? s->mss : tcp_mssopt(inc);
892 if (inc->inc_flags & INC_ISIPV6)
893 mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
894 else
895 mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr);
896
897 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++)
898 continue;
899
900 return (i);
901 }
902
903 /*
904 * Determine the receive window size for a socket.
905 */
906 u_long
select_rcv_wnd(struct socket * so)907 select_rcv_wnd(struct socket *so)
908 {
909 unsigned long wnd;
910
911 SOCKBUF_LOCK_ASSERT(&so->so_rcv);
912
913 wnd = sbspace(&so->so_rcv);
914 if (wnd < MIN_RCV_WND)
915 wnd = MIN_RCV_WND;
916
917 return min(wnd, MAX_RCV_WND);
918 }
919
920 int
select_rcv_wscale(void)921 select_rcv_wscale(void)
922 {
923 int wscale = 0;
924 unsigned long space = sb_max;
925
926 if (space > MAX_RCV_WND)
927 space = MAX_RCV_WND;
928
929 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space)
930 wscale++;
931
932 return (wscale);
933 }
934
935 __be64
calc_options0(struct vi_info * vi,struct conn_params * cp)936 calc_options0(struct vi_info *vi, struct conn_params *cp)
937 {
938 uint64_t opt0 = 0;
939
940 opt0 |= F_TCAM_BYPASS;
941
942 MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE);
943 opt0 |= V_WND_SCALE(cp->wscale);
944
945 MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS);
946 opt0 |= V_MSS_IDX(cp->mtu_idx);
947
948 MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE);
949 opt0 |= V_ULP_MODE(cp->ulp_mode);
950
951 MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ);
952 opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize);
953
954 MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->pi->adapter->vres.l2t.size);
955 opt0 |= V_L2T_IDX(cp->l2t_idx);
956
957 opt0 |= V_SMAC_SEL(vi->smt_idx);
958 opt0 |= V_TX_CHAN(vi->pi->tx_chan);
959
960 MPASS(cp->keepalive == 0 || cp->keepalive == 1);
961 opt0 |= V_KEEP_ALIVE(cp->keepalive);
962
963 MPASS(cp->nagle == 0 || cp->nagle == 1);
964 opt0 |= V_NAGLE(cp->nagle);
965
966 return (htobe64(opt0));
967 }
968
969 __be32
calc_options2(struct vi_info * vi,struct conn_params * cp)970 calc_options2(struct vi_info *vi, struct conn_params *cp)
971 {
972 uint32_t opt2 = 0;
973 struct port_info *pi = vi->pi;
974 struct adapter *sc = pi->adapter;
975
976 /*
977 * rx flow control, rx coalesce, congestion control, and tx pace are all
978 * explicitly set by the driver. On T5+ the ISS is also set by the
979 * driver to the value picked by the kernel.
980 */
981 if (is_t4(sc)) {
982 opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID;
983 opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID;
984 } else {
985 opt2 |= F_T5_OPT_2_VALID; /* all 4 valid */
986 opt2 |= F_T5_ISS; /* ISS provided in CPL */
987 }
988
989 MPASS(cp->sack == 0 || cp->sack == 1);
990 opt2 |= V_SACK_EN(cp->sack);
991
992 MPASS(cp->tstamp == 0 || cp->tstamp == 1);
993 opt2 |= V_TSTAMPS_EN(cp->tstamp);
994
995 if (cp->wscale > 0)
996 opt2 |= F_WND_SCALE_EN;
997
998 MPASS(cp->ecn == 0 || cp->ecn == 1);
999 opt2 |= V_CCTRL_ECN(cp->ecn);
1000
1001 /* XXX: F_RX_CHANNEL for multiple rx c-chan support goes here. */
1002
1003 opt2 |= V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]);
1004 opt2 |= V_PACE(0);
1005 opt2 |= F_RSS_QUEUE_VALID;
1006 opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id);
1007
1008 MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL);
1009 opt2 |= V_CONG_CNTRL(cp->cong_algo);
1010
1011 MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1);
1012 if (cp->rx_coalesce == 1)
1013 opt2 |= V_RX_COALESCE(M_RX_COALESCE);
1014
1015 opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0);
1016 #ifdef USE_DDP_RX_FLOW_CONTROL
1017 if (cp->ulp_mode == ULP_MODE_TCPDDP)
1018 opt2 |= F_RX_FC_DDP;
1019 #endif
1020 if (cp->ulp_mode == ULP_MODE_TLS)
1021 opt2 |= F_RX_FC_DISABLE;
1022
1023 return (htobe32(opt2));
1024 }
1025
1026 uint64_t
select_ntuple(struct vi_info * vi,struct l2t_entry * e)1027 select_ntuple(struct vi_info *vi, struct l2t_entry *e)
1028 {
1029 struct adapter *sc = vi->pi->adapter;
1030 struct tp_params *tp = &sc->params.tp;
1031 uint64_t ntuple = 0;
1032
1033 /*
1034 * Initialize each of the fields which we care about which are present
1035 * in the Compressed Filter Tuple.
1036 */
1037 if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE)
1038 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift;
1039
1040 if (tp->port_shift >= 0)
1041 ntuple |= (uint64_t)e->lport << tp->port_shift;
1042
1043 if (tp->protocol_shift >= 0)
1044 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift;
1045
1046 if (tp->vnic_shift >= 0 && tp->ingress_config & F_VNIC) {
1047 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) |
1048 V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) <<
1049 tp->vnic_shift;
1050 }
1051
1052 if (is_t4(sc))
1053 return (htobe32((uint32_t)ntuple));
1054 else
1055 return (htobe64(V_FILTER_TUPLE(ntuple)));
1056 }
1057
1058 static int
is_tls_sock(struct socket * so,struct adapter * sc)1059 is_tls_sock(struct socket *so, struct adapter *sc)
1060 {
1061 struct inpcb *inp = sotoinpcb(so);
1062 int i, rc;
1063
1064 /* XXX: Eventually add a SO_WANT_TLS socket option perhaps? */
1065 rc = 0;
1066 ADAPTER_LOCK(sc);
1067 for (i = 0; i < sc->tt.num_tls_rx_ports; i++) {
1068 if (inp->inp_lport == htons(sc->tt.tls_rx_ports[i]) ||
1069 inp->inp_fport == htons(sc->tt.tls_rx_ports[i])) {
1070 rc = 1;
1071 break;
1072 }
1073 }
1074 ADAPTER_UNLOCK(sc);
1075 return (rc);
1076 }
1077
1078 /*
1079 * Initialize various connection parameters.
1080 */
1081 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)1082 init_conn_params(struct vi_info *vi , struct offload_settings *s,
1083 struct in_conninfo *inc, struct socket *so,
1084 const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp)
1085 {
1086 struct port_info *pi = vi->pi;
1087 struct adapter *sc = pi->adapter;
1088 struct tom_tunables *tt = &sc->tt;
1089 struct inpcb *inp = sotoinpcb(so);
1090 struct tcpcb *tp = intotcpcb(inp);
1091 u_long wnd;
1092
1093 MPASS(s->offload != 0);
1094
1095 /* Congestion control algorithm */
1096 if (s->cong_algo >= 0)
1097 cp->cong_algo = s->cong_algo & M_CONG_CNTRL;
1098 else if (sc->tt.cong_algorithm >= 0)
1099 cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL;
1100 else {
1101 struct cc_algo *cc = CC_ALGO(tp);
1102
1103 if (strcasecmp(cc->name, "reno") == 0)
1104 cp->cong_algo = CONG_ALG_RENO;
1105 else if (strcasecmp(cc->name, "tahoe") == 0)
1106 cp->cong_algo = CONG_ALG_TAHOE;
1107 if (strcasecmp(cc->name, "newreno") == 0)
1108 cp->cong_algo = CONG_ALG_NEWRENO;
1109 if (strcasecmp(cc->name, "highspeed") == 0)
1110 cp->cong_algo = CONG_ALG_HIGHSPEED;
1111 else {
1112 /*
1113 * Use newreno in case the algorithm selected by the
1114 * host stack is not supported by the hardware.
1115 */
1116 cp->cong_algo = CONG_ALG_NEWRENO;
1117 }
1118 }
1119
1120 /* Tx traffic scheduling class. */
1121 if (s->sched_class >= 0 &&
1122 s->sched_class < sc->chip_params->nsched_cls) {
1123 cp->tc_idx = s->sched_class;
1124 } else
1125 cp->tc_idx = -1;
1126
1127 /* Nagle's algorithm. */
1128 if (s->nagle >= 0)
1129 cp->nagle = s->nagle > 0 ? 1 : 0;
1130 else
1131 cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1;
1132
1133 /* TCP Keepalive. */
1134 if (tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE)
1135 cp->keepalive = 1;
1136 else
1137 cp->keepalive = 0;
1138
1139 /* Optimization that's specific to T5 @ 40G. */
1140 if (tt->tx_align >= 0)
1141 cp->tx_align = tt->tx_align > 0 ? 1 : 0;
1142 else if (chip_id(sc) == CHELSIO_T5 &&
1143 (port_top_speed(pi) > 10 || sc->params.nports > 2))
1144 cp->tx_align = 1;
1145 else
1146 cp->tx_align = 0;
1147
1148 /* ULP mode. */
1149 if (can_tls_offload(sc) &&
1150 (s->tls > 0 || (s->tls < 0 && is_tls_sock(so, sc))))
1151 cp->ulp_mode = ULP_MODE_TLS;
1152 else if (s->ddp > 0 ||
1153 (s->ddp < 0 && sc->tt.ddp && (so_options_get(so) & SO_NO_DDP) == 0))
1154 cp->ulp_mode = ULP_MODE_TCPDDP;
1155 else
1156 cp->ulp_mode = ULP_MODE_NONE;
1157
1158 /* Rx coalescing. */
1159 if (s->rx_coalesce >= 0)
1160 cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0;
1161 else if (cp->ulp_mode == ULP_MODE_TLS)
1162 cp->rx_coalesce = 0;
1163 else if (tt->rx_coalesce >= 0)
1164 cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0;
1165 else
1166 cp->rx_coalesce = 1; /* default */
1167
1168 /*
1169 * Index in the PMTU table. This controls the MSS that we announce in
1170 * our SYN initially, but after ESTABLISHED it controls the MSS that we
1171 * use to send data.
1172 */
1173 cp->mtu_idx = find_best_mtu_idx(sc, inc, s);
1174
1175 /* Tx queue for this connection. */
1176 if (s->txq >= 0 && s->txq < vi->nofldtxq)
1177 cp->txq_idx = s->txq;
1178 else
1179 cp->txq_idx = arc4random() % vi->nofldtxq;
1180 cp->txq_idx += vi->first_ofld_txq;
1181
1182 /* Rx queue for this connection. */
1183 if (s->rxq >= 0 && s->rxq < vi->nofldrxq)
1184 cp->rxq_idx = s->rxq;
1185 else
1186 cp->rxq_idx = arc4random() % vi->nofldrxq;
1187 cp->rxq_idx += vi->first_ofld_rxq;
1188
1189 if (SOLISTENING(so)) {
1190 /* Passive open */
1191 MPASS(tcpopt != NULL);
1192
1193 /* TCP timestamp option */
1194 if (tcpopt->tstamp &&
1195 (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323)))
1196 cp->tstamp = 1;
1197 else
1198 cp->tstamp = 0;
1199
1200 /* SACK */
1201 if (tcpopt->sack &&
1202 (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack)))
1203 cp->sack = 1;
1204 else
1205 cp->sack = 0;
1206
1207 /* Receive window scaling. */
1208 if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323)
1209 cp->wscale = select_rcv_wscale();
1210 else
1211 cp->wscale = 0;
1212
1213 /* ECN */
1214 if (tcpopt->ecn && /* XXX: review. */
1215 (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn)))
1216 cp->ecn = 1;
1217 else
1218 cp->ecn = 0;
1219
1220 wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND);
1221 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1222
1223 if (tt->sndbuf > 0)
1224 cp->sndbuf = tt->sndbuf;
1225 else if (so->sol_sbsnd_flags & SB_AUTOSIZE &&
1226 V_tcp_do_autosndbuf)
1227 cp->sndbuf = 256 * 1024;
1228 else
1229 cp->sndbuf = so->sol_sbsnd_hiwat;
1230 } else {
1231 /* Active open */
1232
1233 /* TCP timestamp option */
1234 if (s->tstamp > 0 ||
1235 (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP)))
1236 cp->tstamp = 1;
1237 else
1238 cp->tstamp = 0;
1239
1240 /* SACK */
1241 if (s->sack > 0 ||
1242 (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT)))
1243 cp->sack = 1;
1244 else
1245 cp->sack = 0;
1246
1247 /* Receive window scaling */
1248 if (tp->t_flags & TF_REQ_SCALE)
1249 cp->wscale = select_rcv_wscale();
1250 else
1251 cp->wscale = 0;
1252
1253 /* ECN */
1254 if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1))
1255 cp->ecn = 1;
1256 else
1257 cp->ecn = 0;
1258
1259 SOCKBUF_LOCK(&so->so_rcv);
1260 wnd = max(select_rcv_wnd(so), MIN_RCV_WND);
1261 SOCKBUF_UNLOCK(&so->so_rcv);
1262 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ);
1263
1264 if (tt->sndbuf > 0)
1265 cp->sndbuf = tt->sndbuf;
1266 else {
1267 SOCKBUF_LOCK(&so->so_snd);
1268 if (so->so_snd.sb_flags & SB_AUTOSIZE &&
1269 V_tcp_do_autosndbuf)
1270 cp->sndbuf = 256 * 1024;
1271 else
1272 cp->sndbuf = so->so_snd.sb_hiwat;
1273 SOCKBUF_UNLOCK(&so->so_snd);
1274 }
1275 }
1276
1277 cp->l2t_idx = l2t_idx;
1278
1279 /* This will be initialized on ESTABLISHED. */
1280 cp->emss = 0;
1281 }
1282
1283 int
negative_advice(int status)1284 negative_advice(int status)
1285 {
1286
1287 return (status == CPL_ERR_RTX_NEG_ADVICE ||
1288 status == CPL_ERR_PERSIST_NEG_ADVICE ||
1289 status == CPL_ERR_KEEPALV_NEG_ADVICE);
1290 }
1291
1292 static int
alloc_tid_tab(struct tid_info * t,int flags)1293 alloc_tid_tab(struct tid_info *t, int flags)
1294 {
1295
1296 MPASS(t->ntids > 0);
1297 MPASS(t->tid_tab == NULL);
1298
1299 t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE,
1300 M_ZERO | flags);
1301 if (t->tid_tab == NULL)
1302 return (ENOMEM);
1303 atomic_store_rel_int(&t->tids_in_use, 0);
1304
1305 return (0);
1306 }
1307
1308 static void
free_tid_tab(struct tid_info * t)1309 free_tid_tab(struct tid_info *t)
1310 {
1311
1312 KASSERT(t->tids_in_use == 0,
1313 ("%s: %d tids still in use.", __func__, t->tids_in_use));
1314
1315 free(t->tid_tab, M_CXGBE);
1316 t->tid_tab = NULL;
1317 }
1318
1319 static int
alloc_stid_tab(struct tid_info * t,int flags)1320 alloc_stid_tab(struct tid_info *t, int flags)
1321 {
1322
1323 MPASS(t->nstids > 0);
1324 MPASS(t->stid_tab == NULL);
1325
1326 t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE,
1327 M_ZERO | flags);
1328 if (t->stid_tab == NULL)
1329 return (ENOMEM);
1330 mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF);
1331 t->stids_in_use = 0;
1332 TAILQ_INIT(&t->stids);
1333 t->nstids_free_head = t->nstids;
1334
1335 return (0);
1336 }
1337
1338 static void
free_stid_tab(struct tid_info * t)1339 free_stid_tab(struct tid_info *t)
1340 {
1341
1342 KASSERT(t->stids_in_use == 0,
1343 ("%s: %d tids still in use.", __func__, t->stids_in_use));
1344
1345 if (mtx_initialized(&t->stid_lock))
1346 mtx_destroy(&t->stid_lock);
1347 free(t->stid_tab, M_CXGBE);
1348 t->stid_tab = NULL;
1349 }
1350
1351 static void
free_tid_tabs(struct tid_info * t)1352 free_tid_tabs(struct tid_info *t)
1353 {
1354
1355 free_tid_tab(t);
1356 free_atid_tab(t);
1357 free_stid_tab(t);
1358 }
1359
1360 static int
alloc_tid_tabs(struct tid_info * t)1361 alloc_tid_tabs(struct tid_info *t)
1362 {
1363 int rc;
1364
1365 rc = alloc_tid_tab(t, M_NOWAIT);
1366 if (rc != 0)
1367 goto failed;
1368
1369 rc = alloc_atid_tab(t, M_NOWAIT);
1370 if (rc != 0)
1371 goto failed;
1372
1373 rc = alloc_stid_tab(t, M_NOWAIT);
1374 if (rc != 0)
1375 goto failed;
1376
1377 return (0);
1378 failed:
1379 free_tid_tabs(t);
1380 return (rc);
1381 }
1382
1383 static inline void
alloc_tcb_history(struct adapter * sc,struct tom_data * td)1384 alloc_tcb_history(struct adapter *sc, struct tom_data *td)
1385 {
1386
1387 if (sc->tids.ntids == 0 || sc->tids.ntids > 1024)
1388 return;
1389 rw_init(&td->tcb_history_lock, "TCB history");
1390 td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history),
1391 M_CXGBE, M_ZERO | M_NOWAIT);
1392 td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0));
1393 }
1394
1395 static inline void
free_tcb_history(struct adapter * sc,struct tom_data * td)1396 free_tcb_history(struct adapter *sc, struct tom_data *td)
1397 {
1398 #ifdef INVARIANTS
1399 int i;
1400
1401 if (td->tcb_history != NULL) {
1402 for (i = 0; i < sc->tids.ntids; i++) {
1403 MPASS(td->tcb_history[i] == NULL);
1404 }
1405 }
1406 #endif
1407 free(td->tcb_history, M_CXGBE);
1408 if (rw_initialized(&td->tcb_history_lock))
1409 rw_destroy(&td->tcb_history_lock);
1410 }
1411
1412 static void
free_tom_data(struct adapter * sc,struct tom_data * td)1413 free_tom_data(struct adapter *sc, struct tom_data *td)
1414 {
1415
1416 ASSERT_SYNCHRONIZED_OP(sc);
1417
1418 KASSERT(TAILQ_EMPTY(&td->toep_list),
1419 ("%s: TOE PCB list is not empty.", __func__));
1420 KASSERT(td->lctx_count == 0,
1421 ("%s: lctx hash table is not empty.", __func__));
1422
1423 t4_free_ppod_region(&td->pr);
1424
1425 if (td->listen_mask != 0)
1426 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask);
1427
1428 if (mtx_initialized(&td->unsent_wr_lock))
1429 mtx_destroy(&td->unsent_wr_lock);
1430 if (mtx_initialized(&td->lctx_hash_lock))
1431 mtx_destroy(&td->lctx_hash_lock);
1432 if (mtx_initialized(&td->toep_list_lock))
1433 mtx_destroy(&td->toep_list_lock);
1434
1435 free_tcb_history(sc, td);
1436 free_tid_tabs(&sc->tids);
1437 free(td, M_CXGBE);
1438 }
1439
1440 static char *
prepare_pkt(int open_type,uint16_t vtag,struct inpcb * inp,int * pktlen,int * buflen)1441 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen,
1442 int *buflen)
1443 {
1444 char *pkt;
1445 struct tcphdr *th;
1446 int ipv6, len;
1447 const int maxlen =
1448 max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) +
1449 max(sizeof(struct ip), sizeof(struct ip6_hdr)) +
1450 sizeof(struct tcphdr);
1451
1452 MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN);
1453
1454 pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT);
1455 if (pkt == NULL)
1456 return (NULL);
1457
1458 ipv6 = inp->inp_vflag & INP_IPV6;
1459 len = 0;
1460
1461 if (EVL_VLANOFTAG(vtag) == 0xfff) {
1462 struct ether_header *eh = (void *)pkt;
1463
1464 if (ipv6)
1465 eh->ether_type = htons(ETHERTYPE_IPV6);
1466 else
1467 eh->ether_type = htons(ETHERTYPE_IP);
1468
1469 len += sizeof(*eh);
1470 } else {
1471 struct ether_vlan_header *evh = (void *)pkt;
1472
1473 evh->evl_encap_proto = htons(ETHERTYPE_VLAN);
1474 evh->evl_tag = htons(vtag);
1475 if (ipv6)
1476 evh->evl_proto = htons(ETHERTYPE_IPV6);
1477 else
1478 evh->evl_proto = htons(ETHERTYPE_IP);
1479
1480 len += sizeof(*evh);
1481 }
1482
1483 if (ipv6) {
1484 struct ip6_hdr *ip6 = (void *)&pkt[len];
1485
1486 ip6->ip6_vfc = IPV6_VERSION;
1487 ip6->ip6_plen = htons(sizeof(struct tcphdr));
1488 ip6->ip6_nxt = IPPROTO_TCP;
1489 if (open_type == OPEN_TYPE_ACTIVE) {
1490 ip6->ip6_src = inp->in6p_laddr;
1491 ip6->ip6_dst = inp->in6p_faddr;
1492 } else if (open_type == OPEN_TYPE_LISTEN) {
1493 ip6->ip6_src = inp->in6p_laddr;
1494 ip6->ip6_dst = ip6->ip6_src;
1495 }
1496
1497 len += sizeof(*ip6);
1498 } else {
1499 struct ip *ip = (void *)&pkt[len];
1500
1501 ip->ip_v = IPVERSION;
1502 ip->ip_hl = sizeof(*ip) >> 2;
1503 ip->ip_tos = inp->inp_ip_tos;
1504 ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr));
1505 ip->ip_ttl = inp->inp_ip_ttl;
1506 ip->ip_p = IPPROTO_TCP;
1507 if (open_type == OPEN_TYPE_ACTIVE) {
1508 ip->ip_src = inp->inp_laddr;
1509 ip->ip_dst = inp->inp_faddr;
1510 } else if (open_type == OPEN_TYPE_LISTEN) {
1511 ip->ip_src = inp->inp_laddr;
1512 ip->ip_dst = ip->ip_src;
1513 }
1514
1515 len += sizeof(*ip);
1516 }
1517
1518 th = (void *)&pkt[len];
1519 if (open_type == OPEN_TYPE_ACTIVE) {
1520 th->th_sport = inp->inp_lport; /* network byte order already */
1521 th->th_dport = inp->inp_fport; /* ditto */
1522 } else if (open_type == OPEN_TYPE_LISTEN) {
1523 th->th_sport = inp->inp_lport; /* network byte order already */
1524 th->th_dport = th->th_sport;
1525 }
1526 len += sizeof(th);
1527
1528 *pktlen = *buflen = len;
1529 return (pkt);
1530 }
1531
1532 const struct offload_settings *
lookup_offload_policy(struct adapter * sc,int open_type,struct mbuf * m,uint16_t vtag,struct inpcb * inp)1533 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m,
1534 uint16_t vtag, struct inpcb *inp)
1535 {
1536 const struct t4_offload_policy *op;
1537 char *pkt;
1538 struct offload_rule *r;
1539 int i, matched, pktlen, buflen;
1540 static const struct offload_settings allow_offloading_settings = {
1541 .offload = 1,
1542 .rx_coalesce = -1,
1543 .cong_algo = -1,
1544 .sched_class = -1,
1545 .tstamp = -1,
1546 .sack = -1,
1547 .nagle = -1,
1548 .ecn = -1,
1549 .ddp = -1,
1550 .tls = -1,
1551 .txq = -1,
1552 .rxq = -1,
1553 .mss = -1,
1554 };
1555 static const struct offload_settings disallow_offloading_settings = {
1556 .offload = 0,
1557 /* rest is irrelevant when offload is off. */
1558 };
1559
1560 rw_assert(&sc->policy_lock, RA_LOCKED);
1561
1562 /*
1563 * If there's no Connection Offloading Policy attached to the device
1564 * then we need to return a default static policy. If
1565 * "cop_managed_offloading" is true, then we need to disallow
1566 * offloading until a COP is attached to the device. Otherwise we
1567 * allow offloading ...
1568 */
1569 op = sc->policy;
1570 if (op == NULL) {
1571 if (sc->tt.cop_managed_offloading)
1572 return (&disallow_offloading_settings);
1573 else
1574 return (&allow_offloading_settings);
1575 }
1576
1577 switch (open_type) {
1578 case OPEN_TYPE_ACTIVE:
1579 case OPEN_TYPE_LISTEN:
1580 pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen);
1581 break;
1582 case OPEN_TYPE_PASSIVE:
1583 MPASS(m != NULL);
1584 pkt = mtod(m, char *);
1585 MPASS(*pkt == CPL_PASS_ACCEPT_REQ);
1586 pkt += sizeof(struct cpl_pass_accept_req);
1587 pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req);
1588 buflen = m->m_len - sizeof(struct cpl_pass_accept_req);
1589 break;
1590 default:
1591 MPASS(0);
1592 return (&disallow_offloading_settings);
1593 }
1594
1595 if (pkt == NULL || pktlen == 0 || buflen == 0)
1596 return (&disallow_offloading_settings);
1597
1598 matched = 0;
1599 r = &op->rule[0];
1600 for (i = 0; i < op->nrules; i++, r++) {
1601 if (r->open_type != open_type &&
1602 r->open_type != OPEN_TYPE_DONTCARE) {
1603 continue;
1604 }
1605 matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen);
1606 if (matched)
1607 break;
1608 }
1609
1610 if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN)
1611 free(pkt, M_CXGBE);
1612
1613 return (matched ? &r->settings : &disallow_offloading_settings);
1614 }
1615
1616 static void
reclaim_wr_resources(void * arg,int count)1617 reclaim_wr_resources(void *arg, int count)
1618 {
1619 struct tom_data *td = arg;
1620 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list);
1621 struct cpl_act_open_req *cpl;
1622 u_int opcode, atid, tid;
1623 struct wrqe *wr;
1624 struct adapter *sc = td_adapter(td);
1625
1626 mtx_lock(&td->unsent_wr_lock);
1627 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe);
1628 mtx_unlock(&td->unsent_wr_lock);
1629
1630 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) {
1631 STAILQ_REMOVE_HEAD(&twr_list, link);
1632
1633 cpl = wrtod(wr);
1634 opcode = GET_OPCODE(cpl);
1635
1636 switch (opcode) {
1637 case CPL_ACT_OPEN_REQ:
1638 case CPL_ACT_OPEN_REQ6:
1639 atid = G_TID_TID(be32toh(OPCODE_TID(cpl)));
1640 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid);
1641 act_open_failure_cleanup(sc, atid, EHOSTUNREACH);
1642 free(wr, M_CXGBE);
1643 break;
1644 case CPL_PASS_ACCEPT_RPL:
1645 tid = GET_TID(cpl);
1646 CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid);
1647 synack_failure_cleanup(sc, tid);
1648 free(wr, M_CXGBE);
1649 break;
1650 default:
1651 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, "
1652 "opcode %x\n", __func__, wr, wr->wr_len, opcode);
1653 /* WR not freed here; go look at it with a debugger. */
1654 }
1655 }
1656 }
1657
1658 /*
1659 * Ground control to Major TOM
1660 * Commencing countdown, engines on
1661 */
1662 static int
t4_tom_activate(struct adapter * sc)1663 t4_tom_activate(struct adapter *sc)
1664 {
1665 struct tom_data *td;
1666 struct toedev *tod;
1667 struct vi_info *vi;
1668 int i, rc, v;
1669
1670 ASSERT_SYNCHRONIZED_OP(sc);
1671
1672 /* per-adapter softc for TOM */
1673 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT);
1674 if (td == NULL)
1675 return (ENOMEM);
1676
1677 /* List of TOE PCBs and associated lock */
1678 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF);
1679 TAILQ_INIT(&td->toep_list);
1680
1681 /* Listen context */
1682 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF);
1683 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE,
1684 &td->listen_mask, HASH_NOWAIT);
1685
1686 /* List of WRs for which L2 resolution failed */
1687 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF);
1688 STAILQ_INIT(&td->unsent_wr_list);
1689 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td);
1690
1691 /* TID tables */
1692 rc = alloc_tid_tabs(&sc->tids);
1693 if (rc != 0)
1694 goto done;
1695
1696 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp,
1697 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods");
1698 if (rc != 0)
1699 goto done;
1700 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK,
1701 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask);
1702
1703 alloc_tcb_history(sc, td);
1704
1705 /* toedev ops */
1706 tod = &td->tod;
1707 init_toedev(tod);
1708 tod->tod_softc = sc;
1709 tod->tod_connect = t4_connect;
1710 tod->tod_listen_start = t4_listen_start;
1711 tod->tod_listen_stop = t4_listen_stop;
1712 tod->tod_rcvd = t4_rcvd;
1713 tod->tod_output = t4_tod_output;
1714 tod->tod_send_rst = t4_send_rst;
1715 tod->tod_send_fin = t4_send_fin;
1716 tod->tod_pcb_detach = t4_pcb_detach;
1717 tod->tod_l2_update = t4_l2_update;
1718 tod->tod_syncache_added = t4_syncache_added;
1719 tod->tod_syncache_removed = t4_syncache_removed;
1720 tod->tod_syncache_respond = t4_syncache_respond;
1721 tod->tod_offload_socket = t4_offload_socket;
1722 tod->tod_ctloutput = t4_ctloutput;
1723 tod->tod_tcp_info = t4_tcp_info;
1724
1725 for_each_port(sc, i) {
1726 for_each_vi(sc->port[i], v, vi) {
1727 TOEDEV(vi->ifp) = &td->tod;
1728 }
1729 }
1730
1731 sc->tom_softc = td;
1732 register_toedev(sc->tom_softc);
1733
1734 done:
1735 if (rc != 0)
1736 free_tom_data(sc, td);
1737 return (rc);
1738 }
1739
1740 static int
t4_tom_deactivate(struct adapter * sc)1741 t4_tom_deactivate(struct adapter *sc)
1742 {
1743 int rc = 0;
1744 struct tom_data *td = sc->tom_softc;
1745
1746 ASSERT_SYNCHRONIZED_OP(sc);
1747
1748 if (td == NULL)
1749 return (0); /* XXX. KASSERT? */
1750
1751 if (sc->offload_map != 0)
1752 return (EBUSY); /* at least one port has IFCAP_TOE enabled */
1753
1754 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI))
1755 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */
1756
1757 mtx_lock(&td->toep_list_lock);
1758 if (!TAILQ_EMPTY(&td->toep_list))
1759 rc = EBUSY;
1760 mtx_unlock(&td->toep_list_lock);
1761
1762 mtx_lock(&td->lctx_hash_lock);
1763 if (td->lctx_count > 0)
1764 rc = EBUSY;
1765 mtx_unlock(&td->lctx_hash_lock);
1766
1767 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources);
1768 mtx_lock(&td->unsent_wr_lock);
1769 if (!STAILQ_EMPTY(&td->unsent_wr_list))
1770 rc = EBUSY;
1771 mtx_unlock(&td->unsent_wr_lock);
1772
1773 if (rc == 0) {
1774 unregister_toedev(sc->tom_softc);
1775 free_tom_data(sc, td);
1776 sc->tom_softc = NULL;
1777 }
1778
1779 return (rc);
1780 }
1781
1782 static int
t4_aio_queue_tom(struct socket * so,struct kaiocb * job)1783 t4_aio_queue_tom(struct socket *so, struct kaiocb *job)
1784 {
1785 struct tcpcb *tp = so_sototcpcb(so);
1786 struct toepcb *toep = tp->t_toe;
1787 int error;
1788
1789 if (ulp_mode(toep) == ULP_MODE_TCPDDP) {
1790 error = t4_aio_queue_ddp(so, job);
1791 if (error != EOPNOTSUPP)
1792 return (error);
1793 }
1794
1795 return (t4_aio_queue_aiotx(so, job));
1796 }
1797
1798 static int
t4_ctloutput_tom(struct socket * so,struct sockopt * sopt)1799 t4_ctloutput_tom(struct socket *so, struct sockopt *sopt)
1800 {
1801
1802 if (sopt->sopt_level != IPPROTO_TCP)
1803 return (tcp_ctloutput(so, sopt));
1804
1805 switch (sopt->sopt_name) {
1806 case TCP_TLSOM_SET_TLS_CONTEXT:
1807 case TCP_TLSOM_GET_TLS_TOM:
1808 case TCP_TLSOM_CLR_TLS_TOM:
1809 case TCP_TLSOM_CLR_QUIES:
1810 return (t4_ctloutput_tls(so, sopt));
1811 default:
1812 return (tcp_ctloutput(so, sopt));
1813 }
1814 }
1815
1816 static int
t4_tom_mod_load(void)1817 t4_tom_mod_load(void)
1818 {
1819 struct protosw *tcp_protosw, *tcp6_protosw;
1820
1821 /* CPL handlers */
1822 t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl);
1823 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2,
1824 CPL_COOKIE_TOM);
1825 t4_init_connect_cpl_handlers();
1826 t4_init_listen_cpl_handlers();
1827 t4_init_cpl_io_handlers();
1828
1829 t4_ddp_mod_load();
1830 t4_tls_mod_load();
1831
1832 tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM);
1833 if (tcp_protosw == NULL)
1834 return (ENOPROTOOPT);
1835 bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw));
1836 bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs));
1837 toe_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1838 toe_protosw.pr_ctloutput = t4_ctloutput_tom;
1839 toe_protosw.pr_usrreqs = &toe_usrreqs;
1840
1841 tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM);
1842 if (tcp6_protosw == NULL)
1843 return (ENOPROTOOPT);
1844 bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw));
1845 bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs));
1846 toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom;
1847 toe6_protosw.pr_ctloutput = t4_ctloutput_tom;
1848 toe6_protosw.pr_usrreqs = &toe6_usrreqs;
1849
1850 return (t4_register_uld(&tom_uld_info));
1851 }
1852
1853 static void
tom_uninit(struct adapter * sc,void * arg __unused)1854 tom_uninit(struct adapter *sc, void *arg __unused)
1855 {
1856 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun"))
1857 return;
1858
1859 /* Try to free resources (works only if no port has IFCAP_TOE) */
1860 if (uld_active(sc, ULD_TOM))
1861 t4_deactivate_uld(sc, ULD_TOM);
1862
1863 end_synchronized_op(sc, 0);
1864 }
1865
1866 static int
t4_tom_mod_unload(void)1867 t4_tom_mod_unload(void)
1868 {
1869 t4_iterate(tom_uninit, NULL);
1870
1871 if (t4_unregister_uld(&tom_uld_info) == EBUSY)
1872 return (EBUSY);
1873
1874 t4_tls_mod_unload();
1875 t4_ddp_mod_unload();
1876
1877 t4_uninit_connect_cpl_handlers();
1878 t4_uninit_listen_cpl_handlers();
1879 t4_uninit_cpl_io_handlers();
1880 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM);
1881
1882 return (0);
1883 }
1884 #endif /* TCP_OFFLOAD */
1885
1886 static int
t4_tom_modevent(module_t mod,int cmd,void * arg)1887 t4_tom_modevent(module_t mod, int cmd, void *arg)
1888 {
1889 int rc = 0;
1890
1891 #ifdef TCP_OFFLOAD
1892 switch (cmd) {
1893 case MOD_LOAD:
1894 rc = t4_tom_mod_load();
1895 break;
1896
1897 case MOD_UNLOAD:
1898 rc = t4_tom_mod_unload();
1899 break;
1900
1901 default:
1902 rc = EINVAL;
1903 }
1904 #else
1905 printf("t4_tom: compiled without TCP_OFFLOAD support.\n");
1906 rc = EOPNOTSUPP;
1907 #endif
1908 return (rc);
1909 }
1910
1911 static moduledata_t t4_tom_moddata= {
1912 "t4_tom",
1913 t4_tom_modevent,
1914 0
1915 };
1916
1917 MODULE_VERSION(t4_tom, 1);
1918 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1);
1919 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1);
1920 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY);
1921